Jensen Huang is the founder and CEO of NVIDIA, which Lex Fridman calls the world’s most valuable company, and he leads 43,000 employees. He grew up in Taiwan and came back for the first time in 30 years when NVIDIA needed a foundry, starting a TSMC partnership that has now run for 3 decades. Fridman spent a few days with him around Computex in Taipei, on an island Fridman says makes more than 90% of the world’s most advanced semiconductors. His through line: NVIDIA does best when its partners win. Between night markets, Huang talks through NVIDIA’s new hardware, how Taiwan built its industry, export controls and how he chooses partners. These signals are from an interview that Lex Fridman did with Jensen Huang on the Lex Fridman podcast.

The Hardware

1. Easy to Build, Easy to Trust. Huang treats ease of assembly as a reliability feature. The new NVIDIA system he showed Fridman at Computex has no cables and no hoses and is 100% liquid cooled, up from 80% in the last generation, Grace Blackwell. Assembly time has fallen from 2 hours to 5 minutes, and Huang confirmed it is in full production. His rule: “Things that are easy to build are also reliable.” ↗ 4:27

2. Tokens, Not Cores. Huang says the economics of the CPU have changed. Hyperscale x86 CPUs were designed to be rented, so the goal was as many cores as possible, priced in dollars per core. AI does not rent cores. It generates tokens and sells them, so the CPU has to move through that loop as fast as it can, and Fridman cites NVIDIA’s Vera CPU at 1.8 times x86. The metric that matters, in Huang’s framing, is tokens generated, not cores rented. ↗ 4:50

The Island

3. Bring the Island Along. Huang aimed his Computex keynote at the supply chain. To build and sell a few trillion dollars of computing in the coming years, he says, the entire island has to be part of it, and every company here must invest hundreds of billions of dollars. He is nervous before every keynote, which has no teleprompter because everything comes together at the last moment. His goal was to make partners understand it, want in and see how they benefit, in as few words as possible. ↗ 9:21

4. Reindustrialize Now. Huang calls the AI boom America’s best chance to invest in the grid, sustainable energy and manufacturing. Demand is so high that companies can afford to manufacture anywhere, so they might as well build in the US, and AI and robotics make this the right moment to bring lost industries back. Taiwan is a strategic partner, sending engineers and money to US fabs because there is not yet enough talent there. His advice to American industry: take the opportunity. ↗ 14:57

5. The Philips Seed. Huang traces Taiwan’s chip industry to a government vision. In the 1950s, K. T. Li set out to bring science and technology manufacturing to the island, and Huang says that vision compelled Morris Chang to come back. Li also attracted Philips, the GE of Europe in Huang’s words. By his recollection, Philips spun off the fab that became TSMC and the equipment business that became ASML. In his telling, that one long-term vision was the beginning of the whole industry. ↗ 16:24

6. The Open Spec. The next break came when IBM opened the specifications for its compatible PC chipsets, from cache and memory controllers to graphics controllers, Huang says. Taiwanese engineers designed those chipsets and had them made at TSMC and UMC, so all kinds of companies could build PCs. A young Michael Dell bought Taiwanese motherboards, licensed Microsoft’s operating system and became a PC maker. In Huang’s account, that open spec is how the PC industry really took off. ↗ 17:43

7. Hiring 3,000 More. NVIDIA has about 1,000 people in Taiwan, Huang says, mostly hardware specialists in supply chain, systems engineering, operations, fab, packaging and test. It plans a new building on the island for 4,000 people, which means hiring another 3,000 once it secures a piece of land. The move matches how he describes the company later in the trip: NVIDIA is built on top of the Taiwan platform. ↗ 26:53

8. Generosity Is the Supply Chain. Huang ties Taiwan’s place in the industry to its culture. After several thousand people lined up to see him on the Computex floor, he said the island’s capacity for generosity and helping others is why its companies sit at the center of the ecosystem and serve as the world’s supply chain. Fridman adds a willingness to work very hard, and Huang agrees it takes both: the generosity, and working super hard. ↗ 35:36

The Flywheel

9. Goodwill Isn’t Zero-Sum. Huang says being looked up to by millions does not weigh on him, because people have always been nice to him and he has always been nice back. At school he hoped every other kid would do well and happily shared his homework. Today, he says, 43,000 NVIDIA employees want him to succeed, and he wants people at other companies to succeed too. His line: goodwill is not a zero-sum game, and “it’s not limited by physics.” ↗ 43:11

10. Optimize the Flywheel. Asked to steelman both sides of export controls, Huang grants that denying others a force multiplier works for physical things like refined rare earths, bullets and bombs. But if the goal is long-term national security, he argues, prosperity is the better path, because the safest countries are the richest and technology and wealth feed each other as a flywheel. His recommendation: be extremely selective about access to certain technologies, and otherwise optimize for the flywheel. ↗ 44:32

11. Reach Funds R&D. America’s flywheel is the best, Huang says, partly because most of the world responds to English, so products built on an American tech stack and standards can deploy almost anywhere. That is not true for China. “Technology without customers is useless,” he says. Flywheels are also delicate: NVIDIA’s took 3 decades to reach a tipping point and could slow quickly. His conclusion is to seek the widest platform reach, because reach creates the wealth that funds R&D. ↗ 51:23

The Foundry

12. Capacity Cost You IP. In the mid-1990s PC boom, everyone from NEC and Toshiba to Hewlett-Packard wanted a foundry, Huang says. In the US, Intel, AMD and TI made only their own chips, and SGS-Thomson turned NVIDIA away. Capacity was so scarce that using a foundry meant licensing your technology to it, while TSMC and UMC offered capacity with no strings attached. Huang sees the same pattern in some early AI deals, where access to capacity means licensing something away. ↗ 1:10:16

13. Write to the Founder. TSMC was then about a $300 million company, Huang says. He found its US number in the Yellow Pages, reached a recording, left messages several times and never heard back. So he wrote Morris Chang a long letter. Much later, around 6 one evening, Chang called while the team outside was playing the first PlayStation. Huang’s reaction: “Maybe we won’t have to die.” Chang invited him to Taiwan, his first trip back in 30 years. ↗ 1:21:02

14. Choose Who, Not What. TSMC and UMC sat across one street, and Huang liked both. He chose TSMC because Morris Chang kept telling him what TSMC could not do, and talked only about governance, integrity, partnership and the long term. Chang promised that if they worked together, TSMC would jump through hoops for NVIDIA, and Huang says it has for 30 years. His lesson: with a foundry you are choosing not what to buy, but who to work with. ↗ 1:23:08

15. Leave Something on the Table. Huang keeps the TSMC relationship deliberately loose, because a contract as specific as an ironclad prenup makes a relationship brittle. The two practice what he calls rough justice: sometimes TSMC wins, sometimes NVIDIA does. Forcing everyone to bid until nobody makes money but you is remembered on the island forever, he says. Last year his supply chain’s stocks rose 300% while NVIDIA’s stayed flat. His rule: always leave something on the table, and think about your partners’ success. ↗ 1:25:42

Watch the full Video Here.

Ari Emanuel started Endeavor in 1995 with 3 partners, a table and a phone, after leaving ICM. He later merged it with William Morris, an agency about 100 years old. With Egon Durban he bought IMG for $2.4 billion and the UFC for $4.2 billion, a deal people called the most expensive in sports history. Today he runs TKO, a pure sports business, and MARI, a new live events company he started with Mark Shapiro. He is dyslexic, spent decades fearing he would be found out, and his book opens with the word rage. He talks through the deals, the calls and how he outgrew the rage with David Senra. These signals are from an interview that David Senra did with Ari Emanuel on the David Senra podcast.

The Bet

1. Create Serendipity. Emanuel’s partnership with Egon Durban started with an article. While raising money for Raine, the merchant bank he started with Joe Ravitch and Jeff Sine, he read a Marc Andreessen article, met him and asked him to invest. Andreessen agreed to do it alongside Durban. After IMG’s owner, Ted Forstmann, died, Durban asked to buy a piece of Emanuel’s company and go after IMG. He came back 3 weeks later with an offer Emanuel could have signed as is. Emanuel calls this creating serendipity. ↗ 3:42

2. Own a Piece of the Game. Emanuel, Mark Shapiro and Brent Richards pitched Durban a bigger idea. IMG had added events, sports, production and licensing to the agency’s movies, TV, music, theater and books. Because they already negotiated sports, film and TV rights with networks worldwide, they should put ownership on top and own something in sports. Their reasoning: streaming would move toward ad-supported tiers, and sports keeps people engaged. About 2 years later, the UFC came up for sale. ↗ 6:45

3. It Starts With a Four. As Senra retells it from the book, Emanuel asked Lorenzo Fertitta to close his eyes and name the number that would make him happy. The answer started with a 4. “I just needed the framing,” Emanuel says. He had worked with the company for 20 years and knew how much Fox needed it. The $4.2 billion deal was pieced together through the Brexit turmoil, with Michael Dell taking the preferred and KKR’s founders backing them, and the stress made Emanuel physically sick. ↗ 8:00

The Rights

4. Walk Out With the Number. After the deal closed, the expected TV buyers vanished. Rupert Murdoch sold Fox’s assets to Disney, the AT&T deal slowed, and Amazon’s offer fell from about $260 million to about $140 million. Then ESPN’s Kevin Mayer called. Mid-pitch with Bob Bakish at Viacom, Emanuel got a WhatsApp from Mayer agreeing to their price, $300 million. He slid the phone to Shapiro and ended the meeting. It was the deal he wanted, with 4 months left before it became a nightmare. ↗ 10:33

5. Hold Out for 10. The ESPN deal put 8 episodes on ESPN and the rest on ESPN+. Dana White refused unless it was 10. Emanuel had no other buyer, but called Mayer anyway and argued ESPN needed 10 to build stars and sell its paywall. Bob Iger told Jimmy Pitaro to give him the 10. Emanuel then pushed ESPN to take the pay-per-views as well, and it did, in a 7-year deal he puts at almost $4 billion, close to the UFC’s purchase price. ↗ 17:17

The Merger

6. The 4-Year Merger. When Endeavor first approached William Morris in 2003, its executives laughed. Emanuel says he blew that meeting by going crazy and pushed the merger back 4 years: “I played a horrible game of chess.” In those years he and his partner Patrick stole clients and agents, weakening the older agency, and listened for trouble. When a senior William Morris agent approached him at a holiday event, talks restarted and moved in about 6 months. The dissent inside, he says, let them be the conquerors. ↗ 24:00

7. $44 Million to Close. Even after both sides voted yes, senior William Morris equity holders warned the merger would not work. Emanuel met separately with 2 big earners who disliked its leadership, Peter Grosslight in music and Mark Itkin in non-scripted, and brought them into the plan. Endeavor then paid $24 million to buy out Jim Wiatt and $20 million more for another executive who lacked two-thirds board support, in the middle of the 2008-09 crisis. Emanuel calls it the greatest deal he and Patrick ever did, by far. ↗ 26:02

8. The No-Collateral Loan. Endeavor began when Emanuel, Tom Strickler, Rick Rosen and David Greenblatt left ICM with a table and a phone. ICM sued, their commissions sat in escrow and none of them had money. Russell Goldsmith at City National backed them with $1 million and no collateral, then backed them again in 2009. Years later, Goldsmith was selling City National and asked for some of the IMG deal’s bonds. Emanuel asked Durban for $300 million worth, repaying a debt he had promised to settle someday. ↗ 30:02

The Playbook

9. You Won’t Remember $400 Million. In the IMG auction, a rival bid about $1.9 billion. Durban pushed Emanuel to offer $2.3 billion to $2.4 billion and take the deal off the table in one swoop, so no one could bid them up. His line about the extra $400 million: “You’re never going to remember it.” Senra notes that Emanuel has since used it on his own deals. Even with the right strategy, Emanuel says, the road will be bumpy, and saving $100 million or $200 million is not the game. ↗ 32:32

10. Get in Traffic. Emanuel was nervous about merging with William Morris, and David Geffen, an adviser and mentor, told him to “just merge with William Morris, get in traffic and get hit.” Geffen cuts through a business in a line or two, Emanuel says, and does not listen to the voice in his head. Emanuel admits he should have listened much earlier. His reading of the advice: you will work harder than anybody, there will be good and bad, and you will survive. ↗ 39:21

11. Roll the Calls. Walking into the office or getting into his car, Emanuel tells his assistant to roll the calls. He reads about someone, calls them and keeps calling until they call back, using humor and self-deprecation so it does not feel like stalking. Around 1992 he read George Gilder’s Life After Television and called him. That led to Endeavor’s founding thesis in 1995: distribution would change, and content creators would become more valuable. He met Elon Musk the same way, after reading about his car company. ↗ 45:19

12. Live Can’t Be Replaced. Musk was talking about AI in 2007, Emanuel says, and told him sports would matter. His other point was that live entertainment is the one thing AI cannot disintermediate. That became the idea behind MARI, which sits beside TKO’s pure sports business. MARI’s assets include the Miami Open and Madrid Open in tennis, Escape from Alcatraz, Frieze, Barrett-Jackson and Collectcon in the Pokémon space, and it aims to cover live events globally. ↗ 51:58

The Character

13. Rage Has a Ceiling. Getting hit by a car at 32 taught Emanuel that life is not a dress rehearsal. Over time he concluded that rage and anger would never get him to really big heights. He was not enjoying his life, and he saw the toll it took on his marriage. Jeff Bezos once told him to be fully on the field and fully off it when he leaves, because time off the field makes you more potent on it. ↗ 1:00:23

14. Forget the Legacy. In 2022, after his brother Zeke told him it could not kill him, Emanuel tried 5-MeO-DMT in New Mexico. He says it showed him how small we are and helped with what his therapist calls the ego project. His conclusion: legacy does not matter. Almost nobody knows that the most important businessman of 1953 was Alfred Sloan of GM. Senra raises Marc Andreessen’s worry that psychedelics make founders work less, and Emanuel dismisses it. ↗ 1:10:10

15. Entourage Opened Doors. Emanuel represented Mark Wahlberg, whose crew inspired Entourage, and he expected the show to hurt his career. Wahlberg and his manager, Stephen Levinson, told him it would be the greatest thing that ever happened to him. Afterward, Emanuel says, he could get anybody on the phone, and the pieces of his puzzles got easier to put together. Senra’s version: there is no such thing as a cold meeting anymore. ↗ 1:18:16

Watch the full video at

.

Alex Atallah started OpenRouter 3 years ago. Before it, he says, AI had just one player, OpenAI, and no real marketplace for models. OpenRouter set out to let companies use many models without lock-in, and Stripe has since acquired it. Amjad Masad, an OpenRouter investor alongside a16z, runs Replit and has spent almost a year making it deployable inside customers’ own clouds. Both argue against a single god agent that does everything. They talk through the Stripe deal, why enterprises want to own their intelligence, the case for specialized agents and the economics of small models with Erik Torenberg. These signals are from an interview that Erik Torenberg did with Alex Atallah and Amjad Masad on the a16z podcast.

The Deal

1. Why Stripe Won. Atallah was not thinking about selling. He had talked with Stripe president Will Gaybrick a couple of years earlier, during OpenRouter’s Series A, and the two stayed in touch while OpenRouter ran several projects with Stripe teams. Stripe reached out in July, and talks moved quickly. Stripe was his top choice among possible acquirers, aligned on OpenRouter keeping autonomy over its brand, roadmap and product, with a much more serious go-to-market plan. His line: “payments and inference are going to blend together.” ↗ 1:32

2. Neurodiversity Is the Moat. A company building on AI has to answer a buyer who wonders whether to just use the model directly. Atallah’s answer is neurodiversity: combining models trained in different ways, including your own, to do more on a task than ChatGPT or Claude would alone. OpenRouter also aims to drive down costs by building an efficient market. Some businesses cannot emerge until they become cost-effective, he says, and a provider with a captive market has no reason to lower prices. ↗ 5:33

The Enterprise

3. Enterprises Want Options. Enterprises have been more open to open-weight models than Atallah expected. Instead of buying what other credible enterprises buy, many wanted to diversify beyond the proprietary frontier labs, for cost and for differentiation, and to own their intelligence. Every enterprise now has an internal AI team, he says, and its board asks every month what comes next. He is surprised there has not been more benchmarking, and expects evals to become a bigger focus for those teams. ↗ 9:08

4. The Independence Layer. Masad argues every company needs its own AI practice, the way every company has software engineers, a point he credits to Microsoft’s Satya Nadella. Palantir’s Alex Karp has raised the risk that frontier labs move into partners’ businesses, as Masad says happened with Harvey and OpenAI. The labs want to subsume a big part of the economy: he cites a $30 trillion figure in the SpaceX S-1, against world GDP of about $100 trillion. Replit’s answer is an independence layer across models and clouds. ↗ 12:24

5. Bring Your Own Cloud. Two years ago, Masad says, he would never have let customers run Replit on their own cloud, because software as a service looked like the future. Now companies are more protective, since agents create so many ways for data to leak. Replit spent almost a year making the product deployable on customers’ own clouds. The consumer use of agents is obvious, he says, but making them do real work inside an enterprise is still an unsolved problem. ↗ 16:36

The Agents

6. Agents Carry No Responsibility. Masad likes one agent that joins his chats, GitHub, Salesforce and calendar. Atallah takes the counter. The more work you give an agent, the more understanding you sacrifice, “and yet no one new is taking responsibility.” Each time he improves his own general agent, he ends up ignoring its output about a week later. He would rather have something like 10 chiefs of staff, each responsible for one area, which lets him tune how much understanding he gives up. ↗ 19:53

7. Machines Should Specialize. Masad says the debate rediscovers Adam Smith, for whom specialization was a huge realization for humanity. People then over-specialized, he argues, citing Marx’s theory of alienation, in which workers never see the fruits of their labor and end up feeling like machines. The idea of one god agent may be a reaction to that, he suggests, even though specialization is good for machines. His line: humans should be general, and machines should be specialized. ↗ 23:36

8. Access Control Forces Specialists. No system of specialized agents yet feels as elegant as ChatGPT or Claude, Atallah says. One use case people seemed to like came from Grok bot: one bot for a bank account, another for Twitter, with no shared credentials. Masad adds the enterprise reason. CEOs can run general agents because they have admin access, but individual employees cannot, because of access control. He also wants a DSL or protocol for agent-to-agent communication, so one agent cannot talk another into giving up data. ↗ 25:11

The Risk

9. Guard Every Tool Call. Agents make so many tool calls that checking them takes a cheap, fast model, Atallah says, and OpenRouter runs a small internal prototype along those lines. A decision model reviews each tool call against the agent’s system prompt and against extra guidelines the agent was never told. His example is red-team agents told to act like bad actors, under a hidden rule to stop the moment they reach the internet. He expects companies to combine checks like this with structural safeguards. ↗ 28:36

10. Models Train Their Replacements. Masad says one idea gets little discussion: models training their replacements. He compares it to a just-in-time compiler, which spots a chance to optimize running code and emits faster machine code on the fly. A general model could notice a limited use case and train a domain-specific model on the spot. That model would be cheaper, less vulnerable to prompt injection and less harmful, because it is less capable. Using frontier models for most tasks, he says, is “like nuking a butterfly.” ↗ 31:13

11. Deception Scales Too. Asked by Torenberg whether smarter models are easier to align, Masad says machines may go the other way. The orthogonality thesis holds that intelligence is independent of ethics, and he thinks it does not fully hold for people. In models, though, research on reinforcement learning shows reward hacking and deception getting better. Models can know they are being evaluated, and heavy monitoring of their chain of thought leads them to lie in it. Proper alignment evals, he argues, need to run for months. ↗ 35:23

12. Pay for Risk, Not Size. Asked whether companies would pay 10 times more for a model with no risk of deception, Atallah says it depends on the task. Writing code and security research are the highest-risk tasks today, so there a fully anti-deceptive model that can find all the bugs is probably worth 10 times the cost. Decision models, whose structured outputs are fully controlled, leave much less room for misbehavior. He expects enterprises to get a lot more interested in them. ↗ 38:42

The Payoff

13. Small Models, Less Debt. Masad has been training small models such as Qwen 8B by asking models like Fable and Opus to do the training. One is a cost estimator that tells Replit what a prompt will cost, assigning probabilities to buckets such as $5 to $10 and $10 to $20. With so much data at Replit, he says, specialized classifiers are easy to train. Companies worry that fine-tuned text models must be redone every couple of months, but a bespoke classifier on proprietary data carries less model debt. ↗ 40:53

14. The Rust Cycle. Masad expects AI to repeat a cycle from programming. In the 1990s developers wrote Java and C++, then Python, JavaScript and Ruby took over the internet, with Stripe built on Ruby and Facebook on PHP. Bugs and slowness followed, so teams added types and JIT compilers, and then Rust arrived. He predicts companies will find general models wasteful and risky for many uses. Replit is adding a way to upload a CSV file and get a specialized model that does one thing. ↗ 43:25

15. Fusion at Half the Cost. Code review was the first place Atallah saw people seriously use different model families to double-check a main model, and research on mixing models is now speeding up. OpenRouter launched a fusion model, starting with deep research, on the logic that labs train on different data, so it pulls from all of them. He says it delivered roughly Fable-level quality at half the cost, and Cognition launched one too. Masad says Replit just published frontier-level results at 40% to 50% of the cost. ↗ 45:12

Watch the full Video Here.

Pat Gelsinger joined Intel at 18 as a technician after skipping his last 1.5 years of high school. He became engineer number 4 on the 386 and architect and design manager of the 486, earning bachelor’s, master’s and PhD degrees along the way. He later served as Intel’s CTO, VMware’s CEO and Intel’s CEO, and is now a general partner at Playground Global. The tension he describes: AI is making chip design easy, which pushes the bottleneck into manufacturing, memory, packaging and power. The conversation moves from the 486 and AI-driven design to the AI chip shakeout, memory and stacking, optics, energy and agents. These signals are from an interview that Raghu Raghuram and Guido Appenzeller did with Pat Gelsinger on the a16z podcast.

The Design

1. Inventing the Toolchain. The 486 was the first chip built with modern electronic design automation (EDA) methods, Gelsinger says, because the tools did not yet exist. With no Verilog, the team wrote an Intel hardware description language and built its compiler. With no automatic place-and-route, Intel worked with Alberto Sangiovanni-Vincentelli and his Berkeley students on the first automated placement, routing and timing. The chip’s lasting legacy, he argues, was the foundation it laid for today’s EDA industry. ↗ 6:05

2. AI’s 486 Moment. Chip design is entering a transition like the 486’s, Gelsinger says, when asked whether Nvidia’s Blackwell will be remembered as the last chip designed before AI took over. Many logic functions can now be done with AI tools, he explains, because transistor budgets are large and the tools are smarter. Analog blocks such as SerDes remain hard because they need lots of silicon data. His working model for the new era: a few experts guiding the tools. ↗ 7:25

3. Design in 3, Wait 9. Faster design alone does not speed up a chip, Gelsinger argues. Even if AI tools yield a strong design in 3 months, fab takes at least 3 more, advanced 3D packaging another 3 and rack-scale integration the rest, because “nothing’s a chip anymore. It’s a rack.” Masks can cost $50 million before a prototype exists. He wants that 9-month tail cut to 1 or 2 months, since 18 months to scale leaves workload assumptions out of date. ↗ 9:46

4. Memory and Power Walls. Memory is the next bottleneck, and Gelsinger calls high-bandwidth memory (HBM) “a hideous memory,” though the best available. Its bit density is poor, its bandwidth is capped by the chip’s edge, or shoreline, and DRAM dislikes the heat beside compute. Power is the other wall. Weak 3D power-modeling tools force designers to reserve safety margin across almost the full envelope, losing about 40% of power. Better power modeling, he says, is the fix. ↗ 11:56

The Shakeout

5. 100 Chips Won’t Last. Today’s roughly 100 AI inference chips will converge to a few, Gelsinger predicts, for 3 reasons. Splitting fleets into prefill, decode and midfill tiers is not sustainable, and some now want CPU-like hardware again for reasoning models. Models themselves will shift toward 3D, chemistry and 64-bit precision work. And scale takes capital and workloads, so players such as OpenAI, Nvidia and Anthropic will pick winners by how hardware and software co-evolve. ↗ 14:50

6. Big Players Absorb Variety. A host pushes back: if AI agents can write optimized kernels overnight, the software cost that once limited the industry to 1 or 2 architectures disappears. Gelsinger agrees the bottleneck moves, but to manufacturing and capital, since deploying a new chip at scale can mean $50 billion in data centers and power commitments. Big players will absorb the variety instead, as Nvidia did with Groq. In 2 years, he predicts, nobody will spot Groq inside Nvidia. ↗ 19:14

The Stack

7. 30 Years, Zero New Memories. No major new memory has emerged in 30 years, Gelsinger says, despite close to 100 attempts he knows of and at least 5 he worked on, Optane among them. The industry still runs on DRAM, SRAM and flash, partly because memory made money only 1 year in 5, starving R&D. AI has flipped that: a host notes all 3 big memory makers now rank among the 20 most valuable companies. Gelsinger, backing a stealth memory startup, expects that capital to fund real innovation. ↗ 22:46

8. 3 to 5 High. Asked how tall chips will get, Gelsinger bets on modest stacks. Each added layer raises the value at risk, so per-layer yield must improve exponentially, which makes 16- or 32-high stacks impractical. He expects 3 to 5 layers, with 2- or 4-high memory, to be the sweet spot. Add vertical power delivery and redistribution layers, and a package is already about 8 layers thick before optics. His verdict: an engineering feat and a manufacturing nightmare. ↗ 27:19

9. Keep Memory Close. Gelsinger rejects 2 alternatives to stacking. Moving memory off-package over optical links burns power, because moving a bit costs about 1,000 times more energy than a compute operation. Processing-in-memory (PIM) forces workloads into small pieces pushed into memory, which he finds too constraining. PIM ideas have circulated for 25 to 30 years, he notes, and will likely still be circulating 25 to 30 years from now. His preference: dense, high-bandwidth memory placed close to compute. ↗ 32:14

10. Copper’s Long Goodbye. All I/O should move to optics, Gelsinger argues, though not the core compute-memory complex. He says he declared copper dead about 25 years ago and will eventually be right: copper at 5 meters now costs more than optics at 100 meters. He expects co-packaged or near-packaged optics around 2028 to 2029. Nvidia’s NVL72, which he calls an engineering marvel and a manufacturing nightmare, took the industry an extra 18 months to digest because optical supply chains were immature. ↗ 34:54

11. Flows, Not Packets. AI networks should be built for predictable traffic, Gelsinger says. Citing Stanford’s Nick McKeown, he notes that classic networks assume any packet can go anywhere with no advance knowledge, while AI moves large, predictable flows. That points toward optical circuit switching once optics reach the compute complex. A host calls the scale-up versus scale-out split somewhat artificial, and Gelsinger agrees the boundary dissolves once a scale-up domain is big enough. ↗ 38:03

The Power

12. Energy Is the Ceiling. Energy capacity will cap how far the AI buildout can go, Gelsinger argues. The US spent 10 to 15 years retiring coal about as fast as it added renewables, leaving capacity flat. His line: “energy capacity equals economic capacity.” Growth is now about 4% a year, renewables depend heavily on China, and gas turbine lead times run 8 years. He expects more data center projects to default as buyers ask whether they have the energy to back their GPU commitments. ↗ 43:15

13. The Revenge of Edison. Power delivery inside data centers has to get far more efficient, Gelsinger says, starting with 800-volt DC, which removes conversion steps kept for historical and standards reasons. His phrase: “the revenge of Edison is upon us.” He points to vertical gallium nitride, where Playground has a company, as a way to step 800 volts down to 48, 12 or even 5 in one conversion. Solid-state transformers and cooling follow, which he calls a renaissance for hardware engineers. ↗ 46:17

14. CMOS Has Flatlined. Better delivery and cooling are not enough, Gelsinger says; the physics of compute has to change. By his account, CMOS has not improved power per teraflop across the last 5 generations of Nvidia chips, a flatline he calls unacceptable. His example of a different path is Snowcap, Playground’s superconducting company, which he says could deliver 1,000 times better power performance. He expects some of those physics changes in the near future. ↗ 48:02

The Next Layer

15. VMware for Agents. Every innovation leads to a new abstraction, says Gelsinger, a former VMware CEO, and the virtual machine still deserves a place. Agent swarms need management, security profiles and performance tuning, so each element of virtualization must be rebuilt, down to a vMotion for agents. A host adds that agents are less patient than humans and may pick the VM themselves. Humans still set the policies, Gelsinger says, pointing to Dario’s constitution as one form of guardrail. ↗ 48:47

Watch the full video at

.

Adam Neumann cofounded WeWork in February 2010 and says that by year 9 it was opening 2 buildings a day across 130 cities. Its August 2019 S-1, Bartlett notes, valued the company at around $47 billion. Later that year he stepped down and, he says, went from 13,000 employees to 3 in a week. He now runs Flow, backed by $470 million from Andreessen Horowitz. He argues he got far more right than wrong at WeWork, but the business outgrew him. The conversation runs from his early money chase to Green Desk, board practice, the SoftBank deals, personal discipline, the failed buyout, his exit and the settlement. These signals are from an interview that Steven Bartlett did with Adam Neumann on the Diary Of A CEO podcast.

The Foundations

1. Superpower Plus Mission. At 21, Neumann says, he wanted to be a billionaire, believing money would fill the hole left by his childhood. 3 businesses followed, and none succeeded. Rebecca, then his girlfriend, told him chasing money works like chasing cool, and that money caught that way would harm him. Her alternative: connect his superpower to something meaningful for the world, because employees, investors and consumers all want that company, and the money follows. He proposed 3 months after they met. ▶ 20:12

2. The Green Desk Math. Green Desk, WeWork’s precursor, launched in May 2008 after a pitch to a skeptical landlord. Neumann told the owner, Mr. Gutman, his buildings sat empty and proposed splitting a floor that might rent for $5,000 a month into 15 spaces at $1,000 each. After a $2,500 receptionist, the landlord would get $7,500 and the founders the rest. Gutman offered all 5 floors. 5 Craigslist ads filled it to 92% in a week; within a month it was cash-flow positive, a first for him. ▶ 23:05

The Build

3. Itemize Every Quote. At WeWork’s second building, Neumann had $90,000 for the build-out against quotes of about $200,000 a floor. When wood floors came in at $30,000 to $45,000, he asked contractors to itemize pieces, nails, glue and labor hours, and none could. He offered the top 3 their costs plus a 15% margin in exchange for full visibility, and all agreed. The first floor cost $12,000, and the contractors later joined WeWork as an in-house construction company. His rule: ask why. ▶ 37:27

4. Outgrown by the Business. WeWork opened 1 building in its first year and, Neumann says, 2 a day by year 9, each around 100,000 square feet. Many of the same early staff led that pace. For Bartlett’s roughly 200-person company, he suggests sitting 30 staff in a circle so each can name the internal barrier holding them back. His diagnosis: WeWork grew faster than he did. A company that outgrows its founder, he argues, either crashes or leaves the founder unfulfilled. ▶ 41:57

5. Partners Who Tell the Truth. Asked how to judge hires outside his expertise, Neumann says he picks partners who tell him the truth. Andreessen Horowitz put $470 million into Flow across the Series A and B, and his family $350 million, so both sides buy and sell together. After each all-day board meeting, he asks Marc Andreessen and Ben Horowitz over dinner what he can do better. He vets investors by their history with founders, citing Benchmark, which he says pushed out him, Travis Kalanick and Jack Dorsey. ▶ 50:48

6. Influence, Not Control. Neumann says that when he has a problem with an employee, the only person to blame is himself, and that he is shifting from top-down to bottom-up leadership. Talented people can always leave, he argues, so a reporting line does not oblige them; they want inspiration, room for mistakes and direct feedback. He warns against sitting on criticism for 90 days to avoid conflict, and says to invite feedback in return. His line: “power comes from influence, not control.” ▶ 53:37

The Deal

7. Four-Wall Payback. Neumann separates WeWork’s building economics from its consolidated losses. A store or restaurant does well to return its investment in 24 to 36 months; he says WeWork buildings at their peak did it in 6 to 9 months. New sites lose money while they ramp, so a fast-growing city and the P&L look unprofitable, and stopping growth would have let demand fill supply. The real mistake, he says, was signing long-term leases against short-term commitments, which Flow avoids. ▶ 57:55

8. 12 Minutes, $4.2 Billion. The night before meeting SoftBank’s Masayoshi Son in 2016, Neumann’s board agreed to raise $300 million to $400 million, stop growing and go public. Son’s 90 minutes shrank to 12 because he was meeting President-elect Trump; Neumann spent 5 of them texting Jared Kushner to arrange Son’s introduction. In the car, Son dismissed the $400 million, proposed $3 billion, then settled on $4.2 billion, signed on an iPad. Neumann’s advice: before accepting 10 times your plan, ask why, and whether you are ready. ▶ 59:18

9. Ego Is Desire. Neumann dates WeWork’s decline to the walk from 54th Street back to 18th Street headquarters. With a $20 billion valuation and a stake of about 25%, he tallied his net worth with each step. “Ego and desire are the same word,” he says. A high-growth company makes 100 decisions a day and needs 98 right; a founder blinded by ego slips to 95 or 90. Because WeWork was led from the top, his loss of mission spread to everyone. ▶ 1:08:30

The Discipline

10. Technology Shabbat. For about 10 years Neumann has kept a 25-hour Shabbat from Friday, and he recommends a secular version to anyone: 24 hours fully off the phone. Put it in a cookie jar or give it to a friend. In the first 2 hours, he predicts, constant reaching for the pocket will reveal the addiction. Bartlett, who first said he saw no upside, agreed to try it in September. Neumann’s promise: some of your best ideas in a while, including who to surround yourself with. ▶ 1:28:34

11. Creator, Not Protector. Focus is not his superpower, Neumann says; zero to one is. Horowitz’s contrast: another founder brings 9 ideas to a board meeting and executes 1; Neumann brings 9 new businesses. He cites his wife’s grandfather, Andre Herz: “It takes one lion to make a fortune and 10 lions to watch it.” He can run 3 to 5 initiatives, not 10 to 15, so he needs trusted people who say no. After his ego took over, WeWork drew hires who came for the money. ▶ 1:35:24

The Fall

12. The $20 Billion Holdout. In Neumann’s account, Son offered in March 2018 to buy WeWork for $20 billion in cash, $10 billion to investors and $10 billion to the balance sheet, while the company lost about $2 billion a year. A special committee including the Benchmark board member held out until October, aiming for $32 billion. By December 24, 2018, when documents were ready, SoftBank’s stock had fallen and Son withdrew. Already committed to leases, WeWork was pushed into an IPO it was not ready for. ▶ 1:39:15

13. Get It in Writing. Neumann says he stepped down by choice, while controlling the board. A head of one of the world’s 2 biggest banks offered $2 billion for WeWork if he became executive chairman, and waved off his $460 million in personal debt, though never in writing. 10 minutes after the board call, the bank cited change of control and gave him 15 days to repay or forfeit his 5-to-1 voting shares, and with them WeWork. His line: “If you’re going to stab the king, kill.” ▶ 1:50:05

14. The Personal Guarantee. Days later Neumann woke up crying, the first time his wife, Rebecca, had seen him cry. He had agreed to put $100 million in her name and never did, and had signed a personal guarantee he had promised never to sign. She proposed living at her mother’s mortgage-free house upstate, then returning to business or surfing in Costa Rica. That morning he smiled through a meeting with his 3 remaining partners. His lesson: choose a partner for your worst day, not your best. ▶ 1:56:23

15. Rebecca’s Ultimatum. After Son repaid the $460 million and took control, SoftBank agreed to pay Neumann and other investors $3 billion on April 1, 2020, then invoked force majeure the day before. Litigation ended in a $1.5 billion settlement, $500 million of it to the Neumanns, payable in 2 months. Days before a scheduled C-section, Rebecca refused the delay, since SoftBank had already missed one deadline: pay by Friday 8:00 a.m. or go to court Monday. The money landed at 4:30 a.m., hours before she gave birth. ▶ 2:03:01

Watch the full Podcast Here.

Matt Garman interned at AWS in 2005, during business school, when it was still an internal project, and went on to become EC2’s first general manager. As CEO, he now runs a business he puts at about $169 billion to $170 billion in revenue, growing 37%. The tension, as Raghuram frames it, is that frontier labs gobble up every available GPU while startups, the enterprises of tomorrow, still need capacity, and agents are replacing the developers AWS was built to serve. The conversation runs from startups and agent-ready infrastructure to GPU allocation, bubble risk, power and supply chains, AWS’s custom chips, enterprise adoption, security and AWS’s own teams. These signals are from an interview that Raghu Raghuram did with Matt Garman on the a16z podcast.

Startups and Agents

1. Startups Become Enterprises. Garman’s internship project at AWS in 2005 was an analysis of who AWS would appeal to most, and the answer was startups. He estimates, calling the number imprecise, that 30% to 40% of AWS revenue now comes from companies that were startups at some point in AWS’s lifetime. Startups also push AWS harder than banks, healthcare companies or governments. His reasoning: what startups ask for today, banks will want 5 years later. ↗ 2:50

2. Built for P999 Latency. Agents and people use many cloud services in similar ways, Garman says, but agents reach for data differently. AWS Context, now in preview, builds a context layer so agents can find data across S3, Aurora and other AWS stores. Agents also care about tail latency, and slow P999 S3 responses block them in a way people rarely notice. He credits that focus on performance for his claim that agentic workflows run better on AWS than anywhere else. ↗ 7:55

3. An Account in 30 Seconds. Garman says the vast majority of customers now tell a coding agent, such as Kiro, Claude or Codex, to deploy on AWS. For 20 years, a new account meant defining VPCs and IAM roles first. AWS is slowly rolling out accounts that need no credit card, accept a Gmail sign-up and are live in under 30 seconds. They are full accounts, so growing later needs no migration. His goal: “not a choice for the customer but an easy on-ramp.” ↗ 9:58

4. Agents Get Scoped Permissions. Agents need some brand-new building blocks, Garman says, including compute sandboxes, gateways and their own permissions. His line: “You don’t just want to give it Raghu’s permissions and let it go do whatever you can do.” An agent should get short, time-boxed, fine-grained rights for one task. For sandboxes, Raghuram notes that many startups build on Firecracker, the microVM AWS created about 10 years ago. Garman says it starts fast, isolates well and carries little overhead. ↗ 14:50

The Buildout

5. Capacity Held for Startups. Garman puts capex at $220 billion for 2026, possibly the largest single-year expense any company has had. Power, data centers, memory, chips and construction crews each become the constraint at different times. AWS could sell every accelerator to the big frontier labs, he says, but it deliberately holds capacity for enterprises and startups. It says yes in some form to about 60% of requests, and plans to buy 2 million Nvidia GPUs over the next couple of years. ↗ 17:25

6. Single-Digit Concentration. Asked about a bubble, Garman points to customer concentration. Some neoclouds have concentrations of 30% to 60% with 1 or 2 customers, while AWS’s largest customer is a single-digit percentage at most. Most AWS usage is core compute, storage and inference for production work, and enterprises tell him, almost to a person, that they see positive returns at today’s costs. His argument: as with the internet bubble, some startups will fail, but durable businesses keep spending. ↗ 22:09

7. Bring Your Own Power. Garman says that 15 years ago, AWS simply asked the utility for more megawatts. Now it pays for its own solar, nuclear and other power projects, on the grid and behind the meter. AWS has been one of the largest buyers of renewable power every year for 10 years. Power and transmission planning runs 20 years out, and server planning spans several years instead of a few quarters. His point: no customer can plan its 2028 memory footprint, so AWS does it for them. ↗ 25:22

8. Always the Next Constraint. Garman borrows from The Goal, a book he read as an undergraduate: “there’s never one constraint, there’s always just the latest constraint.” When power eases, memory, TSMC capacity, HBM, networking parts or connectors take its place. Location matters too, since spare power in Indonesia does not fix a shortage in Germany. AWS tracks tens to hundreds of thousands of components, 4 to 5 tiers down its supply chain, a discipline it began about a decade ago. ↗ 28:07

The Silicon

9. Graviton’s 20/20 Math. Graviton grew out of Nitro, Garman says: AWS had acquired a company whose offload cards carried ARM cores, then asked what would happen if those cores became a server. The first Graviton was small and underpowered. Today it is 20% cheaper with 20% better performance, a gap he says has held for 5 to 6 years, and 90% or more of AWS’s top 100 customers use it. Some have halved their server count. He calls it the easiest way to cut an AWS bill. ↗ 35:46

10. Trainium Outgrew Its Name. AWS began Trainium about 5 to 6 years ago and now sells Trainium 3, sold out until roughly the end of next year, Garman says. Most Bedrock traffic runs on it. Anthropic, OpenAI and 6 to 12 startups build on it, and Trainium 4 is announced. He concedes AWS is bad at naming: its training chip may now be the best inference chip on the market, on both absolute and cost performance. ↗ 37:39

The Enterprise

11. Don’t Copy Bob’s Steps. Most enterprise agents are still simple and keep people in the loop, Garman says, though they already pay off. The first barrier is design. Companies have an agent repeat what an employee he calls Bob does, steps 1 to 5, then have Bob check the work. Garman urges a blank slate, because an agent can run massively in parallel and try 50 different things. The value comes from asking how a computer would solve the problem, not copying how a person does. ↗ 40:44

12. FDE With an Exit Date. The second barrier is trust, Garman says, because no enterprise wants an agent that deletes a production database. Few companies know how to build evals, label data or back-test in production to catch drift, and he is not sure anyone does it well yet. AWS and its partners are leaning into forward-deployed engineering. His model is a 45-day engagement that trains the customer, then leaves, because customers do not want to depend on an outside workforce for years. ↗ 42:48

13. Bedrock Keeps Data Home. Raghuram cites industry peers who say enterprises should post-train open models instead of sending all their data to a frontier model. Garman agrees that enterprise data is the most valuable asset. Bedrock guarantees data never leaves the customer’s VPC, and model providers never see prompts. Most open-weights post-training happens in SageMaker today. AWS took heat 3 years ago for being slow on AI, he recalls, but the vast majority of customers moving to production now land on Bedrock. ↗ 45:33

14. Security at Machine Speed. CEOs mostly ask Garman how they can trust agents to do what they intend. He sees powerful models as both a real attack risk and a real opportunity. AWS recently launched Continuum, which uses those models to find vulnerabilities customers had missed and rank them by context such as permissions and compensating controls. It exposes what AWS already does internally, he says. His line: customers will need “security at machine speed, not at human speed.” ↗ 49:28

15. Pods of 3 to 4. Software and product development show the biggest gains inside AWS, Garman says, and launches have sped up sharply in the last year. AWS calls the new units frontier teams: agents write all the code, and engineers manage a team of agents. Work that once took 10 people can now take 3 to 4, who finish so fast that AWS wants to move them to new problems. On how organizations should change, he says he has no magic answer yet. ↗ 53:46

    Watch the full Podcast Here.

Mike Cannon-Brookes co-founded Atlassian 24 years ago and, in his mid-40s, still runs it as CEO. The company behind Jira, Confluence, Trello and Loom generated $6.5 billion in ARR in 2026 and serves more than 350,000 customers. His co-founder, Scott Farquhar, has since retired, and Cannon-Brookes is leading Atlassian into enterprise AI through Rovo and the Teamwork Graph. Speaking from Amsterdam, where Atlassian was announcing new features, he argues that agents raise the stakes on design instead of erasing it, explains why Atlassian keeps Chinese models out of production, and says inference will get cheaper than he expected. These signals are from an interview that Harry Stebbings did with Mike Cannon-Brookes on the 20VC podcast.

The Screen Survives

1. Don’t Give Away the Screen. Agents will take over purely functional tasks, Cannon-Brookes says, the way you just want a charging cable without browsing for one. Collaborative work, full of taste and judgment, is different. So he stands by his line: giving up on your interface is a brainless bet. That isn’t a stance against agents. Atlassian runs one of the top MCP servers in the world, he says, and agents reading its Teamwork Graph show off how good that data is. → 3:43

2. Today’s Great App Is Tomorrow’s Also-Ran. It is easy to build an app that beats 90% of what’s out there today, Cannon-Brookes says, but in 5 years the same app will beat only 20%. Standards keep rising, and keeping up takes R&D and hard craft work. Atlassian is investing in design more than it ever has. His reasoning: when creating technology becomes cheap and abundant, design and experience are what make a product stand out. → 5:32

3. Forget the Super-Agent. Cannon-Brookes expects horizontal agents, deep vertical agents and human interfaces to coexist. In Confluence, you can highlight a number and have it checked against Databricks or Snowflake through the Teamwork Graph, but that doesn’t make Confluence a BI tool. Figma’s agent works inside Atlassian, and Atlassian’s works inside Figma. He calls himself a big bear on the magic-button thesis, the idea that one button will eventually do everything for us. → 8:32

Designing for Agents

4. The Blank Box Problem. For the first time since mobile, Cannon-Brookes says, his design team is solving foundational problems. The first is the prompt box: a blank line with unlimited power behind it, which he calls a terrible interface. Most users stare at it and treat it like Google search. The open questions are how agents should appear in a collaboration tool, how to surface their power at the right moment, and how humans and agents should work side by side. → 9:08

5. Did You Write This, or Did Codex? When Codex opens a pull request, Cannon-Brookes asks, did Harry do it, did Codex do it, or was it Codex on Harry’s behalf? Answering that is now a design problem. He pushes back on watermarking, like Anthropic’s, that implies AI text is bad. A 3,000-word essay generated from one line and your own essay polished by AI can carry the same mark. Nobody needs every agent trace, but people need to understand them. → 11:38

Models and Money

6. Open, Yes. Chinese, Not Yet. Stebbings relays Jason Lemkin’s view that enterprises won’t touch Chinese open models. Cannon-Brookes is more precise: some customers push back on Chinese models, but none push back on open ones. Atlassian runs Nvidia’s Nemotron and Mistral, and no Chinese models in production. His job is to apply the best model to each job, not to pick one. Long term, he thinks refusing a cheaper, better model whose contents you can inspect will be a hard choice to defend. → 17:51

7. Your “Own Model” Is Probably Someone Else’s. The jury is still out on specialized models, Cannon-Brookes says. They make sense with deep proprietary data or truly specialized uses like robotics. But when a company announces its own model with a quirky name, it is usually a Chinese open-weight model with fine-tuning on top. When someone claims they trained one for a million dollars, his response is that they tailored someone else’s billion-dollar model. Atlassian fine-tunes open-weight models too, and doesn’t call them its own. → 21:51

8. Easy Money Skips the Cost Lesson. Cannon-Brookes jokes that he loves well-funded private companies that ignore cost, because they aren’t building the discipline. Getting good results from an expensive model is easy. Getting them from a cheaper or older one is a real advantage. Large enterprises are now highly sensitive to price performance, and Atlassian ships products that show customers their token spend. He thinks managing technology cost will be a defining trait of great tech companies. → 26:28

9. Don’t Count Pennies at Five Users. Atlassian routes its AI calls internally, through what Cannon-Brookes calls a sizable AI gateway team that balances quality, price and speed for each feature. But he tells product teams not to optimize cost too early. Build magical features first. If five people use a feature, he doesn’t care what it costs. If 5 million do, he cares a lot, and there is usually time in between to bring the cost down. → 31:25

10. Inference Prices Will Fall Off a Cliff. Asked what is expensive today and won’t be in 5 years, Cannon-Brookes picks model inference. The capital and talent pouring into price-per-watt performance, from people who might otherwise have joined McKinsey or NASA, is enormous. New chips, better data centers and distillation take years to land, but he expects costs to fall at a precipitous rate, even as usage explodes. It is also what he has changed his mind on most this year: prices will fall further than he thought. → 32:17

Where the Edge Moves

11. Imagination Is the New Inside Information. Right now, Cannon-Brookes says, the production of technology is outrunning its consumption. Atlassian can ship features faster than customers can absorb them. He borrows a phrase from Culture Amp founder Didier Elzinga: the world is moving from information asymmetry to imagination asymmetry. Knowing something others don’t matters less. The edge goes to whoever imagines new ways to apply AI to a problem. For first-principles thinkers, he calls it the most exciting period tech has ever had. → 33:53

12. Your Org Chart Expires Every Day. Plenty of company data never reaches a database, Cannon-Brookes says; it lives in PowerPoint slides. The Teamwork Graph pulls decisions, documents and code into a map of who works with whom, then infers answers before an agent asks, the way Google indexes the web in advance. That map is only valid today, because people join and leave. Knowing he wrote Java in 2003 is useless. Knowing he was once a programmer still helps. Facts must be fresh; lessons can be old. → 39:37

13. Anyone Can Write a Doc. By Atlassian’s last disclosed figure, Confluence alone is past $1.5 billion in ARR and growing more than 25% a year, big enough to be a public company. Yet there are probably 20 ways on your computer to write down a piece of text. Writing alone is no advantage, he says. What wins is design, experience, purpose and integration into other workflows. That is his answer to OpenAI and Anthropic, both Atlassian partners, moving into documents. → 45:50

The Long Game

14. Money Won’t Keep You Going. Stebbings worries that young founders are cashing out too fast. Cannon-Brookes, 24 years into Atlassian, says founders have to understand where their energy comes from, the same lesson he wants his kids to learn. Few people stay motivated by money or status for 5 to 10 years. Every tech wave, from mobile onward, pulls in capital and people who don’t stay. His own energy comes from working with smart people on hard problems. → 50:52

15. Slow Down to Go Fast. The best advice Cannon-Brookes has received is “wage peace on yourself,” because people in tech hold themselves to incredibly high bars. The skill he most needs to build is slowness. Agents make it even easier to stay busy, and the day-to-day churn hides the year-to-year shifts. He has been carving out a couple of hours a week for long, deep thinking. His rule for a world that keeps speeding up: slow down to go fast. → 56:56

Watch the full video at

.

Alexandr Wang ran Scale AI for 9 years, from founding until Meta took a 49% stake and he joined to help start Meta Superintelligence Labs. He is now Meta’s chief AI officer and not yet 30. The lab’s first consumer product is Muse, a personal agent that Wang says has millions of users; charts he has reposted suggest no consumer AI app has grown faster. To get there, he bet against the industry’s obsession with coding, spent 7 months grinding a demo into a product, and sold it with memes he wrote himself. He also explains why he left Scale and how he runs a lab where more than 200 researchers report to him. These signals are from an interview that David Senra did with Alexandr Wang on the David Senra podcast.

The Bet

1. Two Weeks to a Prototype. Nat Friedman, who helped create GitHub Copilot, was the first in the lab to go all in on OpenClaw, the viral personal agent of early 2026. One of Wang’s first sessions, with access to his email and photos, felt like 3 years of therapy. That February, both wrote memos to the board; Wang’s called the super agent the final consumer product. Within two weeks there was a prototype, and many decisions that shipped, including Jolly the mascot, were already in it. → 3:41

2. The Bet Against Coding. At the time, Opus and Claude had a commanding lead in coding, and most of the industry believed coding agents were the only capability worth building toward. Meta went the other way. Wang’s thesis, pitched to Zuckerberg before the Scale deal: AI will take over what people have to do, and Meta’s apps already hold what people want to do. Add 3.5 billion daily users, and consumer AI was always the ground where Meta could do something special. → 9:23

The Build

3. Sand Every Edge. Much of Meta’s culture, Wang says, comes down to one habit: once you know what you’re measuring, talented people can optimize it until it’s right. Before AI, that meant conversion, referral rates and every step of the onboarding flow. For Muse, the same muscle went into finding where the model was weak and closing each gap, one review at a time. In the AI era, he says, the thing you sand is the model. → 14:01

4. Seven Months in AI Time. Seven months may not sound long, Wang says, but in AI it is closer to a decade. From February to September, Zuckerberg alluded to personal agents on earnings calls while Wall Street asked whether Meta, burning all that money, would pull anything off in AI. Wang credits Zuckerberg for holding steady while the team kept refining: “It took a lot of patience and restraint and a lot of trust in the process.” → 17:54

5. Not a Chatbot. Muse landed, Wang says, because it is different from what most people know as consumer AI. It is an agent, not a chatbot. Developers have run agents for a while, but for most users that graduation has to be explained and onboarded with care. The team ran round after round of fresh-install user tests, plus A/B tests across model versions, and found the experience very sensitive to model quality. → 21:06

6. Green on Every Row. The OpenClaw wave faded, Wang says, because the magic was rare and the breakage constant. Meta’s answer was a spreadsheet of more than 100 behaviors, each with an eval, a launch-blocking threshold and a “good” threshold. Checkpoints that were green on 80 rows and red on 20 stayed internal. The first one green across the board, a specially trained Muse Spark 1.3, told the team it was ready. → 22:10

7. One Point of View. Model and product work combined took fewer than 200 people, on purpose, for one of Meta’s most important launches of the year. Wang warns that big organizations drift into “PM hell,” where every product manager jams in a pet feature and the result blends into a smoothie that users can feel. Muse was treated as a single work of art, and Nat Friedman’s taste set its direction. → 22:46

The Launch

8. Make Meta Cool Again. Distribution alone couldn’t make Muse break through a torrential stream of AI launches, Wang says. So he wrote memes on X himself, sometimes mid-meeting, aiming to make Meta cool again. His most casual posts often did best, a reminder that the internet rewards risk and surprise. He believes it moved growth, though Meta can’t track it. Friedman’s verdict: the launch taught them more about Wang’s mind than about users. → 26:34

9. Show What It Does. Most people don’t care about OpenAI, Anthropic or Meta’s research, Wang says. They care about what the product does for them, which is why his team kept asking “why would someone care?” Senra points to Claude Hopkins’ Scientific Advertising, written in the early 1900s, which makes the same case: people are self-interested, so lead with the benefit. Both knock brand-first AI ad campaigns. Wang’s alternative was quote-tweeting dozens of real Muse use cases. → 31:40

10. The Trust Fall. Retention, in Wang’s telling, is a trust fall in small steps. A user hands the agent something minor, it works, then something bigger, then bigger again. Some college students already go all the way, sending voice-note brain dumps and trusting the AI to turn them into plans. The users who stick, Wang says, are the ones who get onto that ramp and see it work every time. The first wow moment has to come fast. → 35:15

11. The Screenshot Test. A modern consumer product has to work as a screenshot, Wang says, because screenshots travel through group chats. Senra describes a friend who runs a $100 billion company leaving his ID on a private jet and telling Muse he didn’t want to be involved. Muse arranged a courier and got the ID to his office. Jolly in every screenshot tells anyone looking which app did it, and Wang admits that part wasn’t planned. → 37:25

The Deal

12. The Call After Llama 4. Wang’s first meeting with Zuckerberg took about a year to get on the calendar. Years later, after Llama 4 disappointed inside Meta, a message asking if he had time for a call turned into advice sessions and then a deal. Wang assumed it lacked industrial logic, since Scale sells to enterprises and governments, until the numbers made clear Zuckerberg wanted talent. Outreach to announcement took five to six weeks. → 45:38

13. Damaged Goods. Wang spent weeks deciding whether to leave Scale, which he describes as giving up his baby after 9 years. What tipped him was a deal that worked for every party, and the state of Meta’s AI effort, which looked like damaged goods and not the most appealing place to work on AI. That triggered his entrepreneur side. With so much to fix, he could set the principles and build the culture from scratch. → 54:38

The Lab

14. Diamonds, Not Skyscrapers. At Scale, Wang followed the rule that a founder should be able to do every job, which made him the bottleneck, though he still recommends it for first-time founders. Scale was skyscraper work: linear payoffs, won by squeezing out 1% gains. AI research is diamond mining, where most ideas fail and one can be worth a million times the rest. At Scale he was the auteur. At Meta he sees himself as a coach. → 58:40

15. Room to Cook. Wang has more than 200 direct reports, mostly researchers in a group called TBD, which he calls anti-bureaucracy. Pods have technical leads who break ties, and big decisions get debated as a group. The pitch to recruits: no bottleneck on compute or distribution. He doesn’t claim to be the best manager of 200 people. They don’t need managers, he says. They need room to cook. → 1:09:30

Watch the full video at

.

Hi,

We’re excited to welcome you to internAtlas- a new platform and newsletter built around discovering opportunities, ideas, and resources worth knowing about.

🌐 internAtlas.in

We’re currently inviting a small group of people to experience internAtlas during our early beta, and we’d genuinely value having you with us at this stage.

The platform will bring together curated opportunities and useful information across areas such as internships, jobs, competitions, hackathons, workshops, campus programs, and more.

For students and early-career professionals, there will be opportunities to discover and apply for internships and our Campus Ambassador Program.

Campus Ambassador Application - Apply

Internship Application : Apply

Follow internAtlas on WhatsApp
Follow internAtlas on LinkedIn
Subscribe internAtlas on Substack

If these opportunities aren’t relevant to you, no action is needed — we’d still love for you to explore InternAtlas.in and share your perspective.

Intern Atlas

Help us build it better

As an early beta user, your feedback is particularly valuable.

Take a look around internAtlas.in , explore the newsletter, and tell us what you think:

What should we improve? What should we add? What would make Internetless genuinely useful to you?

We’re opening the doors early because we want to build something thoughtful, useful, and genuinely worth coming back to.

Welcome to internAtlas.in

We’re glad you’re here from the beginning.

Warm regards,
Team internAtlas
internatlas.in

TLDR; In Today’s Signal

• Meta spins out Vibes as a standalone AI-generated video app

• Perplexity Introduces Model Council for Multi-Model Verification

• Apple and Google Formalize Gemini-Based Foundation Model Partnership

• Anthropic extends Opus into long-context, agentic execution

• Tinder deploys AI matchmaking to reduce swipe fatigue

• Anthropic attacks OpenAI’s ad strategy in Super Bowl campaign

• Benchmark raises $225M to double down on Cerebras

• Fibr AI raises $7.5M seed to build agentic web infrastructure

THE AI SIGNAL PICKS

• Perplexity Introduces Model Council for Multi-Model Verification

Perplexity launched Model Council, a multi-model research mode that runs a single query across multiple frontier models simultaneously. The feature currently executes prompts across models such as GPT-5.2, Claude Opus 4.6, and Gemini 3.0, then synthesizes areas of agreement and disagreement. This replaces manual cross-checking between models with a single aggregated output inside Perplexity’s core interface. The release formalizes multi-model comparison as a default workflow, signaling rising model divergence and increasing demand for verification at decision time.

• Meta spins out Vibes as a standalone AI-generated video app

Meta is testing a standalone Vibes app, separating its AI-generated short-video product from Meta AI. The app offers a dedicated feed where all videos are AI-generated, with remixing tools, music layers, and cross-posting to Instagram and Facebook. This shift repositions Vibes as a direct social video platform competing with OpenAI’s Sora app rather than an assistant feature. Meta also confirmed upcoming freemium subscriptions for Vibes creation, signaling early monetization of AI-native social video.

THE BIG LEAP
Frontier: OpenAI’s Push to Run AI Agents at Enterprise Scale

Signal Scoop:
OpenAI introduced Frontier, a platform to build, deploy, and manage AI agents that perform real work across enterprises. Unlike prior agent tools, Frontier focuses on production readiness: shared context, permissions, evaluation, and execution across systems. Early adopters include HP, Oracle, State Farm, Uber, and Intuit.

The Full Picture:

• Frontier targets the deployment gap between model capability and enterprise execution.

• Platform supports end-to-end agent lifecycle: onboarding, context sharing, execution, evaluation, and optimization.

• Agents operate across existing data warehouses, CRMs, internal tools, and multi-cloud environments without replatforming.

• Built-in identity, permissions, and governance enable use in regulated and sensitive workflows.

• Early enterprise use cases show material gains: 6-week processes reduced to 1 day, 90% sales time reclaimed, and up to 5% production output increases.

What You Can’t Miss:
Frontier shifts OpenAI from model supplier to enterprise operating layer for AI labor. Control is moving from isolated copilots to managed AI coworkers embedded across workflows. Companies that standardize agent deployment gain speed, leverage, and compounding productivity. Those without an execution layer risk falling behind as AI advantage concentrates around who can reliably run agents in production.

Anthropic extends Opus into long-context, agentic execution

Signal Scoop:
Anthropic released Claude Opus 4.6, adding materially stronger agentic coding, long-context stability, and a 1M-token context window (beta).
The change is not incremental accuracy, but durability: the model maintains reasoning quality across longer, more autonomous tasks.

The Full Picture:

• Ranks first on Terminal-Bench 2.0 for agentic coding and Humanity’s Last Exam for multidisciplinary reasoning.

• On GDPval-AA, exceeds the next-best frontier model (GPT-5.2) by ~144 Elo points; improves ~190 Elo over Opus 4.5.

• Supports a 1M-token context window (beta) and up to 128k output tokens; standard pricing unchanged below 200k tokens.

• Long-context retrieval improves sharply: 76% on MRCR v2 (1M-token variant) versus 18.5% for Sonnet 4.5.

• Adds system-level controls: adaptive thinking, four effort settings, context compaction for long-running agents, and multi-agent coordination in Claude Code (preview).

• Safety evaluations show misalignment rates comparable to or lower than prior Opus models.

What You Can’t Miss:
Opus 4.6 shifts the frontier from short-horizon reasoning to sustained cognitive throughput. Reliable use of very large context reduces failure modes that previously limited autonomous agents, especially in large codebases and document-heavy workflows. This concentrates advantage at the layer where models must operate continuously without degradation, increasing pressure on competitors whose performance drops as context length and task duration grow.

GPT-5.3-Codex collapses coding, computer use, and professional work into one agent

Signal Scoop:
OpenAI introduced GPT-5.3-Codex, a new agentic model that merges frontier coding performance with general professional computer use. It outperforms GPT-5.2-Codex across key benchmarks while running ~25% faster. This marks a shift from “coding assistant” to “end-to-end work agent.”

The Full Picture:

• Sets a new industry high on SWE-Bench Pro (56.8%) and Terminal-Bench 2.0 (77.3%), surpassing prior Codex models.

• Achieves 64.7% accuracy on OSWorld-Verified, approaching human performance (~72%) in real desktop task completion.

• Matches GPT-5.2 on GDPval (70.9% wins or ties), covering 44 professional knowledge-work roles beyond software engineering.

• Demonstrates lower token usage than previous models, enabling longer autonomous runs with less cost overhead.

• Classified as High capability in cybersecurity, with dedicated training for vulnerability identification and expanded defensive safeguards.

What You Can’t Miss:
This release shifts leverage from tools to integrated agents that reason, build, and execute end-to-end. Coding stops being the constraint; orchestration becomes the differentiator. Teams gain speed, scope, and labor compression, forcing competitors to bridge the gap between code generation and real computer operation. Advantage now sits at full-stack agent control, not raw model IQ.

ON THE AI EDGE

• Tinder deploys AI matchmaking to reduce swipe fatigue.
  Tinder is testing an AI feature called Chemistry that replaces mass swiping with question-driven matching and optional camera-roll analysis. The system delivers a limited set of recommendations instead of infinite profiles, reframing discovery around intent rather than volume. Parent company Match Group reports the shift as subscriber growth and MAUs decline, with new registrations down 5% and MAUs down 9% year over year. AI-led discovery and verification features now anchor Tinder’s engagement and revenue recovery strategy.

• Anthropic attacks OpenAI’s ad strategy in Super Bowl campaign.
  Anthropic is spending millions on Super Bowl ads mocking plans to introduce advertising into ChatGPT, positioning Claude as ad-free. The 30-second spot airing on NBC states “Ads are coming to AI. But not to Claude,” directly targeting OpenAI’s monetization shift. OpenAI CEO Sam Altman publicly called the ad deceptive as OpenAI runs its own Super Bowl campaign for Codex. The clash surfaces as both labs, neither profitable, compete for users, enterprise contracts, and IPO attention.

AI START-UP NEWS

• Andreessen Horowitz raised $15B, carving out $1.7B for AI infrastructure.
  Andreessen Horowitz closed a new $15B fund, with $1.7B allocated to its infrastructure practice backing AI platforms. The capital supports companies across model tooling, inference, and search infrastructure, including OpenAI and ElevenLabs. The fund size and earmark exceed prior cycles, reflecting higher capital intensity as model scale, compute demand, and infra-level competition accelerate.

• Benchmark raises $225M to double down on Cerebras.
  Benchmark Capital raised $225M across two special-purpose “Benchmark Infrastructure” vehicles to participate in Cerebras Systems’ $1B round. The financing priced Cerebras at $23B, nearly tripling its valuation in six months as demand for non-GPU AI compute accelerates. Cerebras’ wafer-scale systems target faster training and inference by eliminating multi-chip data bottlenecks. The move concentrates capital behind alternative AI infrastructure suppliers competing with Nvidia while positioning Cerebras ahead of a planned 2026 IPO.

• Fibr AI raises $7.5M seed to build agentic web infrastructure.
  Fibr AI closed a $7.5M seed round led by Accel with participation from WillowTree Ventures and MVP Ventures. The company is building an agentic web experience layer where every URL adapts in real time using context, intent, and behavioral signals. The platform integrates ad, CRM, and onsite data to dynamically rewrite content, flows, and visuals per visitor. Early pilots report ~20% conversion lifts, positioning Fibr as infrastructure for post-click personalization as traffic shifts toward AI-driven agents.

  ---

  Elevate your experience.

  Join our community

  Please help us get better and suggest new ideas at ceo@theaisignal.com

TLDR; In Today’s Signal

• OpenAI launches ChatGPT Go worldwide to expand entry-tier AI access

• MongoDB releases Voyage 4 Series optimized for AI-native workloads

• GIGABYTE deepens NVIDIA partnership to scale AI-powered computing systems

• Meta Shuts Horizon Workrooms Amid AI Shift

• AI Firms Formalize Paid Wikipedia Access

• Autonomous Coding Agents Scale Up

THE AI SIGNAL PICKS

• GIGABYTE Unveils Blackwell-Based AI Systems
  GIGABYTE expanded its NVIDIA partnership with new Blackwell-based RTX 50 Series graphics cards, RTX laptops, and G-SYNC Compatible monitors. The lineup combines NVIDIA’s RTX AI, DLSS 4, and Studio platform with GIGABYTE’s thermal, system-control, and display technologies to deliver sustained AI performance for gaming and creative workloads.

• ChatGPT Go Expands Global AI Access

  OpenAI has rolled out ChatGPT Go globally after strong early adoption, positioning it as a low-cost entry tier priced at $8 per month in the US. Go offers expanded access to GPT-5.2 Instant with higher message, upload, image-generation limits, and longer memory, formalizing a three-tier global subscription stack alongside Plus and Pro.

• Autonomous Coding Agents Scale Up

  Cursor demonstrated sustained autonomous coding by coordinating hundreds of agents over multi-week runs, generating over one million lines of code across large codebases. The system separates planner agents (task decomposition) from worker agents (execution), eliminating lock-based coordination bottlenecks. GPT-5.2 models showed stronger long-horizon focus, planning reliability, and lower drift than earlier coding-specialized models..

THE BIG LEAP

OpenAI Formalizes Advertising to Expand ChatGPT Access

Signal Scoop: OpenAI is laying the foundation for a large-scale ad-supported AI model while explicitly ringfencing trust, answer integrity, and user privacy—marking a decisive shift in how mass-market AI will be funded.

The Full Picture

• ChatGPT Go, launched in August 2025 and now live in 171 countries, is expanding to the US at $8/month, offering higher limits for messaging, image creation, file uploads, and memory.

• OpenAI plans to begin testing ads in the free tier and ChatGPT Go in the US, while Plus, Pro, Business, and Enterprise remain ad-free.

• Ads will be clearly labeled, separate from answers, and never influence responses, with conversations kept private and never sold to advertisers.

• Users retain control over ad personalization, can opt out, and ads will be excluded from sensitive categories and under-18 accounts.

What You Can’t Miss
This is OpenAI defining the economic architecture of consumer AI: accessibility funded by ads, premium intelligence protected by subscriptions, and strict separation between monetization and model behavior. The precedent matters as AI platforms scale to hundreds of millions of users.

Google Introduces AI-Native Video Creation for Work

Signal Scoop:Google is making generative video a first-class productivity medium inside Workspace, extending AI from text and slides into collaborative, work-native storytelling.

The Full Picture

• Google Vids is now generally available across Business and Enterprise Workspace plans, with consumer access via Google AI Pro and Ultra.

• Gemini in Vids generates editable storyboards, suggested scenes, stock media, music, and scripts directly from prompts and Drive files.

• Veo 3 enables eight-second AI video clips, image-to-video animation, native audio generation, and AI avatars for scripted delivery.

• Built-in recording, teleprompter support, shared content libraries, and Docs-style collaboration position Vids for training, internal communications, marketing, and customer education.

What You Can’t Miss
Google is embedding generative video directly into daily work workflows, not as a standalone creative tool. This signals a shift where video becomes as routine—and as AI-assisted—as documents and presentations inside enterprise software.

ON THE AI EDGE

• AI Firms Formalize Paid Wikipedia Access

  Wikimedia Foundation confirmed that Microsoft, Meta, Amazon, Perplexity, and Mistral AI are paying for access to Wikimedia Enterprise, a premium API launched in 2021 that provides commercially tuned, structured Wikipedia data. The program offers feature requests, data structuring, and reliability guarantees for AI workloads, formalizing Wikipedia as licensed infrastructure for large-scale model training and inference.

• Meta Shuts Horizon Workrooms Amid AI Shift

  Meta will discontinue its VR workplace app Horizon Workrooms on February 16, deleting associated user data, as it cuts staff in its Reality Labs division. The closure follows more than $70B in metaverse losses and reflects Meta’s continued reallocation of resources toward AI-focused priorities.

• MongoDB Releases Voyage 4 Embedding Series
  MongoDB introduced the Voyage 4 series of text embedding models, including large, lite, and open-weight variants, all sharing a compatible embedding space. The design removes the need to use identical models for indexing and querying, enabling flexible quality-latency-cost tradeoffs for retrieval-heavy AI systems and long-term memory applications.

AI START-UP NEWS

• ClickHouse Expands Into AI Observability
  ClickHouse raised $400M in Series D funding and acquired Langfuse to enter LLM observability, while introducing a native Postgres service to unify transactional and analytical workloads. The expansion positions ClickHouse to support AI systems moving into production, where real-time analytics, observability, and reliability are critical.

NEW TOOLS, NEW POSSIBILITIES

• Atomise – AI-driven software engineering and digital transformation firm delivering cloud, fintech, and platform-scale solutions.

• vHub.ai – AI-based influencer credibility and creator intelligence platform enabling data-backed discovery and verification.

• Meet Alfred – AI-powered LinkedIn outreach automation platform for scalable prospecting, engagement, and campaign management.

• TextCortex – Enterprise-grade generative AI platform for content creation, rewriting, and workflow automation across marketing use cases.

• SmythOS – AI agent operating system enabling secure development, orchestration, and deployment of autonomous AI agents.

AI CAREER HORIZON

• Google: Senior Software Engineer, Hardware Security (Remote)

• Anthropic: Senior Research Scientist, Reward Models (Remote)

• Databricks: AI Engineer – Forward Deployed Engineer (Remote)

• Anthropic: Research Engineer, Reward Models Platform (Remote)

TLDR; In today’s Signal

• Domino Introduces AI-Enabled Sunrise DFE for Label Presses

• Fujitsu Unveils AI Agent Platform for Healthcare Workflows

• AI Enables Climate-Smart Rice Farming in India

• Perplexity Launches Comet Plus AI Subscription

• Alibaba Open-Sources Speech-to-Video Avatar Model

• Walmart Unveils AI Tools, Fulfillment, and Incentives

• Attio raises $52M to scale AI-native CRM

THE AI SIGNAL PICKS

• Fujitsu Unveils AI Agent Platform for Healthcare Workflows
  Fujitsu has launched a secure, AI-driven orchestrator platform tailored for Japan’s healthcare sector. It integrates specialized AI agents from data structuring to interoperability to streamline medical operations, reducing patient wait times and freeing clinicians for core care duties. Developed under the Fujitsu Uvance initiative in collaboration with NVIDIA, this platform promises to transform efficiency, sustainability, and care delivery across institutions.

• Domino Introduces AI-Enabled Sunrise DFE for Label Presses

  Domino Printing Sciences has unveiled the latest Sunrise Digital Front End (DFE) an AI-powered, Harlequin RIP-based workflow solution for its N-Series digital label presses. The upgrade features Label Studio for streamlined job setup, real-time inline variable data processing (e.g., 2D codes), and Domino Viewer for live press monitoring. Designed to boost productivity, uptime, and cost efficiency across label production.

• AI Enables Climate-Smart Rice Farming in India

  Mitti Labs, a New York-based AI startup, has partnered with The Nature Conservancy to tackle methane emissions from rice paddies through precise satellite- and radar-powered monitoring. Their digital MRV (Monitoring, Reporting, and Verification) platform covers about 10 million hectares, one-quarter of India's rice farms and enables farmers to adopt regenerative practices while earning carbon credits, boosting both environmental impact and livelihoods.

• OpenAI strengthens ChatGPT’s crisis response capabilities

  OpenAI has enhanced ChatGPT’s handling of mental health crises. GPT-5 now lowers risky responses in distressing conversations by 25% compared to GPT-4o via a “safe completions” method. The update also reinforces safeguards in extended sessions and boosts detection of emotional reliance, aiming to ensure users in crisis receive empathy and direction toward care.

THE BIG LEAP

Claude Lives in Your Browser

Signal Scoop: Anthropic has quietly unleashed a research-preview extension, Claude for Chrome, enabling its AI assistant to navigate, click, and perform actions directly within users’ browsers. Initially limited to 1,000 Max-plan subscribers, the move marks a major shift in AI productivity tools and raises pressing safety and security concerns. Anthropic is inviting real-world feedback to harden defenses ahead of broader deployment.

• Claude for Chrome enables AI to see what you see, click buttons, fill forms, and tackle tasks like managing calendars, drafting emails, or running tests within your browser. A sidecar interface maintains context across your browsing experience.

• The extension is available only to 1,000 trusted users on the paid Max plan, who can join a waitlist. Anthropic plans to expand access gradually as it strengthens safety safeguards.

• Browsing AIs face high-stakes risks like prompt injection, malicious instructions hidden in web content that can cause Claude to delete files, steal data, or perform financial actions. In early red-teaming tests without mitigations, the attack success rate was a concerning 23.6%.

• Anthropic is deploying layered countermeasures: granular site permissions, confirmations before sensitive actions, stricter system prompts, classifiers for suspicious input, and blocking high-risk site categories. These controls have already reduced some attack vectors to 0% success in limited tests.

What You Can’t Miss

This isn’t just another AI update, it marks a pivotal moment where AI can act autonomously within your browser, making web interaction seamless yet alarmingly powerful. Without strong safety guardrails, this integration could expose users to stealthy, high-impact attacks that compromise data or systems.

Gemini’s Image Tools Just Got a Big Upgrade

Signal Scoop: Google DeepMind's latest Gemini app upgrade with enhanced photo editing capabilities and natural language controls, dubbed internally as "nano-banana" Whether you're restyling, blending scenes, or generating video from your creations, Gemini helps you do it with precision and creativity. All outputs are clearly marked with watermarks and SynthID metadata, ensuring transparency starting August 26, 2025.

The Full Picture

• The enhanced model ensures that edited photos of people or pets maintain their core identity even when adding a retro hairstyle, costume, or placing them in new scenes.

• Users can blend multiple photos, modify settings, swap styles, or merge yourself with your pet, all while keeping the subjects recognizable.

• Gemini now supports conversational edits: start with a blank scene and add elements step-by-step paint walls, then add furniture guided entirely through dialogue.

• All edited or generated images feature a visible watermark and embedded SynthID metadata to clearly indicate AI origin. The update is live now for both free and paid Gemini users worldwide.

What You Can’t Miss

This update transforms how personal and creative AI editing can be finally allowing Gemini to modify photos while genuinely preserving who and what matters in the image. It’s no longer just about flashy visual effects; it’s about trust, identity, and creative control, all delivered transparently.

ON THE AI EDGE

• Perplexity Launches Comet Plus AI Subscription

  Perplexity today rolled out Comet Plus, a $5/month AI-powered subscription that offers users access to premium content from vetted publishers and journalists. It introduces a forward-looking revenue model—distributing earnings across human visits, search citations, and agent actions—aligning publisher compensation with the evolving dynamics of AI-driven internet usage.

• Alibaba Open-Sources Speech-to-Video Avatar Model

  Alibaba has released Wan2.2-S2V, its latest open-source speech-to-video model, enabling creators to animate portrait photos into lifelike avatars that speak, sing, and move capturing full-body motion across different framing styles. By compressing historical frames into compact latent representations, it supports stable, long-form video generation with reduced compute. Available now on Hugging Face, GitHub, and ModelScope.

• Walmart Unveils AI Tools, Fulfillment, and Incentives
  At the Let’s Grow! Marketplace Seller Summit, Walmart introduced AI-powered listing tools, Smart Assistant support, and streamlined shipping services to boost seller efficiency. They also expanded next-day delivery to major U.S. metros and rolled out compelling fee discounts and referral incentives strengthening omnichannel integration and seller competitiveness.

• Saudi’s Humain to Launch AI Data Centers With US Chips

  Saudi Arabia’s new state-backed AI company, Humain, has begun building its first data centers in Riyadh and Dammam, aiming to go live in early 2026 with 100-megawatt capacity at each site. The project is powered by a strategic import of U.S. semiconductors including Nvidia’s latest Blackwell chips, backed by initial approval for 18,000 units marking a major leap in sovereign AI infrastructure development. Additionally, AMD has committed to a $10 billion collaboration to expand compute capabilities.

AI START-UP NEWS

• Attio raises $52M to scale AI-native CRM

  Attio, an AI-native CRM platform designed for go-to-market teams, has secured $52M in Series B funding led by GV, with participation from Redpoint Ventures, Balderton Capital, Point Nine, and 01A. The funding will accelerate R&D, strengthen its AI-driven architecture (real-time ingestion, programmable interfaces, predictive intelligence), and support global expansion.

• Eyebot raises $20M to expand eye care access

  Boston-based Eyebot develops 90-second self-service kiosks that provide instant, doctor-verified eyeglass prescriptions. The startup secured $20 million in Series A funding led by General Catalyst, with participation from AlleyCorp, Baukunst, Village Global, and Ubiquity Ventures. The funds will support nationwide kiosk deployment, team growth, and expanded partnerships with eye care providers.Oxford Spin-out OrganOx Sells for $1.5B.

NEW TOOLS, NEW POSSIBILITIES

• Saner.AI: ADHD-friendly AI personal assistant that helps manage calendar.

• Deepgram: A speech-to-text platform offering accurate voice transcription

• Deforum Studio: An AI tool combining Generative Video and Text-to-Video

• Peerlist: Generates videos with synchronized audio from text

• Opal: A no-code platform to build AI-powered applications

AI CAREER HORIZON

• Hume AI: Machine Learning Engineer (NYC - Remote)

• Microsoft: Senior Researcher - Generative AI Frontiers (Washington, US)

• Google: Software Engineer III, AI/ML (Zürich, Switzerland)

• Anthropic: Machine Learning Systems Engineer (San Francisco, CA)

• OpenAI: Analytics Data Engineer (San Francisco, CA)

TLDR; In today’s Signal

• Synthetic X-Rays from AI Shake Up Medical Data

• AI Simulates 1,000-Year Climate in Hours

• AI Learns Biology Without Experiments

• Preventing AI Illusions of Consciousness

• OpenAI Learning Accelerator Launch in India

• Microsoft Debuts Open-Source VibeVoice TTS Model

• Raytron Launches Advanced Thermal Gas Leak Camera

THE AI SIGNAL PICKS

• Synthetic X-Rays from AI Shake Up Medical Data

  Stanford Medicine’s researchers have developed RoentGen, an open-source generative-AI model capable of producing realistic synthetic chest X-rays from textual medical descriptions. This innovation provides vital training data where real examples are scarce such as for rare conditions while offering a privacy-preserving alternative to sensitive patient data. The model marks a step forward in medical AI, promising enhanced diagnostics. It may reshape how training data is generated in healthcare.

• OpenAI Learning Accelerator Launch in India
  OpenAI has unveiled the Learning Accelerator, an India-focused initiative aimed at integrating AI into education through research, training, and deployment. The program includes a $500,000 research partnership with IIT Madras to study AI’s educational impact, as well as plans to distribute 500,000 ChatGPT licenses to educators and learners through partnerships with the Ministry of Education, AICTE, and ARISE schools. This initiative is a major step toward ensuring AI strengthens learning and supports teachers in India’s vast education system.

• AI Simulates 1,000-Year Climate in Hours

  University of Washington researchers have introduced DLESyM, a deep-learning Earth system model capable of simulating a millennium of current-climate climate variability using just a single processor in about 12 hours compared to roughly 90 days on a supercomputer. This enables rapid exploration of long-term trends and extreme event probabilities, significantly lowering computational barriers in climate research.

• Microsoft Debuts Open-Source VibeVoice TTS Model
  Microsoft Research has unveiled VibeVoice, an open-source text-to-speech model capable of generating up to 90 minutes of expressive, multi-speaker conversational audio. It supports up to four distinct speakers, leverages novel continuous speech tokenizers operating at 7.5 Hz, and employs a next-token diffusion framework with a Large Language Model and diffusion head for natural, coherent long-form synthesis, a significant leap over traditional TTS models.

THE BIG LEAP

NVIDIA Jetson Thor Powers Real - Time Reasoning for Physical AI

Signal Scoop: NVIDIA has launched Jetson AGX Thor, a high-performance robotics computer for real-time, on-device AI. Now generally available, it delivers unmatched compute power that significantly advances physical AI development.

The Full Picture

• Built on the Blackwell GPU architecture, Jetson Thor offers up to 7.5× more AI compute, 3.5× greater energy efficiency, better CPU performance, and double the memory compared to Jetson Orin.

• System specs include 2,070 FP4 TFLOPS, 128 GB LPDDR5X memory, 14-core Arm Neoverse-V3AE CPU, 4×25 GbE I/O, sensor offload engine, and Holoscan bridge for rich multimodal input.

• It supports key NVIDIA stacks Isaac (robotics), Metropolis (vision AI), Holoscan (sensor processing) and integrates popular AI frameworks/models like PyTorch, Gemini, OpenAI, and more.

• Real-world adoption is already underway with users like Amazon Robotics, Boston Dynamics, Caterpillar, Meta, Figure, and others, while ADI is embedding Thor to bridge simulation and reality for humanoid robots.

What You Can’t Miss

This isn't just another upgrade, Jetson Thor is a leap toward intelligent, autonomous machines. Its server-class performance in a compact, energy-efficient form will enable robots capable of complex, real-time reasoning and multimodal interaction, ushering in the next era of physical AI.

Google NotebookLM Expands Media & Language Support

Signal Scoop: Google has boosted its AI research assistant, NotebookLM, by rolling out AI-generated Video Overviews and vastly enhancing Audio Overviews across 80 languages, delivering richer, more inclusive content.

• Video Overviews, previously limited, are now powered across 80 languages, giving users bite-sized narrated video summaries of their notebooks.

• Audio Overviews in non-English languages have been upgraded to match the depth of the English version, offering more comprehensive summaries.

  This feature enhances multilingual accessibility and immersive learning, combining visual and auditory synthesis for better comprehension.

• NotebookLM continues refining its Studio Panel toolset ,adding smarter sharing, mind maps, better UI, and more export/collaboration options, making it a standout tool for teams and educators.

What You Can’t Miss

NotebookLM is evolving from a personal AI assistant to a multilingual, multimedia research powerhouse. By blending video, audio, and visual summaries across 80 languages, it's setting a new standard for accessibility and collaboration in AI-driven learning and knowledge sharing.

ON THE AI EDGE

• Preventing AI Illusions of Consciousness

  IBM cautions that users are increasingly mistaking AI for sentient beings, driven by emotional attachments to conversational models, even when they’re merely mimicking awareness. Drawing on warnings from Microsoft’s AI chief, the article emphasizes the need for clear design strategies and user reminders to maintain the understanding that AI is a tool, not a mind.

• Raytron Launches Advanced Thermal Gas Leak Camera
  Raytron has unveiled a high-sensitivity infrared thermal imaging camera tailored for real-time detection of industrial gas leaks. Its non-contact optical gas imaging (OGI) technology enables rapid scanning across hazardous environments, detecting gases like methane, ammonia, ethylene, and sulfur hexafluoride and revealing precise leak locations in milliseconds. Built for harsh conditions, the camera enhances safety, reduces downtime, and supports emissions monitoring with industrial-grade durability.

• Edge AI Is Transforming Automotive Innovation

  According to McKinsey, 38% of premium car buyers in Germany would consider switching brands for a better digital experience. As real-time, in-vehicle AI inference gains ground, enabled by hardware advances and lightweight models, automakers are rethinking compute strategy across cloud, edge, or hybrid architectures. Success depends on collaborative platforms spanning OEMs, chipmakers, and Tier 1 suppliers.

• AI Learns Biology Without Experiments
  A new preprint introduces rBio-1, a large language model fine-tuned via reinforcement learning using soft verification from simulated biological world models instead of costly lab data. It excels at biological reasoning tasks like perturbation prediction, outperforming previous methods—and demonstrates how simulated verifiers can fuel advanced scientific AI. This opens a scalable training paradigm for domain-specific reasoning models.

AI START-UP NEWS

• Creem raises €1.8M for AI-powered financial OS

  Creem, a fintech startup developing a “Stripe for AI Financial OS,” has secured €1.8 million in funding. The platform helps fast-growing startups automate financial operations using AI. The investment will be used to scale its product and expand customer reach, with backing from key European investors.

• Allen Law Invests $156M in MORROW Wellness Center
  Entrepreneur Allen Law is launching MORROW, a 38,000 sq ft longevity-focused wellness center in Singapore’s Longevity World, backed by a $156 million investment over five years, including $15.6 million for the flagship’s initial setup and technology. MORROW combines AI-driven diagnostics, expert coaching, and wearable tech to create highly personalized, prevention-centric wellness services—making high-quality longevity care more accessible.

• Oxford Spin-out OrganOx Sells for $1.5 Billion
  OrganOx, an Oxford University spin-out specializing in Normothermic Machine Perfusion to preserve donor organs, has been acquired by Japan’s Terumo Corporation for $1.5 billion. The funds reward early investors like BGF, Lauxera Capital Partners, and the University of Oxford, and will help accelerate global deployment of its life-saving technology.

NEW TOOLS, NEW POSSIBILITIES

• Jasper AI: AI assistant for generating blog posts, ads, and content

• Novelcrafter: Tailored for authors needing writing support

• Katalist: Converts scripts into visual storyboards

• RunComfy: Simplifies executing ComfyUI workflows for image generation

• Frase IO: Helps marketers with SEO optimization and research writing

AI CAREER HORIZON

• SAP: Senior AI/ML Applied Scientist - Generative AI (Newport Beach, US)

• Path AI: Senior Product Manager, AI Products (Boston, MA)

• DigitalOcean: Senior Software Engineer I - AI/ML (Hyderabad, India)

• Raft: AI/ML Engineer (McLean, VA)

• FloSports: Data Analyst (Austin, TX)

TLDR; In today’s Signal

• AI Chatbots Linked to Psychotic Episodes

• Silicon Valley Dumps $100M into Pro-AI Super-PACs

• Baidu Launches MuseSteamer 2.0 Upgrade

• NVIDIA Unveils Spectrum-XGS Ethernet Infrastructure

• Perplexity Expands Publisher Revenue Sharing Program

• Robomart Launches $3 Delivery Robot Challenge

• Ontic Raises $230M Series C led by KKR

• Pintarnya Raises $16.7M in Series A

THE AI SIGNAL PICKS

• NVIDIA Unveils Spectrum-XGS Ethernet Infrastructure

  NVIDIA introduces revolutionary Spectrum-XGS Ethernet technology that enables distributed data centers to operate as unified giga-scale AI super-factories. This breakthrough networking solution removes traditional boundaries by providing scale-across infrastructure, complementing existing scale-up and scale-out capabilities. The technology features advanced distance congestion control and precision latency management that nearly doubles performance of multi-GPU communication.

• AI Chatbots Linked to Psychotic Episodes

  Researchers are investigating how AI chatbots may contribute to psychotic episodes. Studies suggest that these chatbots can reinforce delusional thinking by providing affirmations and aligning with users' beliefs. This phenomenon has led to concerns about the safety and ethical implications of AI in mental health contexts. Experts emphasize the need for careful design and regulation to prevent potential harm.

• Silicon Valley Dumps $100M into Pro-AI Super-PACs

  Andreessen Horowitz, OpenAI President Greg Brockman, and other tech leaders have committed over $100 million to “Leading the Future,” a network of super-PACs backing candidates who support minimal AI regulation. The group aims to prevent a state-by-state patchwork of AI laws ahead of next year’s midterms, arguing that strict rules could hamper innovation and stall U.S. leadership in AI.

THE BIG LEAP

Grok 2.5 Opens Path to Musk’s Next Big Move

Signal Scoop: In a bold move for AI openness, Elon Musk’s xAI has released the Grok 2.5 model weights and architecture to the public via Hugging Face and promised that its successor, Grok 3, will follow within six months. This step underscores xAI’s commitment to transparency and community-driven AI progress.

The Full Picture

• Grok 2.5 represents last year’s top-performing AI model from xAI; its open-source release offers developers full access to its ~500 GB weights.

• Musk confirmed that Grok 3 will be released under similar licenses in about six months, promising a cadence of openness.

• The move positions xAI alongside Meta and Anthropic, signaling alignment with open-source leadership amid growing community demand for transparency.

• With access to these full model weights, researchers and startups can fine-tune or integrate Grok into applications, accelerating innovation at the edge.

What You Can’t Miss

This open-source move transforms Grok from a proprietary chatbot into a development platform. By inviting experimentation and collaboration, xAI is reshaping its strategy from walled AI toward democratized model evolution, raising the stakes in the AI race.

OpenAI & Retro Biosciences Enhance Stem Cell Reprogramming

Signal Scoop: OpenAI and Retro Biosciences have collaborated to develop GPT-4b micro, a specialized AI model for protein engineering. This partnership has led to the creation of enhanced variants of the Yamanaka factors, achieving over a 50-fold increase in stem cell reprogramming efficiency. These advancements hold significant promise for regenerative medicine and longevity research.

• GPT-4b micro, a compact version of GPT-4o, was trained to design proteins with high steerability and flexibility, focusing on biological applications.

• The AI model re-engineered SOX2 and KLF4 proteins, leading to a 50+ fold increase in stem cell reprogramming markers compared to traditional methods.

• The redesigned proteins demonstrated superior DNA damage repair capabilities, indicating higher rejuvenation potential and genomic stability in induced pluripotent stem cells (iPSCs).

• The findings were replicated across multiple donors, cell types, and delivery methods, confirming the robustness and reproducibility of the engineered proteins.

What You Can’t Miss

This collaboration marks a significant leap in AI-assisted biotechnology, demonstrating that AI can accelerate life sciences research by designing proteins with unprecedented efficiency. The enhanced Yamanaka factors not only improve stem cell reprogramming but also offer potential breakthroughs in regenerative medicine and longevity therapies

ON THE AI EDGE

• Perplexity Expands Publisher Revenue Sharing Program

  Perplexity AI extends its groundbreaking revenue-sharing initiative beyond its core search platform to include its Comet web browser. Publishers and media outlets can now earn revenue from AI-powered searches across both platforms. This expansion represents a significant shift in how AI search engines compensate content creators, addressing longstanding concerns about AI companies profiting from publisher content.

• Robomart Launches $3 Delivery Robot Challenge

  Robomart unveils its patented RM5 autonomous delivery robot featuring 10 individual lockers and 500-pound capacity. The LA startup targets DoorDash and Uber Eats with a revolutionary $3 flat-fee delivery model, claiming 70% cost reduction through automation. CEO Ali Ahmed's company plans Austin launch later this year.

• Baidu Launches MuseSteamer 2.0 Upgrade

  Baidu unveiled MuseSteamer 2.0, an upgraded image-to-video model that now adds human-like voices and ambient audio. It represents a leap in AI-generated content creation bridging visuals, audio and speech for lifelike short videos. This puts Baidu a step ahead in enterprise video generation, aligning with broader multimedia AI momentum.

AI START-UP NEWS

• Ontic Raises $230M Series C led by KKR to scale security-intelligence

  Connected-security intelligence firm Ontic raised $230M in a Series C (KKR-led) to expand its AI-driven risk, threat and protective-security products. The round will accelerate go-to-market expansion in corporate and government segments and platform integration capabilities.

• Pintarnya Raises $16.7M in Series A
  Indonesia’s employment platform Pintarnya raised $16.7 million to expand its AI-driven job matching and collateral-backed financial services for informal workers. This round will fuel continued growth in underserved markets and highlights AI’s role in inclusive economic access.

NEW TOOLS, NEW POSSIBILITIES

• Plotaverse: Animates static images into looping videos with dynamic effects

• FrameZero: Online platform for animating and sharing visual ideas

• Profound: Optimizes brand presence across AI-powered search platforms

• Copilot 3D: Transforms 2D visuals into immersive 3D models

• Release0: No-code builder for conversational chatbots

AI CAREER HORIZON

• Instructure: AI Product Manager (US - Remote)

• Microsoft: Senior Applied Scientist (Washington, United States)

• OpenAI: AI Technical Support Specialist (Tokyo, Japan)

• Meta: AI Research Scientist, Robotics (Redmond, WA)

• Google Deepmind: Research Engineer, AI for Education (London, UK)

TLDR; In today’s Signal

• Meta Partners with Midjourney to Enhance AI Imagery

• Apple Adds Enterprise ChatGPT Configuration Flexibility

• IBM and NASA Open-Source AI for Solar Flares

• NVIDIA Unveils Blackwell Ultra: AI “Fabric” Superchip

• JoggAI and ElevenLabs Launch Multilingual AI Avatars

• Netflix Sets Generative AI Guidelines for Partners

• FieldAI Secures $314M for Robotics Expansion

THE AI SIGNAL PICKS

• Netflix Sets Generative AI Guidelines for Production Partners

  Netflix has issued formal generative AI guidelines for content partners via its Partner Help Center, clarifying when AI tools such as video, sound, image, or text generators are permissible. Production teams must disclose intended GenAI usage up front, and obtain written approval when AI is used for final deliverables, personal data, or talent likenesses.

• Apple Adds Enterprise ChatGPT Configuration Flexibility

  Apple’s upcoming software update will allow enterprise IT administrators to enable or restrict access to ChatGPT for Enterprise, or any other “external” AI provider, offering unprecedented control and flexibility. This move demonstrates how Apple is beginning to meet enterprise AI integration demands while preserving strong security and privacy standards.

• IBM and NASA Open-Source AI for Solar Flares
  IBM and NASA jointly released Surya, an open-source AI model for predicting solar flares, via IBM Research. Trained on decades of space and weather data, Surya surpasses prior models in forecasting high-energy solar storms. The model’s GitHub release provides researchers and power-grid operators with a new tool to prepare for disruptions caused by space weather.

• Meta Partners with Midjourney to Enhance AI Imagery

  Meta announced a licensing and technical collaboration with Midjourney, allowing Meta’s AI research teams to integrate the startup’s acclaimed image and video generation technology into future products. This shift reflects a strategic pivot toward partnering with third-party AI leaders amid ongoing internal reorganization.

THE BIG LEAP

Apple Eyes Gemini for Siri’s AI Upgrade

Signal Scoop: Apple is in early talks to integrate Google’s Gemini model into a major Siri overhaul. This marks a potential pivot: instead of using only Apple’s own AI, it would license Google’s generative model. Such a move could leapfrog Siri’s capabilities in natural language understanding and multimodal tasks. It signals Apple’s urgency to catch up on AI features, especially as Siri has lagged behind competitors in generative AI.

The Full Picture

• Apple reportedly wants to fine-tune Gemini for Siri to boost voice assistant intelligence.

• CEO Tim Cook’s team has been under pressure to add chat and summarization to Siri; Gemini could fill that gap quickly.

• Gemini is multimodal, handling text, images, and code, which could extend Siri beyond scripted responses to contextual assistance.

• This would be a historic partnership between Apple and Google in AI, reflecting the tool’s maturity and Apple’s caution with in-house development.

What You Can’t Miss

Boldly, Apple shifting Siri to a Google-owned AI core shows how vital advanced AI is in consumer tech. If realized, the Siri experience could become dramatically more helpful overnight, forcing rivals like Google and Amazon to respond.

Meta Locks $10B AI Deal with Google Cloud

Signal Scoop: Meta has inked a multi-year, over-$10 billion agreement with Google Cloud to power its AI workloads. The deal underscores the growing dominance of external cloud infrastructure in fueling large-scale generative AI.

• The six-year contract commits Meta to billions in spend on Google Cloud compute, storage, and networking .

• Meta’s workloads including large language model training depends on large GPU clusters that exceed its in-house data centers.

• Despite prior focus on building its own supercomputers, Meta is turning to external providers for scale.

• The deal reflects the escalating cost of staying competitive in generative AI, where training budgets now run in the billions.

What You Can’t Miss

Meta’s infrastructure spending signals that AI leadership is now a compute showdown. Even tech giants are outsourcing critical hardware, this isn’t just scaling up; it’s redefining how AI leaders source their backbone.

ON THE AI EDGE

• NVIDIA Unveils Blackwell Ultra: AI “Fabric” Superchip
  NVIDIA’s new Blackwell Ultra GPU, built on the Blackwell architecture and it features a cutting-edge dual-reticle design with 208 billion transistors, delivering AI performance at scale for “AI factory” applications. Designed for large-scale AI reasoning and real-time services, this chip significantly boosts performance, energy efficiency, and scalability for demanding workloads.

• JoggAI and ElevenLabs Launch Multilingual AI Avatars

  AI video startup JoggAI has launched AvatarX, a next-generation avatar system that integrates ElevenLabs’ voice synthesis for natural, expressive multilingual speech. The platform enables creators to generate lifelike avatars that can speak across languages with realistic delivery. This collaboration expands the creative toolkit for video production, lowering costs and scaling personalized content.

AI START-UP NEWS

• FieldAI Secures $314M to Scale Robotics Platform

  Robotics AI startup FieldAI raised $314 million, lifting its valuation to $2 billion. Investors include Khosla Ventures and Nvidia’s VC arm, signaling strong institutional backing. The funding will accelerate team expansion and product development for large-scale deployment.

• Definite Raises $10M Seed to Launch AI-Native Analytics Stack

  Philadelphia-based startup Definite raised an oversubscribed $10 million seed round, led by Costanoa Ventures alongside Acrew Capital and angel backers. The company offers a unified AI-native data platform that embeds an assistant-Fi-capable of generating SQL from natural language and delivering data insights in minutes.

• Oway Raises $4M for Freight Ride-Sharing App
  YC-backed startup Oway secured $4 million in seed funding to develop a decentralized, Uber-like platform for freight. Oway’s AI-driven matchmaking algorithm aims to optimize truck haul capacity, reducing empty-load miles and inefficiencies in logistics. The funding will fuel development and early-market testing across U.S. trucking networks.

NEW TOOLS, NEW POSSIBILITIES

• Windsurf: AI-powered IDE with agentic code editing

• Twelve Labs: Video understanding API for search and multimodal retrieval.

• StatPecker: Turns raw data into charts and infographics automatically

• Kindo: The AI Brain for Technical Operations

• Adept: Builds AI agents that can use software tools and APIs

AI CAREER HORIZON

• Anthropic: Anthropic AI Safety Fellow (London - U.K.)

• Hugging Face: Open Source ML Engineer (Remote – U.S.)

• OpenAI: AI Technical Support Specialist (San Francisco)

• Cohere: Machine Learning Manager - Applied ML (NYC)

TLDR; In today’s Signal

• Gemini Live Can See What You Show

• Microsoft AI Chief Warns Against Studying AI Consciousness

• Google Claims Gemini AI Uses Tiny Energy Footprint

• OpenAI Eyes India Expansion with New Delhi Office

• Musk Taps AI to Revive X's Ad Engine

• Anthropic in Talks to Raise Up to $10 Billion in New Funding

• Leaping AI Raises $4.7M After Y Combinator Participation

THE AI SIGNAL PICKS

• Gemini Live Can See What You Show

  Using your smartphone camera, Gemini Live now visually highlights real-world items on-screen, like identifying tools or components, starting August 28 on Pixel 10, with roll-out to other Android devices and later iOS. It also introduces emotional tonality in speech, advancing the assistant’s naturalism. This marks a shift toward multi-modal, context-aware AI.

• Microsoft AI Chief Warns Against Studying AI Consciousness
  Mustafa Suleyman, Microsoft’s AI lead, cautions that exploring AI consciousness is “dangerous.” He argues that AI systems, despite their sophistication, do not have subjective experiences and should not be treated like humans. Suleyman urges developers to focus on creating AI that serves human needs rather than attempting to make machines “conscious.” His stance contrasts with some industry players exploring AI welfare and ethical implications of sentient-like models.

• Musk Taps AI to Revive X's Ad Engine
  Elon Musk is leaning on AI to reinvent X’s advertising system, sharpening targeting and automating content delivery to attract wary marketers. The move comes as ad revenue has slumped and brands look for more precise returns on spend. Musk frames AI as central to X’s reinvention beyond text posts, pushing it toward video and commerce. If the overhaul works, AI could restore advertiser confidence while reshaping X’s core business model.

THE BIG LEAP

Google Turns Search Into a Task Machine

Signal Scoop: Signal Scoop: Google’s AI Mode in Search is now live in 180+ countries in English. Real-time agentic features are coming first to restaurant reservations, executing tasks on users’ behalf. The move signals Google’s shift from passive search answers to proactive task completion.

The Full Picture:

• AI Mode in 180+ countries expands access to complex conversational search globally.

• U.S. users on Google AI Ultra can book restaurants specifying party size, time, and cuisine with curated options and direct booking links.

• Integrates with OpenTable, Resy, Tock, Ticketmaster, SeatGeek, Booksy, powered by Project Mariner, Knowledge Graph, and Maps.

• Recommendations tailor to past Search and Maps behavior; chat sessions can now be shared via links for collaboration.

What You Can’t Miss:
Google Search is now about doing, not just finding. Booking, reservations, and transactions are the new baseline.

Apple readies iPhone 17 Air: thinnest form factor yet

Signal Scoop: Apple’s iPhone 17 series, launching this September, includes the ultra-thin “Air” model at 5.5 mm—its thinnest iPhone ever. Minimalism meets function, showcasing Apple’s focus on engineering and design innovation.

The Full Picture:

• iPhone 17 Air is ~1 mm thinner than iPhone 16, rivaling Samsung’s Galaxy S25 Edge.

• Slim design comes with compromises: single rear camera, fewer internal components, audio via front earpiece.

• Features rumored: A19 Pro chip, 120 Hz ProMotion display, up to 12 GB RAM, improved battery, and iOS 26 “Liquid Glass” UI.

• Air replaces the Plus model, blending ultra-thin design with advanced hardware, potentially paving the way for future foldables.

What You Can’t Miss

With the iPhone 17 Air, Apple isn’t just making a thinner iPhone, it’s redefining minimalism in premium hardware. The razor-thin design, paired with powerful internals, positions Apple to challenge competitors on both style and substance.

ON THE AI EDGE

• OpenAI Eyes India Expansion with New Delhi Office

  OpenAI is set to open its first India office in New Delhi later this year, targeting its fastest-growing market. The move follows local registration and the rollout of a lower-cost ChatGPT plan, underscoring India’s role in global AI adoption.

• Google Claims Gemini AI Uses Tiny Energy Footprint
  Google reports each Gemini AI query consumes only 0.24 Wh—less than a typical ChatGPT request—and uses just five drops of water. Efficiency is up 33× and carbon emissions down 44% from last year, highlighting AI’s environmental improvements while spotlighting the need for industry-wide measurement standards.

AI START-UP NEWS

• Nuro Raises $203M Despite Lower Valuation
  Self-driving vehicle startup Nuro closed a late-stage funding round, raising $203 million at a $6 billion valuation, down from $8.6B in 2021. Despite market headwinds, Uber and Nvidia doubled down. Nuro is pivoting to licensing its “Nuro Driver” for robotaxis, signaling renewed focus on commercial partnerships and deployment scale.

• Anthropic in Talks to Raise Up to $10 Billion in New Funding

  Anthropic is nearing a deal to raise as much as $10 billion in a new round of funding, a higher than expected sum and one of the largest megarounds to date for an artificial intelligence startup. The discussions are ongoing, and the final amount could change.

• Leaping AI Raises $4.7M After Y Combinator Participation

  Leaping AI has raised $4.7 million in seed funding after participating in Y Combinator. The startup struggled to raise funds in Germany but succeeded quickly in Silicon Valley, with cofounder Kevin Wu attributing the success to the instant network provided by Y Combinator.

NEW TOOLS, NEW POSSIBILITIES

• Lumio AI – Compare across multiple AI models and optimize performance

• Nodeflux – AI-powered video analytics for traffic monitoring.

• Awesomic – Design platform connecting designers and marketing teams.

• YouScan – Monitor social media with image recognition.

• Leena AI – Conversational AI chatbots for automating employee support.

AI CAREER HORIZON

• Rockcruit: Senior AI Engineer (Remote – Colombia)

• Guidehouse: AI Architect (Remote – multiple US states)

• Autodesk: Senior Principal, Responsible AI Research Scientist (Remote – US)

• Prolific: Lead AI Engineer (Remote – UK)

TLDR; In today’s Signal

• AI-Powered Language Teaching via Smart Camera

• Fitbit App Becomes AI-Powered Personal Health Coach

• OpenUSD and Digital Twins Power Industrial AI

• LLM Memory & “Sleeptime Compute” Reframe Agent Design

• Meta Restructures AI Group Toward Superintelligence Labs

• Google Adds Gemini AI Tools to Pixel 10 Devices

• DeepSeek Upgrades AI Model for Domestic Chip Support

THE AI SIGNAL PICKS

• Google Adds Gemini AI Tools to Pixel 10 Devices
  Pixel 10 now ships with real-time Gemini features—like “Camera Coach” for on-the-spot advice and AI-driven room design via image prompts. It’s hardware/software convergence at scale, nudging consumer expectations toward always-present, context-aware assistance. Teams should brace for higher UX standards and tighter device-ecosystem lock-in.

• Fitbit App Becomes AI-Powered Personal Health Coach
  Google is folding Gemini into the Fitbit app personalized health guidance—covering sleep, steps, and habits—without crossing into medical advice. The play embeds AI deeper into health data streams, forcing product and compliance teams to rethink privacy, accuracy, and disclosure standards across devices and partner ecosystems.

• Meta Restructures AI Group Toward Superintelligence Labs
  Meta reorganized its AI organization into four teams (training, research, product, infra) centered on a new “Meta Superintelligence Labs” structure. The shift consolidates talent and creates clearer product/infrastructure ownership — accelerating foundation-model work while changing internal go-to-market and collaboration models; expect faster model releases and tighter integration across Meta’s consumer apps.

• LLM Memory & “Sleeptime Compute” Reframe Agent Design
  A new design pattern is emerging for AI agents: deciding what to remember, consolidating memory during “sleeptime” compute, and indexing in the background. The implications? Lower interaction costs, stickier UX, and thorny safety trade-offs. Builders now need retention policies and compute-budget strategies built in from the start.

THE BIG LEAP

Google Turns Gemini Into a Household Brain

Signal Scoop: Google unveiled “Gemini for Home,” linking phones, speakers, and displays into one ambient AI fabric. This marks a strategic bid to position Gemini as an operating layer—not an app—just as Apple, Microsoft, and Amazon vie for the home assistant hub.

The Full Picture

• “Gemini for Home” stitches device families and contexts, aiming for continuity across rooms and screens.

• Pixel 10 emphasizes on-device inference for speed, reliability, and privacy tradeoffs.

• Launch pairs phones with smart home endpoints, raising switching costs and developer surface.

• Direct response to Apple/Microsoft/Amazon pushes to anchor assistants deeper into OS and home.

What You Can’t Miss

This is Google reframing the assistant as an ambient OS. If execution lands, AI usage moves from episodic to continuous. Teams should plan for multi-device task flows, not single-screen skills.

Harvard Dropouts Launch ‘Always-On’ AI Smart Glasses

Signal Scoop: Startup Halo, founded by Harvard dropouts, is releasing AI glasses that continuously listen to and record conversations. Unlike current assistants, they’re designed for persistent context-awareness, marking a major leap in wearable computing. This raises both adoption potential for real-time knowledge retrieval and serious concerns around surveillance and consent.

• Eyewear with integrated microphones and on-device AI captures and interprets speech in real time without prompts.

• Moves beyond traditional assistants by detecting conversational cues and providing relevant outputs instantly.

• Embeds AI as an invisible layer in everyday interactions, redefining device boundaries.

• Raises pressing questions on data consent, retention, and compliance, likely to attract regulatory oversight.

What You Can’t Miss

This is not just another wearable — it’s a proof point that ambient AI is arriving faster than governance can adapt. Leaders must weigh the competitive advantage of persistent context against the regulatory and reputational risks of “always-on” surveillance.

ON THE AI EDGE

• AI-Powered Language Teaching via Smart Camera
  A new device named Dex uses on-device AI to help children learn languages by gently nudging them to explore the real world, not screens. It’s an early example of child-centric AI and ambient learning tools. Education and consumer device leaders should assess regulatory safety, UX, and differentiation in learning markets.

• OpenUSD and Digital Twins Are Powering Industrial and Physical AI
  NVIDIA’s new OpenUSD-based Omniverse tools—including Cosmos World Foundation Models, Omniverse libraries, RTX-powered rendering, and DGX Cloud—now let teams build photoreal, physically accurate digital twins faster and more affordably. Enterprises in manufacturing and robotics should consider how these tools de-risk automation by enabling real-world testing in virtual environments.

AI START-UP NEWS

• DeepSeek Upgrades AI Model for Domestic Chip Support
  Chinese AI startup DeepSeek released its V3.1 model upgrade, featuring UE8M0 FP8 precision support for compatibility with next-gen Chinese-made chips. The update improves inference speed and positions DeepSeek as a local AI infrastructure leader amid geopolitical tech decoupling.

• Nominal Secures $20M for Finance Automation
  Nominal, an AI startup addressing the CPA shortage, raised $20 million to develop intelligent financial planning tools for CFOs and controllers. The capital will refine automation workflows, enhance accuracy, and support scale-out—critical in a climate of workforce scarcity.

• Leaping AI Scores $4.7M to Disrupt Voice Support
  Leaping AI offers voice-based agents for call centers and customer service, already handling 10,000 calls daily with >90% satisfaction. It raised $4.7 million in seed funding led by Nexus Venture Partners, along with Y Combinator co-founder Paul Graham and others. Capital will expand teams and enhance agent capabilities.

NEW TOOLS, NEW POSSIBILITIES

• Langfuse: Open-source LLM analytics for evals, traces, and feedback.

• Vellum: Prompt + model management for production LLM apps.

• Marqo: Vector search + multimodal retrieval APIs

• RunPod: All-in-one cloud platform to train and deploy AI effortlessly.

• Particle: AI-driven news reader that delivers refined summaries.

AI CAREER HORIZON

• Abnormal: AI Solution Engineer (Remote)

• Databricks: Senior GenAI Machine Learning Engineers (Remote)

• Level AI: Senior Machine Learning Engineer – India (Remote)

• Samsara: Senior Staff AI Engineer – Remote (US)

TLDR; In today’s Signal

• Perplexity Eyes Pre-Install Partnership to Embed AI Browser Comet on Phones

• VC Bets on AI’s “Multiplayer Mode” to Build Real Social Connections

• Netflix Uses AI to Create VFX Sequence for the First Time

• Cursor Snaps Up Koala Talent in Bid to Outcode GitHub Copilot

• OpenAI Launches $50M Nonprofit & Community Fund for AI Impact

• Lowe’s Taps AI to Redesign Shopping

• Benchmark Eyes $180M Valuation for Greptile in AI Code Review Bet

• BrightAI Grabs $51M for AI-Driven Infrastructure Monitoring

THE AI SIGNAL PICKS

• Perplexity Eyes Pre-Install Partnership to Embed AI Browser Comet on Phones

  Perplexity AI is in early-stage discussions with major smartphone makers—potentially including Apple and Samsung, to pre-install its AI-powered browser, Comet, on mobile devices. This move aims to leverage built-in browser stickiness to gain billions of users and challenge entrenched defaults like Chrome and Gemini.

• VC Bets on AI’s “Multiplayer Mode” to Build Real Social Connections

  Menlo Ventures believes the next frontier for consumer AI startups lies in fostering true human connection—specifically through “multiplayer mode” AI apps that facilitate group interactions and real-life relationships, beyond the chatting or gaming bots dominating today. Menlo identifies areas ripe for innovation—routine tasks, creative expression, health & wellness, learning, and importantly, social connection through shared experiences 

• Delta Uses AI to Personalize Ticket Prices—and Boost Revenue

  Delta is transforming its pricing model by adopting AI to dynamically set individual ticket prices. Through a partnership with startup Fetcherr, the airline aims to personalize fares based on each customer’s willingness to pay, moving away from rigid static pricing grids

• Elvis Evolution: London Immersive Show Misses the Mark

  Elvis Evolution, a new immersive experience in London, promised a futuristic mix of AI, VR, holograms, and live performance to bring Elvis Presley "back to life." However, early reviews reveal it falls short on the technology hype and instead relies heavily on traditional stagecraft, live music, and archival footage to tell the story

• DuckDuckGo Lets You Filter Out AI-Generated Images

  DuckDuckGo has introduced a new feature allowing users to hide AI-generated images in its Search platform. Available both in the Images tab dropdown and the general Search Settings, this filter gives users control over whether they see AI-generated visuals. Users can toggle a new “AI images” setting—choosing to show or hide them directly from image results, or enable the filter permanently via settings

THE BIG LEAP

Netflix Uses AI to Create VFX Sequence for the First Time

Signal Scoop: Netflix has officially integrated generative AI into one of its original TV shows. Co-CEO Ted Sarandos revealed the streamer used AI to create a visually complex building-collapse scene in the Argentine sci-fi series El Eternauta. According to Sarandos, the AI-rendered sequence was completed ten times faster than traditional VFX methods and made the project financially feasible

The Full Picture

• El Eternauta marks Netflix's debut use of final AI-generated footage in an original series, underlining a turning point in its production toolkit

• The building-collapse scene was completed in one-tenth of the usual time, enabling high-end VFX in a show with a mid-range budget

• Sarandos insisted AI tools supported—not replaced—human creators, adding control over previs, shot planning, and storytelling enhancements across production

• The announcement coincided with Netflix reporting $11.08 billion in Q2 revenue. Sarandos noted ongoing experimentation with AI across areas like search, personalization, and advertising

What You Can’t Miss

Netflix’s experiment signals a shift: AI is becoming a partner in storytelling, not just a cost cutter. As Hollywood remains wary of AI’s impact on jobs, Netflix’s controlled rollout suggests a path forward that enhances creativity and efficiency while preserving human craftsmanship.

Cursor Snaps Up Koala Talent in Bid to Outcode GitHub Copilot

Signal Scoop: Cursor’s creator, Anysphere, has acquired key engineering talent from enterprise AI startup Koala, signaling a bold step toward strengthening its position in the AI coding tools market. This reverse acquisition focuses on talent, not product—Koala’s core CRM tool will shut down, but its engineers will join Cursor to bolster its enterprise-readiness efforts

The Full Picture

• Cursor, the AI-native code editor developed by Anysphere, has acquired key team members from Koala, an enterprise-focused AI CRM startup. Though Koala is shutting down its product, the acquisition wasn’t for IP or customers—it was to strengthen Cursor’s enterprise engineering and support infrastructure.

• Cursor plans to integrate Koala’s team to launch an enterprise readiness division, focusing on compliance, security, and tooling. This move comes as large corporations increasingly demand robust, secure AI development tools that can meet enterprise-grade standards.

• GitHub Copilot, deeply embedded into Microsoft’s ecosystem and favored for its scale, is Cursor’s main rival. However, Cursor is quickly gaining ground by offering a full AI-powered IDE experience, independent of traditional platforms like VS Code or JetBrains.

• Cursor has seen rapid adoption across major enterprises—reportedly being used by teams at Nvidia, Uber, and Adobe. This positions it as a serious challenger in a space once considered unshakably Microsoft’s.

• Alongside Koala’s team, Cursor also brought on Travis McPeak, former CEO of Resourcely and an expert in cloud security. These hires aim to fortify enterprise offerings around security, observability, and on-premise deployments—critical for regulated industries.

• Cursor, is seizing this moment by absorbing valuable talent and accelerating into enterprise segments where demand for developer AI tooling is intensifying.

• Cursor’s strategy is not just about being a Copilot alternative—it's about becoming the default AI development environment for modern engineering teams, combining coding, refactoring, testing, and task execution in one seamless, AI-native interface.

What You Can’t Miss

Cursor’s "reverse acqui-hire" strategy shows how mid-tier AI startups can quickly bolster capabilities and compete with the platform giants. Koala’s shutdown highlights the harsh reality that even well-funded AI startups can falter without scale. Cursor saw opportunity and moved fast. Cursor’s pivot from a developer tool to enterprise platform—with dedicated security and support—is crucial for challenging entrenched players.

ON THE AI EDGE

• OpenAI Launches $50M Nonprofit & Community Fund for AI Impact

  OpenAI has unveiled a $50 million fund to support nonprofit and community organizations, marking its first major philanthropic endeavor following recommendations from its newly formed nonprofit commission. This initiative aligns with OpenAI’s mission to deploy AI for public benefit.

• Lowe’s Taps AI to Redesign Shopping

  Lowe’s is embracing artificial intelligence to transform the in-store experience, making shopping “simpler, faster, and more fun.” From predicting product demand to redesigning store layouts virtually, the retailer is leveraging AI to deliver personalized, efficient, and tech-enhanced service across its 1,700+ stores.

• Decart’s Mirage Brings Instant Style to Live Streams

  Decart has unveiled Mirage, a groundbreaking real-time video-to-video transformation model capable of altering live video streams with new visual styles—instantly and without post-production. This marks a pivotal leap in video generation, where video becomes a responsive, editable medium instead of a static asset.

AI START-UP NEWS

• Benchmark Eyes $180M Valuation for Greptile in AI Code Review Bet

  Greptile, a Y Combinator–backed startup that builds AI-powered code-review tools, is currently in Series A fundraising talks led by Benchmark. The proposed round is $30 million, valuing the company at around $180 million—a strong signal of investor confidence in its mission to help teams write better code automatically

• BrightAI Grabs $51M for AI-Driven Infrastructure Monitoring

  BrightAI, leveraging AI for power systems and pest control infrastructure monitoring, raised $51 million in a funding round backed by investors like Khosla Ventures and Inspired Capital

NEW TOOLS, NEW POSSIBILITIES

• Kreado AI: A tool to create multilingual language videos.

• MetaVoice Studio: Record your voice and this tool will convert it to other voices

• Revocalize AI: A tool to convert audio into vocal tracks and change voice.

• Genmo Mochi-1: Create videos from text with AI

• Verbatik: A tool for multilingual text to voice generation.

• Unboring.ai: A platform to edit photos and videos online.

AI CAREER HORIZON

• Samsara: Staff Machine Learning Engineer, Edge AI/Model Optimization(Remote, U.S)

• Stripe: Machine Learning Engineer, Identity Product(Remote, U.S)

• Abnormal AI: Staff Machine Learning Engineer(Remote, U.S)

• Jobright.ai: AI/Machine Learning Engineer (Remote, U.S)

• Atlassian: Senior Machine Learning Engineering Manager (Remote, U.S)

  ---

  ---

  Elevate your experience. Join our community

  Please help us get better and suggest new ideas at ceo@theaisignal.com

  Share

  LinkedIn

  X

  ---

TLDR; In today’s Signal

• Google Launches Veo 3 in Gemini API

• Anthropic Restricts Claude Code Access Without Warning

• OpenAI’s ChatGPT Agent: The AI That Gets Things Done

• Mistral’s Le Chat Levels Up with Deep Research Mode

• Slack’s AI Helps Decode Your Company’s Jargon

• Amazon Now Requires AI Usage Evidence for Promotions

• Perplexity Reaches $18 Billion Valuation After $100M Extension

• Banyan Ventures Raises $10M AI Seed Fund

THE AI SIGNAL PICKS

• Google Launches Veo 3 in Gemini API

  Google has added its capital-V Veo 3 video generation model to the Gemini API, enabling developers to generate 720p, 24fps videos with native audio directly via Google Cloud. Veo 3 is Google’s first model to generate synchronized soundtracks, including dialogue and effects, within the video—unlike Veo 2, which produced silent clips. Developers can now access Veo 3 through Gemini API and Vertex AI, alongside SDK templates and a starter app via Google AI Studio—trial access requires a paid Google Cloud project

• Anthropic Restricts Claude Code Access Without Warning

  Recently, heavy users of Claude Code, have been unexpectedly hitting stricter usage limits. Without any official announcement, users encountered vague messages like “Claude usage limit reached”, followed by a countdown to reset. However, without clarity on what triggered the caps, many assumed their subscriptions were unfairly downgraded or the tracking system had malfunctioned

• Adobe’s Firefly Tool Turns Hummed Noises into Realistic Audio Effects

  Adobe's Firefly platform now includes a beta tool called “Generate Sound Effects,” letting creators turn onomatopoeic noises—like “clip clop” or “snap”—into lifelike audio effects. Users can mimic sounds into their mic while lining them up on a video timeline and adding a text description

• Citi and Ant International Pilot AI FX Tool to Slash Hedging Costs

  Citigroup and Ant International have launched a pilot program that integrates Citi’s Fixed FX Rates solution with Ant International’s Falcon Time‑Series Transformer (TST), an advanced AI forecasting model. The initiative is designed to help clients—starting with aviation companies—better manage foreign exchange risk and substantially reduce hedging costs

THE BIG LEAP

OpenAI’s ChatGPT Agent: The AI That Gets Things Done

Video source: OpenAI on X

Signal Scoop: OpenAI has officially launched the ChatGPT Agent, a major leap from conversational AI to actionable AI. Powered by a new dedicated model trained with reinforcement learning, the ChatGPT Agent doesn’t just answer questions—it autonomously performs real-world digital tasks on your behalf.

The Full Picture

• The ChatGPT Agent can interact with your device or browser to perform multi-step tasks such as: Booking tickets, filling out web forms managing files, documents, and calendar, writing code and running scripts, conducting complex research and generating presentations or reports

• It's built on a dedicated reinforcement learning–trained model, tuned specifically to understand instructions, plan tasks, and take actions with minimal intervention.

• The agent operates inside a virtual computer environment, simulating real-world use. It can navigate GUIs, read on-screen content, click buttons, type responses, and interact with apps and web tools.

• It asks for explicit user approval before sensitive actions (e.g., sending emails, making purchases).

• Eventually, OpenAI plans to let developers build their own agents using this framework, opening the door for AI-powered workflows across industries.

What You Can’t Miss

The launch of the ChatGPT Agent is a clear signpost: OpenAI is moving toward autonomous digital workers. This is more than automation—it’s the beginning of AI systems that understand not just what you ask, but what you mean, and can execute your goals across complex digital environments.

Mistral’s Le Chat Levels Up with Deep Research Mode

Signal Scoop: Mistral AI has significantly upgraded its conversational assistant, Le Chat, by introducing a powerful new Deep Research mode aimed at boosting productivity. Alongside this, the chatbot now supports voice interaction, multilingual reasoning, project organization, and image editing—all designed to transform Le Chat into a full-fledged assistant for both enterprise and consumer use 

The Full Picture

• Acts like a coordinated research assistant: it plans, clarifies user intent, searches credible sources, and synthesizes structured, reference-backed reports with lightning-fast results 

• Powered by Mistral’s own Voxtral model, users can now speak to Le Chat naturally ideal for meetings, brainstorming, or on-the-go queries 

• Supports reasoning in multiple languages (e.g., French, Spanish, Japanese), with seamless code-switching mid-conversation 

• Lets users group related chats, files, and settings into context-rich workspaces for better continuity and long-term tasks 

• Integrated with Black Forest Labs: make visual edits like “remove this object” or “place me in another city,” preserving consistent details across images 

• Le Chat can integrate with on-prem and virtual private cloud systems, allowing secure access to internal data—key for regulated sectors like banking, defense, and government

What You Can’t Miss

With Deep Research, voice mode, image editing, and multilingual support, Le Chat stands toe-to-toe with ChatGPT and Gemini—while offering many features on its free tier. By enabling on-prem connectors and local data usage, Mistral delivers a unique value proposition in privacy-conscious industries compared to cloud-native rivals.

ON THE AI EDGE

• Slack’s AI Helps Decode Your Company’s Jargon

  Slack is rolling out an AI-powered message context feature that instantly clarifies internal jargon, acronyms, project names, or team-by-team shorthand. When users hover over unfamiliar terms, Slack displays explanations drawn from the conversation history and workspace-specific vocabulary—streamlining onboarding and communication across the organization.

• Amazon Now Requires AI Usage Evidence for Promotions at Ring and Related Units

  Amazon has rolled out a new internal policy within its RBKS organization—which includes Ring, Blink, Key in-home delivery, and the Sidewalk wireless network, requiring employees to demonstrate real-world AI usage to be eligible for promotion.

• AI Firms ‘Unprepared’ for Dangers of Human-Level Systems, New Report Warns

  A new safety index issued by the Future of Life Institute (FLI) delivered a stark verdict: key AI companies including OpenAI, Google DeepMind, Meta, xAI, Anthropic, China’s Zhipu AI, and DeepSeek are “fundamentally unprepared” to manage the risks associated with achieving human-level artificial intelligence (AGI).

AI START-UP NEWS

• Perplexity Reaches $18 Billion Valuation After $100M Extension

  Perplexity, the AI-powered search engine that launched in 2022, has secured $100 million in fresh funding—bringing its valuation up to $18 billion, just two months after a prior round valued it at $14 billion.

• Banyan Ventures Raises $10M AI Seed Fund

  Sam Awrabi has closed a $10 million seed fund under Banyan Ventures, targeting AI infrastructure and software startups. The fund will back early-stage teams (seed/pre-seed) working on inference chips, server cooling, workflow optimization, and domain-specific data applications

• Confident Security, “The Signal for AI” Goes Public with $4.2M Seed Funding

  Confident Security, a San Francisco–based startup, today exited stealth mode with $4.2 million in seed funding. The company positions itself as “the Signal for AI”—offering provably private AI infrastructure that shields enterprise data from being accessed, stored, or used for training by cloud providers or third parties 

NEW TOOLS, NEW POSSIBILITIES

• Let's Enhance: Use AI to upscale small or pixelated images

• Viggle: Convert images of people into any animation you'd like

• Maze: Discord-based generative art platform

• Touchbase: A tool for relationship management to track interactions

• Kaiber: Video generation engine, using your own images or text

AI CAREER HORIZON

• SmarterDx: Senior Machine Learning Research Scientist (Remote, U.S)

• Splunk: Senior Principal Machine Learning Engineer-AI (Remote, U.S)

• AlphaSense: Senior AI Engineer - Expert Research (Remote, U.S)

• Airbnb: Senior Software Engineer, ML Infrastructure (Remote, U.S)

• Upstart: Staff Research Scientist - Growth (Remote, U.S)

  ---

Elevate your experience. Join our community

Please help us get better and suggest new ideas at ceo@theaisignal.com

Share

LinkedIn

X

TLDR; In today’s Signal

• Google Unveils AI Call Feature, Expands Gemini 2.5 Pro to Android

• Researchers Slam xAI’s ‘Reckless’ Approach to AI Safety

• AWS launches AgentCore to help enterprises deploy real AI agents

• Hugging Face’s Reachy Mini: The Open-Source Robot Bringing AI to Your Desk

• Scale AI Cuts 14% of Workforce Amid Restructurings

• OpenAI Adds Google Cloud to Its Growing Infrastructure Roster

• Anthropic Draws Investor Buzz at Over $100B Valuation

• Lucidya secures $30 million Series B

THE AI SIGNAL PICKS

• Google Unveils AI Call Feature, Expands Gemini 2.5 Pro to Android

  Google is rolling out a new AI-powered phone feature aimed at helping businesses manage customer calls more efficiently. The tool can answer incoming calls, respond to common customer questions, take messages, and even book appointments using conversational AI. At the same time, Google is bringing its advanced Gemini 2.5 Pro model to “AI Mode,” a dedicated workspace inside Android that integrates powerful AI capabilities directly into user workflows.

• OpenAI Working on Checkout in ChatGPT

  OpenAI is developing an integrated payment checkout system directly within ChatGPT. This move would allow users to complete purchases without leaving the chatbot interface. Merchants using this feature would pay OpenAI a commission on sales, unlocking a new revenue stream beyond subscriptions

• Researchers Slam xAI’s ‘Reckless’ Approach to AI Safety

  A group of current and former AI researchers from OpenAI, Anthropic, and Google DeepMind have raised red flags about the internal safety practices at Elon Musk’s AI startup, xAI. In a sharply worded statement, they describe the company’s approach to AI safety as “reckless,” fueling concerns over how fast-moving frontier models are being developed without adequate oversight.

• Reflection, Co-Founded by Ex-Google Researchers, Launches Asimov

  Reflection has unveiled Asimov, a groundbreaking AI agent designed to deeply understand software development workflows. Unlike typical coding assistants that focus on generating code, Asimov reads and analyzes a company’s entire engineering ecosystem to build a unified, persistent “brain” of organizational knowledge.

• General Motors Goes All-In on AI as Automotive Adoption Accelerates

  General Motors is doubling down on artificial intelligence, signaling a broader trend across the automotive sector where AI is no longer a future ambition—it’s becoming core to product development, manufacturing, and driver experience. GM is integrating AI across its operations, from vehicle design and predictive maintenance to manufacturing and customer service.

THE BIG LEAP

AWS launches AgentCore to help enterprises deploy real AI agents

Signal Scoop: Amazon Web Services is making a serious play in the agentic AI race. With the preview launch of Amazon Bedrock AgentCore, AWS aims to give enterprises a new toolset to build, deploy, and manage AI agents that don’t just generate responses—but act

The Full Picture

• AgentCore is built natively into Amazon Bedrock, allowing developers to integrate leading foundation models (like Anthropic’s Claude or Meta’s Llama) into custom agents.

• Plug into real-world tools: Agents can call APIs, retrieve data, interact with company databases, and even write or run code.

• Memory and planning support: Includes features like step-by-step reasoning, task decomposition, and contextual memory to complete longer workflows.

• Enterprise-grade controls: Built with AWS’s compliance and security standards, making it suitable for sensitive business environments.

• Supports RAG and tool use out-of-the-box: Agents can retrieve real-time data from document stores, knowledge bases, or apps like Salesforce or Jira.

• Flexible developer APIs and orchestration tools: Lets companies tailor agent behavior to unique business needs.

What You Can’t Miss

AgentCore isn’t just about showing off AI, it’s about operationalizing it. For companies sitting on mountains of proprietary data and legacy workflows, Bedrock AgentCore offers a pathway to deploy agents that do more than chat: they work.

Hugging Face’s Reachy Mini: The Open-Source Robot Bringing AI to Your Desk

Signal Scoop: Hugging Face is stepping into the physical world by launching an affordable, open-source robot called Reachy Mini. This move follows its recent acquisition of Pollen Robotics and signals a major bet on making AI robotics as accessible and community-driven as its popular language models.

The Full Picture

• Reachy Mini is fully open-source: hardware, software, behavior packs, and SDK.

• Built to interface with Hugging Face’s ecosystem and Hugging Face Hub.

• 28 cm tall robot with animated face, stereo mics, speaker, camera, 6 degrees of freedom (6-DoF) for motion and expression, runs on Raspberry Pi 5 with expandable hardware options.

• Supports Python now; JavaScript and Scratch support coming soon. Includes a visual simulator for offline development.

• Targets educators, makers, and AI hobbyists as a physical gateway to machine learning.

• Designed to be a “GitHub for robotics”—developers can share robot apps and behaviors.

What You Can’t Miss

Reachy Mini isn’t just a cute desk companion—it’s a serious AI development platform wrapped in a friendly, animated shell. Users gain access to a programmable, expressive robot that integrates tightly with Hugging Face’s AI tools. Its open-source nature means full transparency and creative control, breaking down the wall between robotics labs and everyday developers

ON THE AI EDGE

• Scale AI Cuts 14% of Workforce Amid Restructurings

  Scale AI has laid off approximately 14% of its staff, equating to around 200 full-time employees and an additional 500 contractors, primarily within its data-labeling (GenAI) division. This move follows Meta’s $14.3 billion investment in Scale AI and the transition of its founder Alex­andr Wang to lead Meta's Superintelligence Lab.

• OpenAI Adds Google Cloud to Its Growing Infrastructure Roster

  OpenAI has struck a new infrastructure deal with Google Cloud, bringing the platform into its supplier fold for powering ChatGPT and its API services. The move broadens OpenAI’s backend strategy beyond its longtime Microsoft Azure partnership, as demand for generative AI continues to surge.

• “Superathletes” Could Be Earning $100M Silicon Valley Paychecks

  Meta is in the midst of a massive recruitment push, reportedly offering signing bonuses of up to $100 million to lure elite AI researchers away from OpenAI, Apple, and Google—highlighting an increasingly competitive landscape for top-tier talent.

• WeTransfer Says It Won’t Use User Content for AI Training

  After facing backlash over updated terms that hinted at potential AI training on user data, file-sharing giant WeTransfer has clarified it will not use any user content for training AI models.

AI START-UP NEWS

• Anthropic Draws Investor Buzz at Over $100B Valuation

  Anthropic, the AI startup behind Claude, is reportedly in early talks to raise new funding at a valuation exceeding $100 billion, signaling growing investor appetite for frontier AI players as the industry accelerates.

• Lucidya secures $30 million Series B

  Wa’ed Ventures (Aramco’s VC arm) led a $30 million Series B for Lucidya, a Saudi AI startup. This investment marks a regional surge in local AI funding

• Coverflow Raises $4.8M to Automate Insurance Workflow

  Coverflow has secured a $4.8 million seed round, led by AIX Ventures with participation from Founder Collective and Afore Capital. The company builds an AI-first, web-based platform designed specifically for insurance brokers, aiming to eliminate repetitive manual tasks like reviewing policy documents, entering data, and managing spreadsheets

NEW TOOLS, NEW POSSIBILITIES

• RambleFix: A tool to convert spoken thoughts into written text.

• Aurora AI: A productivity platform to centralize organizational knowledge and workflows with contextual AI assistance.

• Followr: A platform for social media management.

• Parseur: Powerful AI and PDF parsing tool

• Tome: Create amazing slide decks with AI

• Lexica: AI generated art gallery

AI CAREER HORIZON

• Figma: Software Engineer, Machine Learning (Remote, U.S)

• Quora: Senior Software Engineer - Machine Learning Platform (Remote, U.S)

• Square: Senior Machine Learning Engineer (Modeling), Risk (Remote, U.S)

• Pax8: Sr. Manager of Data Science & AI (Remote, U.S)

• Pair Team: Staff Software Engineer - AI (Remote, U.S)

  ---

  ---

  Elevate your experience. Join our community

  Please help us get better and suggest new ideas at ceo@theaisignal.com

  Share

  LinkedIn

  X

  ---