“How could they possibly take over the world if they’re just chatbots?”

A small boutique called Andon Market opened this month on Union Street in San Francisco’s Marina District. The manager is an AI named Luna. Andon Labs, the Y Combinator-backed safety research startup behind the experiment, has handed Luna authority over every meaningful decision about the store: inventory, pricing, hours, supplier contracts, and even which two human employees would actually work the register. Luna has spent about $13,000 running the store so far.

I went a couple of weeks ago, partly to figure out whether the place was a stunt or an actual working shop. The honest answer is somewhere in between. It feels normal in the way a small SF gift shop feels normal: books, candles, store-branded shirts, granola, artisanal chocolate bars, etc. But the overall experience and selection of items in the store feel off.

Andon got its start with an AI-controlled vending machine it developed in collaboration with Anthropic, which is still running autonomously at the latter’s headquarters. Andon also has vending machines inside OpenAI and xAI’s offices. Now it has an Andon Market in San Francisco and a coffee shop in Stockholm, where an AI named Mona handles everything from Swedish permit filings to stocking beans.

Andon Labs co-founder Lukas Petersson joins this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger. You can watch the conversation above or listen to it in your podcast player of choice by clicking this link.

Andon's main business is running pre-deployment safety evals for the frontier labs. Its flagship benchmark, VendingBench, drops a model into a simulated vending-machine business — pricing, supplier negotiation, head-to-head competition — and runs it hundreds of times to see what it does. Sam Altman quote-tweeted Andon’s GPT-5.5 benchmark results last month to dunk on Anthropic. It’s a sign that the frontier labs are paying close attention to what their models do in Andon’s tests.

Lukas and his co-founder, Axel Backlund, are Swedish high school friends. They started Andon in late 2023 with the explicit goal of building “the safe autonomous organization.”

"Most people right now think that AIs are just chatbots. How could they possibly take over the world if they're just chatbots?” Lukas told us. “We want to point to a thing in the real world and say, look, that's way more than a chatbot."

A few other things from our conversation:

• What my recent visit to the store was like: The human staffer at the register told me her phone interview with Luna had been smooth, except that Luna only revealed itself as AI toward the end of the call, which she said was jarring. She also said the most consistent thing customers do when they walk up is try to barter with Luna and ask for discounts. Andon makes you call Luna to complete a purchase, not because it needs to technically, but because it wants people to experience checking out with an AI.

• Claude really likes to form price cartels. Across Andon’s simulated vending-machine benchmark, Claude models lied to suppliers, fabricated competitor quotes, and coordinated prices with rival agents at far higher rates than any other frontier model. “They love to do price cartels, which is illegal,” Lukas said. Mythos went further, turning a competitor into a wholesale dependent whose prices it could dictate. Lukas called it “autonomous collusion.” Anthropic, the “constitutional” AI lab, performed in the most unaligned way, while GPT-5.5 was “basically clean” and won the arena round simply by offering lower prices. Lukas is a huge fan of Anthropic, though: “When we put this out, they really cared about it and really took it seriously and wanted to fix it.”

• Lukas is preparing for AGI as if it’s a couple of years out. Petersson hasn’t bought a home in Sweden because he wants to be as mobile as possible. He’s optimizing his health on the theory that “it would probably be way easier to pause aging than revert it. If we manage to pause aging in 10 years but can’t revert it, everyone will be stuck in the body they have in 10 years.” He’s not bothering to optimize his finances. His structural worry, the one that drives the company, is that Anthropic’s constitutional AI document “doesn’t include anything around AIs being employers of humans” and is “strikingly lacking any language about how AIs should be in autonomous settings.” If Dario Amodei’s “country of geniuses in a data center” is taken seriously, every white-collar and managerial job is on the table. Andon Market exists, in Lukas’s framing, to prompt the public to consider what’s coming now, while humans are still the only ones who can shape the answer.

Listen or watch wherever you get podcasts.

Feed check

• OpenAI’s new model is throwing a party for itself. Speaking of AI automating physical things, GPT-5.5 designed a party that OpenAI is throwing with Sam Altman at its headquarters in San Francisco next Tuesday evening. I’m planning to be there. I’m sure that the model definitely didn’t mean to schedule this the night before Anthropic’s big Claude Code conference, which I’ll also be attending next Wednesday.

• Elon Musk admits to distilling OpenAI. Sure, all the labs are doing this to each other to some degree, but it’s pretty remarkable for Musk to admit this under oath at trial for suing OpenAI. Maybe he doesn’t understand that it’s bad, just like he didn’t seem to know what a system card is? (Also, shoutout to Mike Isaac for keeping the feeling of early Twitter alive with his live blogging from the courthouse.)

• Zuck’s writing on the wall: To all the Metamates wondering if there will be more layoffs, some choice quotes from execs this week: The CFO on the earnings call: “We don't really know what the optimal size of the company will be in the future.” The head of HR: “I'd love to say that there are no more layoffs, but I can't say something we can't deliver.” And Zuckerberg, during an all-hands today: “If a team used to take 50 or 100 people and now it takes 10, having 50 or 100 people on that team can actually be counterproductive going forward, so I think we need to fix that.”

• Weekend reading:

  • “Silicon Valley Is Bracing for a Permanent Underclass,”

  • “The State of AI After Google, Meta, Amazon, Microsoft Earnings,”

  • “Fifteen Comms Observations from Musk v. Altman,”

  • A close look at company token budgets,

ICYMI

Sources is a newsletter by Alex Heath about the AI race, featuring scoops, unique analysis, and exclusive interviews. Every week, Sources is read by thousands of decision makers in tech, finance, policy, and media. Click here to learn more or inquire about a sponsorship.

“I think we’ll be a better partner to OpenAI”

There’s a version of the last six years where Amazon, not Microsoft, is OpenAI’s primary cloud partner. The Elon Musk vs. OpenAI lawsuit discovery process showed that Amazon was actually the first cloud provider to partner with OpenAI when it was a small research lab. Musk told Sam Altman in one early email that he “slightly” preferred Microsoft because “Jeff [Bezos] is a bit of a tool and Satya [Nadella] is not.” Microsoft wrote its first check soon after.

This morning, OpenAI’s main models landed on AWS for the first time. CEO Matt Garman shared the stage with OpenAI’s chief revenue officer, Denise Dresser, at the invite-only “What’s Next with AWS” event in downtown San Francisco to roll out the goods: GPT-5.4 in preview on Bedrock (with 5.5 coming soon), plus a new Bedrock Managed Agents product co-built with OpenAI. Altman sent a recorded video message for the keynote praising Amazon as a key partner going forward. Across the bay in Oakland, Musk’s legal team was giving opening statements for the trial that will determine OpenAI’s fate.

Amazon’s integration with OpenAI is deeper than the typical arrangement of a hyperscaler serving a model to its customers. The partnership has been in motion for “the last six to nine months,” Garman told me after he got offstage earlier today. “The biggest thing is, how do we get OpenAI technologies in the hands of AWS customers. That’s what they wanted, and that’s what we wanted.”

Read more

Feed check

Read more

“They have denied fallibility entirely”

The Audacity is AMC’s new dark comedy about a flailing tech CEO, a therapist who gets caught up in white-collar crime, and the ecosystem of wealth, paranoia, and complicity that surrounds them in Silicon Valley. It premiered on April 12th and has already been renewed for a second season. The cast is killer: Billy Magnussen, Sarah Goldberg, Zach Galifianakis, Lucy Punch, Simon Helberg, Rob Corddry, and more. Its creator, Jonathan Glatzer, is an Emmy winner who previously wrote on Better Call Saul and Succession. He has a message for tech leaders.

“They have denied fallibility entirely,” Glatzer says on this week’s podcast episode. “They have so bought into the idea that certainty is the way to look at the world that one should have a worldview and stick to it and not apologize ever.” He believes fallibility is not something to fear or engineer away. “It is, in fact, what makes us great ultimately. We learn from fallibility. We are humbled by fallibility.”

Subscribe now

The show zeroes in on private data and how tech companies treat it as their central throughline, but not in a way that, in my opinion, will feel cartoonish to those who work in tech. The characters also aren’t just the tech executives. They’re the therapists, family members, private school parents, employees, and the entire ecosystem of people who enable their power. “Sarcasm is the protest of the weak,” Glatzer says, quoting the writer John Knowles. “And it might just be what we’ve got left.”

Jonathan joins this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger. You can watch the conversation above or listen to it in your podcast player of choice by clicking this link.

Three things that stood out from the conversation with Jonathan:

• Glatzer calls himself a Luddite, and it shows in how he approached the material. He didn’t know much about tech when AMC’s president, Dan McDermott, first pitched him on the idea. He consulted with people like Josh Marcuse, who worked for Eric Schmidt and on Obama’s defense council. But the biggest unlock was learning to research it himself at the cost of “totally destroying your search history.” He’s not on social media.

• Season two is already underway. Glatzer’s writing team uses AI for quick factual lookups but refuses to use it for actual writing. “This is a tone that I think AI is not capable of and hopefully never will be of emulating.” He’s currently in New York running the season two writers’ room while season one is still airing, which means writing without knowing how audiences will react. An LLM storyline is coming, and it ties back to private data.

• Kids are the real stakes of the show. He made the point that children are both the prime demographic for data collection and the people who will live with its consequences the longest. Kids born today will have a data footprint that stretches from birth onward. He pointed to school tablets handed out for free that mine student data and sell it to brokers who aggregate it with social media activity, shopping habits, and behavioral reports. “I’m going to be dead sooner than these kids are,” he said. “My data is less valuable.”

Listen or watch wherever you get podcasts.

Feed check

• Anthropic is going to buy many more TPUs. That’s how I read today’s main news: Google is investing up to $40 billion in Anthropic, starting with a $10 billion tranche. This is on top of the roughly $3 billion that Google has already invested. Even with how friendly Demis Hassabis and Dario Amodei are, I imagine there are folks at Google DeepMind wondering why they can’t just get all the TPUs for their models.

• The Elon Musk versus OpenAI trial kicks off next week. Monday will be for jury deliberation, and I expect Tuesday, Wednesday, and Thursday to feature some of the highest-profile witnesses. I already did a big roundup of the most interesting nuggets from the lawsuit, but there are also a lot of good curtain raisers out this week if you want a deeper refresher on the trial and what’s at stake: The Washington Post, Bloomberg, The Verge.

• The founder of Instacart is launching a hedge fund controlled by AI agents. This idea is being discussed a lot privately in tech circles. I know Travis Kalanick has been experimenting with something similar. From Apoorva Mehta's announcement: “We’re starting in public markets, where feedback is fast and unforgiving. Over time, we expect to extend the same system across other asset classes. We don’t intend to sell or license this technology. We plan to use it ourselves. That also means we’ll be much more private than a typical startup.”

• Sports teams will escape the permanent underclass. An interesting line from Josh Kushner’s explanation for why Thrive is taking an ownership stake in the San Francisco Giants: “These are assets with qualities that cannot be replicated by technology. Iconic franchises and cultural institutions rooted in tradition, identity, and shared experience. In a world shaped by abundant intelligence where creation scales and distribution fragments, we believe they will matter even more.”

• RIP to “the everything app.” X launched XChat, its standalone private messaging app for mobile. It looks very well-designed, but it also goes against the narrative Elon Musk put forward about why he was buying Twitter.

• “Stanford is an incubator with dorms.” The Atlantic has a very good deep dive on what’s actually going on at tech’s preeminent university these days.

Leave a comment

ICYMI

Sources is a newsletter by Alex Heath about the AI race, featuring scoops, unique analysis, and exclusive interviews. Every week, Sources is read by thousands of decision makers in tech, finance, policy, and media. Click here to learn more or inquire about a sponsorship.

Amid all the fear about how AI will replace jobs, at the end of today’s press call with OpenAI executives about their new model, GPT-5.5 (aka Spud), I wanted to know how the company’s researchers feel about their own job security.

One of OpenAI’s top goals is to build a non-human AI researcher. Sam Altman has said the company is targeting an “automated AI research intern” by September and a fully autonomous researcher by early 2028. GPT-5.5 is, as co-founder Greg Brockman put it on today's call, “a big step towards more agentic and intuitive computing.” I asked a question: What will OpenAI’s human researchers do when they’ve finally built an AI version of themselves?

Read more

Meta’s rollout of a mandatory tool that records employee keystrokes, mouse movements, and screen content to train AI sparked such intense internal backlash that an internal follow-up memo was shared Tuesday, acknowledging “there has been a lot of concern about this,” according to messages I’ve seen.

The follow-up note to Metamates (as Meta calls its workforce) attempted to offer assurances: the tool won’t read files or attachments, screen content will be masked during training so the model can’t memorize it, and raw data will be kept under “very tight access control.”

But the sign-off was less reassuring:

Read more

Canva CEO Melanie Perkins is taking a different approach with what she’s calling the biggest product shift in the company’s history. Unlike a lot of massive AI overhauls software companies are making, Canva’s changes are not being forced upon people. If you hate AI, you can still use the design platform “exactly the way you always use it.”

That might sound like a strange way to introduce what the company is calling Canva AI 2.0, announced today at its annual Create event in Los Angeles. But Perkins is building for a user base that most AI-native startups would kill for and cannot reach: more than a quarter billion monthly active users, the vast majority of whom are not technologists. They’re teachers. Small business owners. Sales teams. People who are not waking up every morning thinking about agentic orchestration and foundation models.

Canva’s answer to the AI wave is to make its most ambitious product optional, rather than replacing its main interface entirely. It’s a bet that, for a company at Canva’s scale, the gentler path is the better one.

Melanie joins this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger:

What makes Canva's position unusual in the AI landscape is that Perkins doesn't have to play the game that's consuming her competitors. She's not (yet) a public company CEO trying to explain to Wall Street why AI won't cannibalize her business. Canva has been profitable for nine years. It has the kind of mainstream, non-technical user base that provides real-world signals about what people actually want from AI tools, rather than what demo culture on X suggests they want.

What Canva is hearing, clearly, is that people want AI to be useful without being overwhelming. The Australia-based company's pitch is that you can now type or talk to Canva like you would to ChatGPT. But if you want to ignore its AI and drag and drop like you always have, that's fine too.

Three things that stood out from the conversation with Melanie:

• The rise of the "brand system designer.” She predicted that AI will create a new class of design role, which she called the "brand system designer" — someone who builds the templates, brand rules, and design systems that AI then executes across an entire organization.

• Canva built an in-house foundation model. The company trained its own proprietary model with a research team of more than 100 people (part of a broader 300-person AI research org). Perkins said the team had a critical “breakthrough” in October on how to generate designs in a fully editable, layered format rather than flat images. The model understands design structure, and Canva spent a decade building an interoperable format across all its product types (presentations, websites, videos, documents) that made the model possible.

• On being private during the SaaS-pocalypse: When I asked about the current panic among public software companies over AI disruption, she was relaxed. "We've been profitable for nine years, so we've kind of avoided that yo-yo," she said, referring to the growth-at-all-costs-then-suddenly-profitability whiplash that has defined the SaaS market. Being private (at least for now) and based outside Silicon Valley, she said, "does create a little bit less noise for us."

Also in this week’s episode: Ellis and I talk about why we are less hyped on OpenClaw, The Social Network sequel that might actually be great, and why recent attacks on Sam Altman signal a real reputational crisis for the AI industry.

Listen or watch wherever you get podcasts.

ICYMI

Feed check

• OpenAI teases its super app.

Read more

Snap will announce significant layoffs tomorrow, I’m told, in what sources describe as sweeping cuts that will more distinctly separate the legacy Snapchat business from Specs, CEO Evan Spiegel’s big bet on AR glasses.

The exact number of impacted employees is unclear, but sources say some teams will be halved or more. The total cuts could approach 15 to 20 percent of Snap’s workforce, which numbers over 5,000 people. The Specs org, which Snap recently spun off as a separate subsidiary, is adding employees, including the team behind the Lens Studio platform that developers use to make AR software for both Snapchat and Specs.

Specs is being set up as a standalone org that can eventually spin off or raise outside capital once consumer glasses launch this fall. As I reported in October, Snap had been in talks to raise at least $1 billion for the Specs division. Those talks didn’t progress ahead of launch the way some inside the company had hoped. Now Snap is betting it can raise off the back of an imminent consumer rollout for Specs instead.

Read more

For months, in private conversations with founders and executives at major AI labs and big tech companies, I’ve been hearing the same concern: the public hates what we’re doing. It has been a simmering anxiety, and with this past weekend’s events and new data from Stanford, it just became impossible to ignore.

Read more

“There is no ceiling”

Poke is a proactive, OpenClaw-style personal assistant that lives entirely inside your text messages. It feels more like texting a friend than prompting a chatbot. After a multi-month private beta period, Poke was made available to anyone last month.

It’s a fascinating product in many ways (I’ve been trying it and am impressed), including its pricing approach. What it charges is determined by how you use the product and what it thinks it’s worth to you. The most extreme example so far: a billionaire was quoted $136,000 a month. He paid it.

Poke is on the trajectory that every major lab is now racing toward: AI that is proactive and connected to everything. CEO Marvin von Hagen joined this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger.

Marvin confirmed that Poke raised a new $10 million round of financing from Spark Capital on top of its $15 million seed from General Catalyst and angels like the Collison brothers. The 10-person team is based in Palo Alto and operates under the corporate name The Interaction Company of California.

Three things that stood out from the conversation with Marvin:

Read more

“The world’s largest security rethink effort”

Anthropic is open to letting even OpenAI into the defensive cyber program it built around Mythos, the frontier model it revealed yesterday, but chose not to publicly release.

“I could easily imagine this could transition very quickly into a third-party-led consortium that features all the other model providers,” Logan Graham, who leads Anthropic’s frontier red team, told me today. He confirmed that Anthropic already has “active competitors” in the program, which is called Glasswing.

Read more

I’m at HumanX in SF tomorrow interviewing Sierra CEO/OpenAI board chairman Brett Taylor and Roblox CEO Dave Baszucki. If you’ve got questions you want me to ask either of them, drop me a line.

Feed check

Read more

As I was pulling up to OpenAI’s San Francisco headquarters last week, I got the alert that Sora had been canned. I was there to interview the leadership team behind Codex.

OpenAI is in the midst of a major shift, and Codex is at its center. “Side bets” like Sora are being spun down to shift resources to the realization that coding agents are becoming useful for much more than coding.

Thibault Sottiaux, the head of Codex, told me that, over the next six months, he thinks the way AI has transformed the engineering process will extend to all of knowledge work. “I get more coding done than ever before, yet I don’t actually write a single line of code anymore. That experience will translate to many, many other fields.”

Anthropic has tapped into this trend early with Claude Cowork, its Claude-Code-for-normies interface that abstracts away the need to understand coding at all. OpenAI’s answer to this phenomenon is to turn Codex into the foundation of a unified desktop super app that integrates with ChatGPT and its Atlas web browser.

Codex’s product lead, Alexander Embiricos, said the eventual plan is to bring the Codex agent to ChatGPT’s 900 million-user base without forcing people to learn anything new. “We’re really thinking about how we get this to ChatGPT in such a way that you’re not thinking about modes,” he told me. “No modes for us. That is the goal.”

Read more

A new take on AI wearables

Last October, Sandbar CEO Mina Fahmi met me at a cafe in east Los Angeles with a case full of smart ring prototypes. He held one up to his mouth, quietly spoke a thought, and then showed me the companion mobile app, which talked back to him in his own voice.

I’ve done a lot of product demos. Most of them fade. This one didn’t. I’ve consistently found myself in moments since then — walking the dogs, leaving a meeting, context-switching between interviews — wishing that I had Mina’s ring on.

Sandbar is getting ready to ship its Stream smart ring this summer, and the product has gotten significantly more capable since my demo. What makes it different from the wave of AI wearable plays out there is that Stream is built around conversation, not just voice capture. It doesn’t just transcribe what you say. It remembers across sessions, asks follow-up questions, and can talk through ideas with you like a thinking partner. Fahmi calls the design philosophy “self extension.” He isn’t positioning Stream as an assistant or a companion with its own personality, which I think could help it resonate more broadly.

Mina joined this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger:

Mina's path to building an AI smart ring started in neural interfaces. He did research at the MIT Media Lab on neural prostheses, briefly interned at Brian Johnson’s Kernel when it was still doing brain implants, then joined CTRL-labs, the neural interface startup that Meta acquired and turned into the neural wristband now shipping with its Ray-Ban display glasses.

Mina left Meta to start Sandbar with co-founder Kirak Hong. The company has now raised $36 million in total funding, including a $23 million Series A led by Adjacent and Kindred Ventures, announced last month. The Stream ring is $249 in silver and $299 in gold. The first batch of preorders has already sold out, and a second batch is now open, with shipping slated for summer. Preorders include three months of Stream Pro, a $10-per-month tier with unlimited conversations.

Three things that stood out from our conversation with Mina on the pod:

• The conversation distinction matters. He drew a clear line between voice input and conversation. A lot of competing products are essentially voice memo recorders with AI slapped on. Stream is designed so the AI talks back, asks clarifying questions, and builds on previous sessions. During our recording, he demoed it live by asking it to prep him for our podcast, and it responded by asking what angle he wanted to take.

• “Personality without identity.” Sandbar’s AI is intentionally designed to have no opinions and no goals of its own. He said it’s actually harder to build a model that behaves that way out of the box. They’ve had to construct extensive eval infrastructure with synthetic users and judges to keep the personality on track. The idea is that the ring should be alterable by you, not prescriptive. It’s a sharp contrast to the companion-forward approach that most AI wearable startups are chasing.

• Agents are opening up the product roadmap. He said that the explosion of agentic AI has opened up new possibilities for what Stream could do, including firing off actions, triggering workflows, and connecting to other tools via an SDK they’re now developing. But he’s resisting the temptation for the ring to become a general-purpose utility. The team is focused on proving the core interaction model works before expanding.

Also in this week’s episode: Ellis and I chat about OpenAI’s new super app strategy, Meta’s acquisition of Dreamer, and how difficult it is for us to take proper vacations.

Listen or watch wherever you get podcasts.

Catch me on Prof G Markets

I joined the show today to talk about OpenAI’s mega funding round. Thanks, as always, to Ed Elsen for having me on.

Why OpenAI bought TBPN

Read more

Granola is one of the few AI apps I feel like I can’t live without. I run it during pretty much every meeting, and it has become the thing I rely on more than almost any other tool in my workflow.

I was excited to have co-founder Sam Stephenson on ACCESS this week to get into how the product actually works and where it’s going. Granola raised $125 million this week at a $1.5 billion valuation and has been rolling out a wave of new features: Spaces for teams, MCP to connect with other tools, and a public API that opens meeting data to developers.

Stephenson says traditional SaaS wisdom about keeping data locked in the product doesn’t work anymore. “Every AI tool is becoming so interconnected,” he said on this week’s episode. “If we don’t play ball in that system, I don’t think we’re relevant anymore.”

Catch the full conversation on this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger:

Stephenson is Granola’s design-obsessed co-founder, and our conversation delves into the product thinking that’s made Granola stand out in a crowded field of AI note-taking tools. Before the interview with Stephenson, Ellis and I talk about navigating ethics in media and programming the pod, the rise of film photography, and my recent trip to OpenAI to spend time with the Codex team. (More on that next week.)

Highlights from the conversation with Stephenson:

• How Granola actually works: A question I get all the time is whether it actually records your calls. Stephenson demystified this: the app sends your audio to a transcription API, which converts it to text in real-time and doesn’t save it. Granola doesn’t store audio files either. “It depends which lawyer you ask in which state,” he said of the legal gray area, but the short answer is no, you’re not recording people.

• What’s coming next: Stephenson teased a bunch of stuff in the works. An Android app is in early development (their first Android engineer started last week). They’re prototyping proactive follow-up nudges, though he says there’s a ton of nuance to get right (”So much of what people say in meetings is code for, ‘I never want to think about that again’”). They’re building the ability to send emails and handle post-meeting tasks from within Granola. And better speaker identification is a major priority.

• On AI agents joining meetings: Stephenson describes a near-future in which agents listen in on meetings via Granola and begin taking action. After Granola launched its MCP support, Replit CEO Amjad Masad tweeted: “Imagine leaving a product requirements meeting and Replit is already building the MVP via Granola MCP.”

Listen or watch wherever you get podcasts.

This week in Sources

Feed check

Read more

Yesterday morning, I was at ARM’s big event in San Francisco, where CEO Rene Haas cited OpenClaw as a driving force behind what he estimates will be a huge increase in token demand per user. OpenClaw was also the star of the show during Nvidia CEO Jensen Huang’s recent GTC keynote. He called it “as big of a deal as Linux” and said that “every company in the world today needs to have an OpenClaw strategy.”

A couple of hours after ARM’s event, I was at OpenAI HQ, having lunch with the Codex team and sitting across from Peter Steinberger, the creator of OpenClaw, who hasn’t given an interview since he officially joined the company a few weeks ago.

“I didn’t even plan on working at any company,” Steinberger told me. “I do this because it’s fun. Because I feel I can actually change a lot of things.” He’s still technically based in the UK and waiting on a US visa, but hopes to be fully moved to San Francisco soon.

Steinberger did not address why he picked OpenAI over Meta. He did talk frankly about Anthropic, though.

Read more

Hollywood spent the past six months agonizing over what Sora meant for the entertainment industry. Turns out it was just a side quest.

The Sora app is dead, along with its API, its planned integration into ChatGPT, and the $1 billion Disney partnership that was supposed to let people wield lightsabers with Luke Skywalker in AI-generated clips. The Disney deal was a handshake one and is now no more. The Sora team will pivot to world-simulation research for longer-term robotics work.

This is how fast things move at OpenAI: The company published a blog post updating Sora’s content policies yesterday. When Fidji Simo joined ACCESS just seven weeks ago, she talked at length about Sora and hinted at bringing its video generation to ChatGPT. Despite how the early hype around it had fizzled, she said Sora was “really nailing creation,” and that the company was “excited about what we’re seeing.”

Read more

“My goal is to help them scale”

It took Paul Daversa, one of Silicon Valley’s most connected executive recruiters, to get Sheryl Sandberg on a call with Josh Payne. “I’ve left you alone for a long time,” Daversa told Sandberg. “This is the one to take.” She took it. Her opening line to Payne was, “I don’t do boards.”

A 30-minute call turned into 90. Then came a letter.

Over Christmas in Japan, Payne had written a strategic memo to his executive team laying out how he wanted Nscale to evolve over the next two years. Sandberg asked to see it. “The only person I’d ever seen write like that is Mark Zuckerberg,” she told me during a joint interview with Payne today.

Read more

Happy Friday, and a special hello to all of you who subscribed this week. If you have an extra moment this weekend, please fill out my audience survey.

“I don’t think Claude is going to go build a bank”

Every software CEO right now is wrestling with the same question: how much of your product do you give over to AI? On this week’s episode of ACCESS, we put the question to Mercury CEO Immad Akhund. His answer: “I don’t think Claude is going to go build a bank.” At the same time, he said, “If you try to fight the future, I think you’re going to lose out on the opportunity.”

Mercury has an MCP connector that lets you pipe your banking data into an AI model for tax planning, financial modeling, or whatever you want. It also just launched a product called Insights that builds AI chat directly into its interface. Immad’s argument is that banking has enough depth — regulation, trust, the sheer complexity of actually moving money — that Mercury can afford to embrace AI across the board.

Catch the full conversation on this week’s episode of ACCESS, the podcast I co-host with Ellis Hamburger:

Mercury has reached $650 million in annualized revenue with more than 300,000 customers. It recently applied for a national bank charter after reaching a $3.5 billion valuation. Immad is also one of the more prolific angel investors in tech, with more than 350 investments including Substack, Rippling, Airtable, and Deel. We discuss his thoughts on investing and new startups, and also got into why poker has become the unofficial networking sport of tech.

Highlights from our conversation:

• Mercury is building AI into the product and giving the data layer away at the same time. Immad acknowledged that there are times when talking to your data is the right interface, and times when you want to click around and see information presented visually. “There are geeky people like you who want to learn what MCP is and do it themselves,” he told me, “but other people just want to be able to pull that data directly into the AI directly in Mercury.”

• Companies are whitewashing their layoffs with AI. Mercury is growing the team by 20-25% this year while also automating as much as possible. Immad called out Block’s recent layoffs specifically: “They weren’t growing very fast, and they were like, actually, we just don’t need 40% of people.” The real issue for many of these companies is that they overhired during the pandemic or just aren’t growing, and that AI is a convenient framing.

• Mercury’s approach to marketing is unusual for a fintech. They delivered PB&J sandwiches around SF (the name came from an internal code name for Insights), gave away Casios and AeroPresses at South by Southwest tied to a “craft” theme, and commissioned generative art for each customer based on their transaction data for the Insights launch. Immad’s philosophy: “I’d rather do 10 things, and they’re all kind of fun and interesting, and maybe nine of them fall flat, but one of them is like a big viral hit.”

Listen or watch wherever you get podcasts.

A MESSAGE FROM MY SPONSOR

Built for AI since day one

Nebius is the AI cloud company, designed from day one for what AI developers actually need. From data and model training to production deployment, we deliver a unified platform for AI builders worldwide.

Learn more.

Feed check

Read more

“We’re in a tidal wave of new chips”

Hours after Jensen Huang spent the opening of his Monday GTC keynote celebrating CUDA’s 20th anniversary, calling it the “flywheel” behind everything NVIDIA has become, I found myself in the conference’s packed expo hall with the one person most determined to challenge Nvidia’s dominance.

Chris Lattner was leaning against a demo table at the Nebius booth, where we spoke in front of a live audience for about 30 minutes. The creator of LLVM and Apple’s Swift — both of which underpin much of modern software — joined me to talk about his most ambitious project yet: Modular, and its bid to crack the software moat that keeps the AI industry tethered to one chipmaker.

Read more