Few developers are better positioned to judge Apple’s AI reset than Ryan Jones. He’s an Apple Design Award winner who has spent seven years building Flighty into one of the most successful indie app businesses. (And one of my favorite iPhone apps, period.)

Jones and I were both at Apple Park on Monday for WWDC, and his read on the keynote is that it was a cleanup lap. He thinks the new Siri plays to advantages nobody else has, starting with the on-device index of your iMessages, photos, and voicemails that Google, Meta, and everyone else will never get access to. He points to Joanna Stern’s demo, in which Siri surfaced an old voicemail from her uncle as a contextual suggestion, as the kind of moment no other assistant can pull off.

Apple is also still way behind the state-of-the-art AI assistants. The new Siri, for example, has no persistent memory between sessions, and the on-device indexing took days to complete on Jones’s iPhone running the iOS 27 beta.

Jones explained the unique way Apple is letting Siri understand content on a device’s screen. Siri’s context system is built on roughly 100 predefined “entity” types that apps can “donate” data to. He contrasts that rigid, deterministic model with MCP, where you hand over content and let the model figure out what it is. It’s also part of why Flighty runs Gemini server-side instead of Apple Intelligence: when a parse fails, his team can see why and retrain on the edge case, while Apple’s stack is a black box.

We also got into the other big shift in Apple’s AI strategy: its souring relationship with OpenAI. Apple is leaning harder into Google as its model partner while weaponizing its privacy stance against OpenAI, and the arc looks a lot like what happened with Facebook. Deep integration first, mortal enemies a few years later. Around 2019, Apple used privacy to kneecap Facebook, and the exact same playbook appears to be running against OpenAI now.

Jones also revealed what’s next for Flighty: an AI-built connection guide, landing in beta at the end of June, that gives you granular walking directions between gates based on your terminals, seat number, and passport. The infinite matrix of airports, nationalities, and destinations made it impossible to build before, but as he put it, it’s “right up AI’s alley.”

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.

Weekend links

• The SpaceX roadshow presentation

• The New York Times: “About 20 New Billionaires Could Be Minted by 3 Mega-I.P.O.s”

• Wired reports: “‘Tell Him He’s a Piece of Shit’: Meta’s New AI Unit Is a Total Mess”

• Vanity Fair: “Aaron Sorkin’s First Interview About His Social Network Follow-Up”

• Reuters: “Anthropic v. OpenAI: Behind the bitter battle for the future of AI”

• “From AGI to ASI”

• CNBC: “Prometheus is the bulk of my time, says Co-CEO Jeff Bezos”

• Simon Willison: “Claude Fable is relentlessly proactive”

• Financial Times: “AI and the productivity paradox”

• Kayvon Beykpour: “Beyond Tokenmaxxing”

• Jasmine Sun’s latest banger: “It’s a scary time to be twenty-two.”

• “Europe 2031: What getting AI wrong means for us”

• “I looked at 1,680 Anthropic resumes. Here’s who they actually hire.”

• Andy Masley: “Why I think panic about local impacts of data centers is just a panic”

• Matt Stoller asks: “What Would It Look Like If the AI Bubble Popped?”

Catch me on Prof G Markets

Thanks to Ed Elson for having me on this week to unpack WWDC and the tepid investor reaction.

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.

Subscribe now

In addition to my Monday evening of programming with Yahoo, I’m excited to announce that I’ll be participating in a panel titled “Attention, AI and the Future of Marketing” with Joe Marchese of Human Ventures and Bryson Gordon, Microsoft’s CVP of global marketing, at the OpenAI Villa during Cannes Lions. Let me know if you want to attend.

If you believe the recent press cycle, corporate America has decided that tokenmaxxing, or burning through AI tokens as a proxy for productivity, has run its course. After spending Tuesday afternoon with leaders from Databricks, Cognition, Pinecone, LangChain, and others at Nebius’ first Inflection forum in San Francisco, I can tell you there’s no such consensus among the people actually building this stuff.

During his opening remarks, Nebius CRO Marc Boroditsky made the pitch that the industry needs to move from tokenmaxxing to “valuemaxxing”: making tokens count rather than counting tokens. The leaders from other AI companies that Nebius put onstage didn’t exactly agree.

The spending split

During one of the early panels, The Information’s Amir Efrati set the stage by comparing this moment to the early public cloud era, when AWS customers were saying, “Wait a minute, I just spent $20 million more than I expected to this year.” Back then, he noted, $20 million was a lot of money.

From there, the panel split:

• Databricks VP Nikita Shamgunov said customers are demanding cost controls after “the wheels started to come off” of tokenmaxxing, while admitting that inside Databricks, “the budget is unlimited for engineers.”

• Cognition CEO Scott Wu took the bullish view on tokenmaxxing: “The GPUs are expensive, but they’re not that expensive,” relative to paying humans, he argued, and it’s “much, much cheaper for any of these models to write 10,000 lines of code.” His caveat: 10,000 useless lines are “a waste of money entirely,” so the discussion at Cognition is starting to shift to measuring outcomes, not consumption. Wu also argued the cost of waiting to embrace AI has changed: “It’s not really okay to be two years late anymore.” Companies that migrated to the cloud a decade late turned out fine. He doesn’t think AI laggards will get that grace.

• When Efrati asked how much of enterprise AI spending on AI is actually experimental. Wu called it "a pretty small minority at this point," since the workloads that scale are the ones that demonstrably work, with coding leading the way. DataRobot CPO Venky Veeraraghavan, whose company works with traditional enterprises like Chevron, countered that, outside of coding, the supermajority of that spend remains experimental, with companies hunting for a use case that sticks.

The agent horizon

Wu sees the agent trend continuing on its current curve. Task length has gone from 20 seconds of autonomous work two years ago to several hours now. The question is how far that extends. Cognition customers have run Devin sessions for multiple weeks. (”I would not recommend that.”)

He teased what’s next: handing an agent an open-ended initiative, like cutting infrastructure costs, and letting it scope its own project over weeks. When Efrati asked if that was a product preview, Wu didn’t say no.

The economics of all this got a live illustration later in the day, when Nebius' VP of ecosystem strategy, Anubhav Maheshwari, walked through building a healthcare compliance agent. The version running on a closed frontier model cost $637 per run, prompting a Slack message from his engineer: "Is this for real?"

Swapping in open-weight models, first DeepSeek and then Nvidia's new Nemotron, brought a single run down to $24 and cut the runtime from an hour to 15 minutes. His takeaway, and a theme all afternoon, was that open models are closing the gap quickly, and the harness around a model now matters as much as the model itself.

Hallway chatter

A few more things I heard from attendees between sessions:

• Everyone at the compute layer wants to know how much Nvidia Vera Rubin supply they can actually get and, just as importantly, how much their competitors are getting. The sorting process is underway and fairly cutthroat.

• The consensus is that AI compute will remain constrained until next year, with a possible softening mid-2027. Some think relief doesn’t really come until late 2027 or even 2028.

• Google’s reported $920 million-a-month deal for Blackwell capacity from SpaceX is being read as a very high premium on a short leash, with an effectively 90-day rolling commitment. That’s what a tight market looks like.

(Disclosure: I attended Inflection as part of a paid partnership with Nebius, which had no editorial control of this story.)

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

• Anthropic nerfed Fable for distillation. I’ve been seeing lots of chatter about the model card disclosure that Anthropic has beefed up Fable’s resistance to use in LLM development and AI research. A few other Anthropic things:

  • Dario Amodei only has one direct report. There are several interesting nuggets in Emily Chang’s new deep dive on Anthropic for Bloomberg’s The Circuit show, including that Dario’s only report is his chief of staff, while Daniella Amodei manages everyone else.

  • He has a new essay. This one is titled “Policy on the AI Exponential” and starts with a Hobbit reference from Lord of the Rings.

  • Two Fable early-tester vibe checks worth reading: Ethan Mollick’s and the write-up from the crew at Every.

• Sam Altman’s message to OpenAI employees about an IPO is ominous. “Altman said that if the company’s technology advances at a rapid pace to the point where the AI itself is able to create new AI—known as recursive self-improvement—that would lessen the chance of a quicker public listing. ‘The faster the potential RSI takeoff looks like it could be, the more it could be advantageous to delay an IPO,’ because the ‘technology and the world may change in surprising ways, and there might be good reasons to be a private company during that time,’ Altman said.”

• Another new neolab: “Igor Babuschkin, an xAI co-founder who previously worked at OpenAI and Google DeepMind, announced the launch of River AI on Wednesday. The startup, which also includes former xAI and Tesla Inc. staffers, seeks to develop AI agents that learn from users and are controlled by them, Babuschkin said. The AI will understand ‘your style, your preferences, and your goals,’ he said in a statement, and will do ‘the work as if you had been supervising the whole time.’”

• The SpaceX IPO is going to make it rain for Musk’s inner circle. “Antonio Gracias, an investor who has built a $65 billion stake in the company, exemplifies the economic upside of being Mr. Musk’s ally.”

• The Social Network sequel trailer looks corny as hell. Jeremy Strong’s Mark Zuckerberg impression is incredible, but I almost laughed out loud at my desk at the “she’s disrupting” line at the end. I’ll still be seeing this movie in the theater when it premieres, though I’m not expecting it to be anywhere near the caliber of the original.

• The tech and media moguls who were invited to Sun Valley this year.

• I’ve got access to Substack’s MCP server. The team at Substack hooked me up with beta access, which I’ve been using with Fable in Claude Cowork to create in-depth growth and strategy reports. On the go, I can prompt with questions like “What has this person read recently?” or “Who signed up from my last issue?” So far, the Substack MCP can only surface analytics/list data. It’s going to get really interesting if content creation tools are added, but for now, I already feel like I have a superpower.

Share

My TBPN hit

Thanks to John Coogan and Jordi Hays for having me on yesterday to recap WWDC. I probably spilled too much about how I’m thinking about using AI to help distribute my reporting. More on that to come…

Last week on ACCESS

Ellis Hamburger and I sit down with Jordan Taylor, founder of Vizcom, to discuss the AI design tool that lets anyone turn a rough sketch into a finished product. We discuss how Jordan built Vizcom from a viral Reddit post during COVID, why a drawing is still a better creative input than a text prompt, and how hundreds of thousands of designers are already using it to design shoes, cars, and other physical products.

And we, of course, had to talk about the Ferrari Luce.

Listen or watch wherever you get podcasts.

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.

Subscribe now

I’ve been in Cupertino all day at Apple Park for WWDC. My takeaways are below. Keep reading for a scoop on Anthropic's next big model release.

Right after Apple’s WWDC keynote finished, I joined dozens of other journalists, influencers, and analysts in a dimly lit theater at Apple Park to hear an AI 101 class and Q&A with Craig Federighi and other Apple execs. Tim Cook and incoming CEO John Ternus sat in the front row.

Apple has never done an on-the-record press Q&A like this after a big keynote. There was no live broadcasting, but below are my notes and takeaways:

Read more

Shoutout to Blackbird for helping me get around the city and setting me up with some great dinners. And thanks to everyone who came out to the new media party I co-hosted at a16z HQ last night. I’m not a sports guy, but go Knicks.

I’ve got a shorter feed check issue today. I’m headed back to the west coast this weekend for Apple’s WWDC on Monday. Ping me if you’ll also be in Cupertino and want to say hi.

In other news, I’m excited to be co-hosting an evening of conversations about AI, the future of the web, and the creator economy with Yahoo at Cannes Lions in a few weeks, featuring Yahoo CEO Jim Lanzone, Beehiiv CEO Tyler Denk, After School’s Casey Lewis, and Cat Goetze of CatGPT fame. If you’ll be on the Crosette on June 23rd and want to attend, ping me.

Feed check:

Read more

Friday’s scoop revealing Scout, Microsoft’s first proactive AI agent for Copilot, buried the real news that Satya Nadella announced onstage at Build today: Microsoft is fully embracing OpenClaw.

When Scout is released more widely later this summer, it will be powered by OpenClaw, and Microsoft will contribute its security guardrails back to the project’s open-source ecosystem. As Peter Steinberger, the creator of OpenClaw, responded to me on X today, Scout is “not OpenClaw-like, it is the OpenClaw gateway.”

Microsoft getting in bed with OpenClaw makes a lot of sense. You only welcome a growing, open framework onto your turf when you’re confident you control the ground it stands on. In this case, Microsoft is doing what it does best: being a platform company rather than trying to own too much of the stack.

Microsoft also gets to ride the agent wave in a way its main hardware rival won’t. Even with OpenClaw’s initial buzz driving a surge in Mac Mini purchases, it’s highly unlikely Apple will create a white-glove experience for OpenClaw like Microsoft has with Scout and MXC on Windows, which runs OpenClaw in a contained, safer environment. A big question I have now is how much Apple will lean into agents at WWDC next week, if at all…

A few more observations from Build:

Read more

Wojciech Zaremba recently bought a copy of “House on Fire,” a 2011 memoir by epidemiologist William Foege about the campaign that wiped smallpox off the planet. He’s using it as a guidebook for executing what is about to become one of the largest philanthropic efforts of all time.

Zaremba is one of OpenAI’s least well-known co-founders. He has spent more than a decade at the company across a range of efforts, from leading its early robotics efforts to starting the team that guides OpenAI’s personality and what became reasoning models. In March, he left the frontier research world to run AI “resilience” at OpenAI’s nonprofit foundation.

Zaremba and I spoke ahead of a post that the OpenAI Foundation published Monday morning titled “Resilience in the Age of AI,” which names four areas it will initially fund: biosecurity, cybersecurity, model safety, and AI’s effect on kids. After $100 million for fighting Alzheimer’s with AI in April and $250 million for researching “economic futures” last week, the initial $25 billion grant machine Zaremba helps oversee is spinning up.

“Even if OpenAI makes the safest models out there, that doesn’t solve the problem of jobs [going away],” Zaremba told me during an interview. His use of the word “resilience” is telling; it implies that, even with fear of AI at an all-time high, he and OpenAI are comfortable publicly framing AI’s advancement as inevitable and something to be withstood. “There is no way to align fire. There is no way to align electricity,” he said. “We need to be resilient to the outcomes.”

Read more

Microsoft is building a single app that folds its scattered Copilot tools into one place: chat, coding, the Cowork task assistant, and a new agentic layer called Autopilot, per my sources.

Fortune’s Sebastian Herrera first reported on the Copilot super app project this morning. The leaked screenshot above, which I’m publishing for the first time, shows what it actually looks like, including a previously unreported, proactive agent named Scout.

Read more

For the past two decades, Drew Houston has occupied a unique place in the firmament of Silicon Valley. He was the first founder to take a Y Combinator company public, and he famously turned down an early acquisition offer from Steve Jobs. His announced plan this morning to eventually step back as Dropbox’s CEO and become executive chairman comes as AI is rewriting the era of tech he came up in.

In a 2009 meeting at Apple, Jobs told Houston that Dropbox was "a feature, not a product." Houston proved Jobs wrong by turning down a nine-figure acquisition offer and building something that hundreds of millions of people use.

He also half-proved Jobs right. Dropbox never reached the scale of YC contemporaries like Reddit, Airbnb, or Stripe, and file storage has become a tentpole feature within the bigger work suites — Google Workspace, Microsoft 365 — that Dropbox competes with for dollars.

Houston insisted to me that he’s not stepping back now because of any structural changes in Dropbox’s business. "I think five years ago I would have been talking about needing to grow paying users and increase our cash flow," he told me in a joint interview with Dropbox’s new co-CEO, Ashraf Alkarmi. "I'm sure five years from now, Ashraf will be saying the same kind of things."

Read more

I asked Nikesh Arora to put his investor hat back on. He was Masa Son’s president and COO at SoftBank before taking over Palo Alto Networks in 2018. I expected a defense of software. Instead, on whether Wall Street investors are right that AI will gut a chunk of the SaaS world, he gave me an unusually candid read for a prominent tech CEO: “I think they’re getting more right than you think.”

We talked on May 14 at an intimate lunch I co-hosted with AlixPartners and a room of tech leaders in San Francisco. His own stock had been swept into the SaaSpocalypse. Palo Alto dropped about 7% on its February earnings.

Most software, Arora argued during our conversation, is a pile of data wrapped in a user interface that a product manager hand-built to anticipate how you'd want to reach and interact with it. His bet is that the roughly $1 trillion flowing into AI infrastructure over the next year will yield models capable of generating that interface on the fly, collapsing the long tail of hand-designed UIs. "80% of UI will go away," he predicted.

Read more

Brian Chesky could have walked onstage at Airbnb’s Summer Release event on Wednesday, declared the company a MCP node, and thrown the gates open to AI agents.

What he did instead was add to his own interface, rather than give it over to AI. The new Airbnb now sells car rentals, groceries, airport pickups, luggage storage, boutique hotels, and more. While many pre-AI software companies are racing to integrate with AI chatbots, I came away from Airbnb’s event with the impression that the company is trying to become its own island, where travelers do more and more without ever leaving.

After his keynote, I asked Chesky where Airbnb would open up to agents and where it wouldn’t. He said the first thing to settle is whether MCP is even the right standard for connecting with AI, or whether travel needs its own protocol.

Read more

I reported last week that Google’s next model wouldn’t push the frontier and instead land in the ballpark of GPT-5.5 and Claude Opus 4.7. That’s what Gemini 3.5 Flash, released across Google’s platforms on Tuesday at I/O, turned out to be.

Across my conversations with senior executives, I think Google is positioning for a moment when companies review their token bills, ask what they’ve been buying, and look for more efficient alternatives.

The tokenmaxxing hangover

During the main I/O keynote, CEO Sundar Pichai correctly noted that many companies have burned through their annual token budgets (it’s May) and framed Gemini 3.5 Flash as the off-ramp. A trillion-tokens-a-day customer could save $1 billion-plus annually by moving 80% of its workloads to 3.5 Flash, he said.

‘Tokenmaxxing’ has become a meme in Silicon Valley, driven by OpenAI and Anthropic’s API economics, and people who have been spending heavily on tokens are starting to realize that much of it is wasteful. Google clearly sees the hangover coming.

Read more

Sources say that Google plans to announce a new Gemini model at its annual I/O conference on Tuesday. I’m told the release will land roughly in the class of OpenAI’s recent GPT-5.5. and well short of Anthropic’s Mythos, which, despite not being widely available, has reset how every lab talks about what “leading” means.

Read more

Sources say Sam Altman has recently discussed starting a new AI compute company. He’d fundraise for the effort, which would be majority-owned by OpenAI but not anchored solely to its business.

Read more

Read more

Anthropic’s surprise deal to use Elon Musk’s Colossus data center is incredibly awkward. Arguably, it’s also unprincipled.

Read more

On Tuesday evening, I went from an off-the-record happy hour with Anthropic leaders to OpenAI HQ for a party thrown by (and for) its newest model, GPT-5.5. The crossover felt a bit like passing through the DMZ. Several attendees at both events naturally wondered whether Sam Altman was trying to upstage Anthropic’s Claude developer conference the next day.

Read more