In partnership with

In Today’s Issue:

🍉 Meta's 'Watermelon' model quietly catches GPT-5.5

🏨 A hotel where every worker is a robot

⛪ Peter Thiel vs the Pope (and China)

🛡️ AI is finding software flaws at record scale

🗳️ Silicon Valley's nine-figure push into the midterms

✨ And more AI goodness…

⚡ The Signal

Meta told its own staff two opposite things in one town hall, and both are true.

Superintelligence chief Alexandr Wang said Meta's in-training model, codenamed Watermelon, has finally matched OpenAI's GPT-5.5 on key benchmarks. Minutes later, Mark Zuckerberg admitted the agentic AI Meta actually promised investors has "not accelerated in the way we expected." That gap, between a model that can clear benchmarks and a product that can act reliably, is the real story of 2026. Sam Altman and Dario Amodei have both walked back their white-collar job apocalypse timelines, and this year's clearest measurable gains have come in narrow places like security research. The benchmark race is tightening again. The product race is where it now gets hard.

All the best,

Kim Isenberg

(Getty Images)

⛪ Thiel Calls the Pope a Chinese Agent Over AI Rules

Peter Thiel told the Aspen Ideas Festival that the Pope is effectively doing Beijing's bidding on AI. The billionaire investor said Pope Leo XIV is "working for the Chinese Communists" because the pontiff's call to regulate AI would slow only the United States in its race against China. Speaking alongside Francis Fukuyama in a session titled "Humanity at the End of History," Thiel framed AI caution as a gift to Beijing, extending the "Antichrist" theme he has pushed all year: a world government that seizes power by promising to shield humanity from threats like AI. He also warned the crowd of a coming "democratic socialist takeover" of the Democratic Party.

👉 tl;dr: A tech billionaire reframes AI safety as surrender to China, from the Vatican to Aspen.

Learn More

(CNBC)

📉 Palantir's CEO Melts Down Live on CNBC

A CNBC segment meant to cover Palantir's new Nvidia government deal turned into a rambling, nearly 20-minute rant. Palantir CEO Alex Karp accused the industry of building models that have been "completely over, irresponsibly over-sold," and when anchor Becky Quick noted that he sounded angry, he shot back: "No. This is the voice of American business that is channeled through me!" Between the stutters and a digression about wanting to teach at Berkeley, viewers were left unsure whether they had watched blunt candor or a live meltdown.

👉 tl;dr: The loudest AI-bubble warning this week came from one of the bubble's biggest beneficiaries, unscripted.

Learn More

(Pudu Robotics)

🏨 China Opens the First Hotel Run Entirely by Robots

A 44-room hotel on an artificial island off Shenzhen will run with no human staff at all. Every check-in, bag delivery, drink, and cleaning task will be handled by a fleet of robots from Pudu Robotics, coordinated by a shared AI layer called PuduFM 1.0 and PuduAgent that uses vision-language-action models so the machines hand tasks to each other in real time. Pudu's CTO Cong Guo promises "no service gaps and no human interruptions." Trial operations are targeted for late 2026, with first guests in early 2027, a deliberate contrast to Japan's Henn-na Hotel, which pulled its robots after they kept breaking.

👉 tl;dr: Physical AI's most complete test yet is a building that fires no one because it never hired anyone.

Learn More

🛡️ Turn the Bug-Hunter on Yourself

❝

The same kind of model now surfacing software flaws at record scale can also audit your own digital life before someone else does.

Why it helps: You do not need to be an engineer to have an AI reason about where you are exposed, from reused passwords to oversharing apps to a home router last updated years ago.

Try this: "Act as a practical security auditor for a non-expert. Ask me 8 questions about how I use email, passwords, my phone, and home Wi-Fi. Then give me a prioritized, plain-English checklist of the 5 changes that most reduce my risk, and tell me exactly how to do each one."

🎬 Watch This

❝

Yuval Noah Harari, the historian who wrote Sapiens, used his 2026 Tanner Lecture at Oxford to make a quietly unsettling argument that AI is starting to "hack the operating system of civilization." Human power rests on large-scale cooperation between strangers, and that cooperation is held together by bureaucracy, the laws, money, and contracts that let people trust each other. AIs, Harari says, are "native bureaucrats": they can remember every rule, transaction, and precedent far more perfectly than any human, which makes them uncannily suited to quietly run the machinery of modern life, who gets a loan, who gets into university, who gets sentenced, even who a strike targets. It is less a warning about killer robots than about paperwork, and it is worth 40 minutes of your weekend.

Hosted at Linacre College, University of Oxford.

(Getty Images via TechCrunch)

"I'm delighted to be wrong about this. I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened."

– Sam Altman, CEO of OpenAI

Anthropic has reportedly told the US Senate that operators linked to Alibaba ran the largest known "distillation attack" on its Claude models. According to a letter from Anthropic's policy team, the operators used roughly 25,000 fraudulent accounts to pull nearly 29 million exchanges out of Claude between April 22 and June 5, targeting its most valuable skills: agentic reasoning, coding, and long-horizon task completion. A distillation attack extracts answers from a stronger model to train a weaker one, a way to copy capability while sidestepping the export controls that cover raw model weights. Anthropic is urging Congress to penalize the practice; Alibaba has not publicly responded, and the claims have not been independently verified.

Meta's 'Watermelon' Catches GPT-5.5 as Its Agents Lag

❝

The Takeaway

👉 Meta's superintelligence chief Alexandr Wang told an internal town hall that the company's in-training model, codenamed Watermelon, has matched OpenAI's GPT-5.5 on key benchmarks.

👉 Watermelon reportedly uses "an order of magnitude more compute" than its predecessor, codenamed Avocado, and has not yet shipped.

👉 In the same meeting, Mark Zuckerberg admitted agentic AI has "not accelerated in the way we expected" over the last four months.

👉 The backdrop: roughly 8,000 layoffs, 7,000 staff moved into new AI units, and up to $145 billion in 2026 capex.

Meta gave its own staff two opposite messages in a single town hall: a model that has finally caught the frontier, and an agent roadmap that keeps slipping. Meta's superintelligence chief Alexandr Wang said the company's next model, in training under the codename Watermelon, now matches OpenAI's GPT-5.5 on key internal benchmarks, though he did not say which ones. Wang said Watermelon uses "an order of magnitude more compute" than its predecessor, codenamed Avocado. If the claim holds once the model ships, it would be the first time this year Meta has drawn level with the frontier it has trailed behind OpenAI, Google, and Anthropic.

(Alexandr Wang, Meta's chief AI officer. Getty Images via TechCrunch)

The same meeting carried a soberer message. Mark Zuckerberg conceded that the "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," and that Meta's reorganization around AI "haven't come to fruition yet." Earlier this year the company cut about 8,000 jobs, moved more than 7,000 people into four new AI groups, and guided to as much as $145 billion in 2026 capital spending. He admitted the restructuring was not as "clean" as it should have been.

(Mark Zuckerberg. Andrew Caballero-Reynolds / AFP via Getty Images)

Put together, the two messages define Meta's bet. A bigger, far more expensive model can close the benchmark gap. Turning that raw capability into reliable agents, the products Zuckerberg actually promised investors, is the part that keeps sliding. Watermelon is a claim about the model; Zuckerberg's caution is a claim about the product. Both are still unproven, and both now ride on the same $145 billion.

Why it matters: A benchmark tie with GPT-5.5 would pull Meta back into the frontier conversation for the first time in a year, but only if Watermelon actually ships and the agents built on it finally work. For now, Meta is asking investors to fund the gap between the two.

Sources:
🔗 Business Insider
🔗 Reuters via Yahoo Finance
🔗 TechCrunch
🔗 SiliconANGLE

Scale AI support on AWS, see how July 9

Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.

You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.

Save your spot

❝

The chart: Epoch AI plots monthly high- and critical-severity CVEs (publicly disclosed software vulnerabilities) from 21 major vendors, including Adobe, Apple, AWS, Cisco, Google, Microsoft, NVIDIA, and Oracle. For years the high line sat around 150 to 300 a month; after Anthropic's Claude Mythos Preview shipped in April 2026 it spikes almost vertically, with the latest month near 1,150 high and 350 critical, together roughly 1,500.

The lesson: This is the clearest real-world sign yet that frontier models are already superhuman at one narrow thing: finding bugs in code. Anthropic says Mythos Preview alone surfaced more than 10,000 high- or critical-severity flaws, and the disclosure curve now tracks model releases, not researcher headcount.

The caveat: Disclosure is not exploitation, and reporting practices vary a lot between vendors, so part of the jump is better-organized reporting rather than pure new discovery. A found-and-patched bug is defense; the same capability in other hands is offense. The chart shows the tool, not who is holding it.

🗳️ Silicon Valley Buys Into the Midterms

❝

⚡ Bottom line: A pro-AI super PAC network called Leading the Future has built a war chest of over $100 million to shape the 2026 US midterms.

💡 Why it matters: For the first time, the AI industry is spending at real scale to reward or punish politicians purely on how they vote on AI rules.

🔎 What it means: 2026 becomes a test of whether AI policy gets written by voters and regulators, or by the people building the models.

The AI industry has discovered retail politics. Leading the Future, a super PAC backed chiefly by venture firm Andreessen Horowitz and OpenAI president Greg Brockman, has assembled a war chest reported at over $100 million for the 2026 midterms, and has already spent more than $20 million across races from Texas and Georgia to Illinois and Montana. Marc Andreessen and Ben Horowitz each put in $12.5 million, and Brockman and his wife matched them.

(Marc Andreessen, co-founder of Andreessen Horowitz. Getty Images via TechCrunch)

Its goal is blunt: oppose any politician whose agenda would "stifle innovation" or "enable China to gain global AI superiority." The playbook borrows from crypto's Fairshake PAC, which spent heavily in 2024 and won, and the message to lawmakers is that a wrong vote on AI now carries a price at the ballot box.

(Anthropic CEO Dario Amodei, whose company funds the opposing camp.
Getty Images via TechCrunch)

The industry is not united. Anthropic has gone the other way, backing a group called Public First Action with $20 million on the argument that most Americans want more AI guardrails, not fewer. So the fight is no longer just AI companies versus Washington; it is one AI faction spending against another over how hard the technology should be regulated, even as most of the resulting attack ads never mention AI at all.

Learn More