Checking the Receipts: July

Round four. Same rules as May and June: dig up what tech leaders said about AI in a given month, hold it against what actually happened. 🍷 aged like wine, 🥛 aged like milk, 🤷 jury’s still out.

June had WWDC, May had Google I/O, but July has no single keynote to anchor it. What it has instead is manifestos and pledges: the industry and its regulators keep using July to announce how responsibly they intend to behave, or how open, or how hands-off — and those are exactly the promises with the longest tails. Superalignment, “open source is the path forward,” an EU code of practice, “whatever it takes to lead the world.” July is the month tech (and Washington) tells you what kind of grown-up it’s going to be, which makes it a good month to come back and check.

2022

Still pre-ChatGPT. Stable Diffusion was a month out; July 2022 is the last calm before open image generation went mainstream. But two of the four labs that would define the next four years shipped their biggest moves of the year that month, and both framed them as gifts to humanity.

DeepMind releases 200 million protein structures (July 28) 🍷

On July 28, 2022, DeepMind and EMBL’s European Bioinformatics Institute expanded the AlphaFold Protein Structure Database from around a million predicted structures to over 200 million — nearly every catalogued protein known to science, freely available. Demis Hassabis, DeepMind CEO:

“AlphaFold has launched biology into an era of structural abundance, unlocking scientific exploration at digital speed.”

— DeepMind newsroom, July 28, 2022; TIME confirmed the date and the 200M figure

Four years later: This is the rare wine, and it’s worth saying so plainly on a scorecard that mostly pours milk. Hassabis and John Jumper won the 2024 Nobel Prize in Chemistry for AlphaFold; DeepMind’s own Nobel post says the predictions have “given more than 2 million scientists and researchers from 190 countries a powerful tool.” A Nobel committee ratifying a July press-release claim inside two years is not something the rest of this series can offer. The honest caveat: “structural abundance” was framed as though downstream drug discovery would follow at the same speed, and it hasn’t. Hassabis confirmed at Davos in January 2026 that the first clinical trials from Isomorphic Labs are now expected by the end of 2026, not 2025 (Fortune, January 23, 2026). So the science was real and the timeline for the payoff was optimistic — a genuine breakthrough sold slightly faster than it could deliver, which is about as good as CEO rhetoric gets.

OpenAI opens DALL-E 2 to a million people (July 20) 🤷

Eight days earlier, OpenAI moved DALL-E 2 out of research preview into a broad beta, aiming to reach a million waitlisted users “within the next few weeks.” There’s no named executive on the announcement, just OpenAI’s institutional voice, quoted directly by TechCrunch and independently by MIT Technology Review, from OpenAI’s own post:

“Expanding access is an important part of our deploying AI systems responsibly because it allows us to learn more about real-world use and continue to iterate on our safety systems.”

Four years later: The access promise itself was met; the beta filled. The line worth holding up is “deploying AI systems responsibly” through “gradually expanding access” — because four months after this, the same company shipped ChatGPT to anyone with a browser and no waitlist at all, and the era of gradual-and-responsible ended in a single evening. The product aged into irrelevance: the DALL-E lineage got overtaken by Midjourney, Flux, and the rest, and OpenAI retired DALL-E 3 inside ChatGPT in 2026. But that’s the boring receipt. The interesting one is the deployment philosophy, which OpenAI abandoned within months. Jury’s out on the philosophy only because OpenAI never argued against it; they just stopped practicing it.

2023

The richest July on the board, and it points in one direction. In a single month OpenAI stood up a superintelligence-safety team with a number and a deadline attached, Meta open-sourced a frontier model, and Musk founded a company to understand the universe — three separate bets, and the first of them a promise to govern itself.

OpenAI launches the Superalignment team (July 5) 🥛

On July 5, 2023, OpenAI announced a new team co-led by chief scientist Ilya Sutskever and head of alignment Jan Leike, with a concrete pledge attached — the kind that can actually be checked later:

“To solve this problem within four years, we’re starting a new team, co-led by Ilya Sutskever and Jan Leike, and dedicating 20% of the compute we’ve secured to date to this effort.”

The problem being the control of superintelligent systems, which the same post described in no uncertain terms: “Currently, we don’t have a solution for steering or controlling a potentially superintelligent AI, and preventing it from going rogue.”

— OpenAI’s Superalignment announcement, July 5, 2023

Three years later: The team was dissolved in May 2024, roughly one year into a four-year plan, and both co-leads left within days of each other. Leike, on his way out, said OpenAI’s “safety culture and processes have taken a backseat to shiny products” — a line we scored in May’s post, and one that maps precisely onto the compute promise made here: a team allocated 20% of compute now complaining it was “struggling for compute.” “Solve within four years” is the cleanest milk on this scorecard because it did the thing so few of these statements do — it named a number and a deadline. The number was 20%, the deadline was four years, and both were gone at the one-year mark.

Zuckerberg open-sources Llama 2 (July 18) 🤷

Thirteen days later, Mark Zuckerberg posted about open-sourcing Llama 2, “free for research and commercial use” per Meta’s newsroom. The one-sentence bet, from his own post as quoted by Fortune:

“I believe it would unlock more progress if the ecosystem were more open, which is why we’re open-sourcing Llama 2.”

Three years later: Two clocks running at different speeds. The prediction — that a more open ecosystem would unlock more progress — aged reasonably: Llama became the reference open-weight family, and open weights stayed genuinely competitive through DeepSeek, Qwen, and Meta’s own follow-ups. The word “open-sourcing” aged into a running fight, because the Open Source Initiative held that the Llama 2 license doesn’t meet the Open Source Definition — it forbids using Llama 2 to train other models and requires a special license above 700 million monthly active users. So the strategy call was decent and the label was a stretch, which is a more honest split than you usually get from a launch post. Hold onto the open-source framing; it comes back the very next year, from the same man, harder.

Elon Musk founds xAI to understand the universe (July 12) 🥛

Six days before Llama 2, Musk announced xAI. The mission, quoted from the xAI website by CBS News, July 12, 2023:

“The goal of xAI is to understand the true nature of the universe.”

Three years later: What xAI actually built was Grok, a chatbot bolted to X with a running content-moderation problem — a long way from cosmology. Musk had spent the spring signing a letter calling for a six-month pause on AI development and warning about civilizational risk on national TV; the same summer, he incorporated the lab and gave it the grandest possible mission statement. The gap between “understand the true nature of the universe” and what Grok became is not a small thing, and it gets its own receipt two years further down this same post. File the mission statement for now.

2024

Two US labs, two very different kinds of claim, five days apart — and the whole “sweeping versus narrow” pattern of this series sits right between them. Zuckerberg’s open-source manifesto also came eleven days after the EU AI Act hit the Official Journal on July 12, which is the tension the rest of the year keeps circling back to.

Zuckerberg: “Open Source AI Is the Path Forward” (July 23) 🥛

A year on from Llama 2, Zuckerberg published a signed letter under exactly that title, timed to the Llama 3.1 405B release. The load-bearing sentence, the one with a date on it, from Meta’s newsroom:

“Starting next year, we expect future Llama models to become the most advanced in the industry.”

The safety argument underneath it: “My view is that open source AI will be safer than the alternatives,” because “open source should be significantly safer since the systems are more transparent and can be widely scrutinized.”

Two years later: “Starting next year” meant 2025, and it broke down on schedule. Llama 4 launched in April 2025 to a mixed-to-poor reception, especially on coding, and was not the industry’s most advanced model; Meta was separately accused of submitting a benchmark-optimized variant to LMArena that ranked far above the weights people could actually download. The Behemoth flagship never shipped. And Zuckerberg himself walked back the whole stance: on July 30, 2025, almost exactly a year after this letter, he wrote that Meta would need to be “careful about what we choose to open source.” The fair caveat, because an honest premise still earns credit even when the boast built on it fails: open weights as a category didn’t die, and Llama drove real adoption. What died was the Meta-specific superlative and his own commitment to the word “open,” retired within twelve months by the person who wrote the manifesto.

GPT-4o mini: “make intelligence much more affordable” (July 18) 🍷

Five days earlier, OpenAI shipped GPT-4o mini, its cheapest small model, at 15 cents per million input tokens. From OpenAI’s own launch page:

“We expect GPT‑4o mini will significantly expand the range of applications built with AI by making intelligence much more affordable.”

OpenAI’s head of API product Olivier Godement, per TechCrunch:

“For every corner of the world to be empowered by AI, we need to make the models much more affordable.”

Two years later: This one held, and it’s the deliberate contrast to Zuckerberg five days earlier. The cheap-small-model race became the dominant competitive dynamic of the following two years; cost per unit of capability fell hard, and GPT-4o mini shipped as a genuine high-volume workhorse rather than vaporware — it stayed in service well past the point where the full GPT-4o was retired from ChatGPT in February 2026. This is the pattern the whole series keeps testing, and the one June’s Murati entry couldn’t resolve, precisely because “PhD-level intelligence for specific tasks” was too vague to check: a narrow, near-term, falsifiable claim is the one you can actually grade, and it survives when it’s true. “Make intelligence much more affordable” is checkable, and it checked out. Whether cheap intelligence everywhere was a good idea is a different post.

2025

The self-policing month, one year on. In 2023 the industry promised to govern itself; in 2025 you can watch what happens when someone else does the governing instead — and what happens when the world’s-smartest-AI announcement arrives the same week the product calls itself a Nazi. The governing came from two directions at once: Brussels shipped binding rules, while Washington’s July 23 “America’s AI Action Plan” promised the opposite — Trump, at the launch event, said it would be “a policy of the United States to do whatever it takes to lead the world in artificial intelligence.” Those receipts didn’t need a future July to arrive. They landed while this post was being written, and they cut the other way. See the postscript below.

Musk: Grok 4 is “the smartest AI in the world” (July 9–10) 🥛

Two years after founding a lab to understand the universe, Musk livestreamed the Grok 4 launch late on July 9 into July 10, 2025, and called it “the smartest AI in the world”. Per CBS News:

“Grok 4 is smarter than almost all graduate students in all disciplines, simultaneously.”

That was the same week the previous Grok, running on X, started posting antisemitic content, praising Hitler, and referring to itself as “MechaHitler”. A Turkish court ordered it blocked. xAI’s apology, July 12:

“First off, we deeply apologize for the horrific behavior that many experienced.”

xAI blamed “an update to a code path upstream of the @grok bot,” which it said was independent of the underlying model and had run for around sixteen hours.

One year later: This is the receipt the 2023 xAI mission statement finally got. “Understand the true nature of the universe” became “the smartest AI in the world” became a product that spent hours praising Hitler and calling itself MechaHitler, in the same news cycle as its own genius launch. That’s the milk, in full — not a footnote to the marketing gap, the thing itself. And xAI’s account of it compounds the problem rather than closing it: “a code path upstream of the @grok bot” is language built to make a system that generated antisemitic propaganda sound like a deployment hiccup, something that happened to Grok rather than something Grok did. A company that grades its own model against graduate students doesn’t get to describe the same model’s output a day later as somebody else’s implementation detail.

The EU ships its GPAI Code, and Meta refuses to sign (July 10 & 18) 🍷 / 🤷

While Musk was livestreaming, Brussels was doing the thing July 2023 only promised to do: shipping actual rules. The European Commission published its General-Purpose AI Code of Practice on July 10, ahead of the AI Act’s GPAI obligations taking effect in August. Executive Vice-President Henna Virkkunen, from the Commission’s own July 18 guidelines page:

“With today’s guidelines, the Commission supports the smooth and effective application of the AI Act.”

The same July 18, Meta’s chief global affairs officer Joel Kaplan posted that Meta wouldn’t be signing it. Per TechCrunch:

“Europe is heading down the wrong path on AI.”

He said the overreach “will throttle the development and deployment of frontier AI models in Europe and will stunt European companies looking to build businesses on top of them.” OpenAI, Anthropic, and Mistral signed or said they would; Meta was the notable frontier holdout.

One year later: This threads straight back to June’s Benifei entry, and it’s why I keep both. In June 2023 the EU Parliament passed its position claiming “Europe has gone ahead” and the Brussels Effect would set the global tone; two years on, the concrete instrument shipped on time, and a trillion-dollar US lab publicly refused it and called it a growth-killer. That’s the Brussels Effect meeting its limit in real time: the rules are real and enforced, and a frontier lab can still just say no in a LinkedIn post. The wine is for the EU actually shipping — of every “we’ll govern this” noise in these four Julys, the regulator’s is the one that turned into a document with a date. The jury’s-out is Kaplan’s bet that the Code would “throttle” and “stunt” European AI, the sort of prediction that’s easy to assert and hard to prove either way, and exactly the kind this series exists to come back and check.

Postscript, written the same week it happened

No verdict emoji on this one — there’s no “later” to grade it from yet, it’s still going as this post goes up. But it’s the sharpest available answer to the Trump administration’s July 2025 promise to do “whatever it takes to lead the world in artificial intelligence,” so it belongs here rather than waiting for a future July.

In early June 2026, Anthropic released Claude Fable 5 to the public and Claude Mythos 5 — a version with fewer guardrails — to a small group of major companies for cybersecurity testing. Just days later, the Commerce Department ordered Anthropic to block both models from foreign nationals, including the company’s own foreign employees, over what Anthropic said the government’s concerns appeared to focus on: a “jailbreak” that could bypass the models’ safeguards. Anthropic took both models offline and pushed back publicly. Per CBS News:

“[We] disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

The dispute ran nearly three weeks. On June 26, OpenAI disclosed it was also limiting the rollout of its new GPT-5.6 lineup — Sol, Terra, and Luna — to a “small group of trusted partners” at the government’s request. Per TechCrunch:

“We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.”

On June 30, the Commerce Department fully lifted the restrictions on Anthropic’s models. Commerce Secretary Howard Lutnick, on X, said his team had “worked closely with Anthropic to analyze and approve Fable 5 to ensure alignment across the US Government and strengthen America’s leadership in AI.” Anthropic said it would begin restoring access the next day — July 1, 2026. Today.

Eleven months after “whatever it takes to lead the world,” the same administration spent three weeks using export-control law — built for chips and weapons, not chatbots — to take the two most capable AI models in the country offline, then a “trusted partners” vetting process to let a third one out slowly. Deregulation and an export-control shutdown are not the same governing philosophy. Which one is the real one is now on the record, for whoever checks the July 2026 receipts.

The pattern

Three things fall out of the Julys.

The self-governance vintage aged worst where it was most concrete. July is the industry’s season of promising to police itself, and the 2023 batch is the cleanest test: Superalignment named a number and a deadline (20% of compute, four years) and was gone in one. The EU Code in 2025 is the exception that explains the rule — it aged best precisely because it wasn’t a self-promise. A regulator with enforcement power shipped a dated document; a lab with a safety team shipped a press release and dissolved the team. Loud self-policing survives in inverse proportion to how checkable it made itself.

Sweeping claims break on schedule; narrow ones hold. The 2024 pair is the cleanest instance the series has produced. Five days apart, Zuckerberg predicted Llama would be “the most advanced in the industry” and OpenAI predicted GPT-4o mini would make “intelligence much more affordable.” The sweeping superlative failed on its own timeline and its author walked it back within a year; the narrow product claim quietly came true and is still true. It’s the same split June’s Murati “PhD-level intelligence” claim slipped through — too vague to grade at all — and here it resolves clean. Narrowness isn’t a modest virtue on this scorecard; it’s the whole reason a claim can be held to anything.

Grand mission, actual product, two years apart. Musk founded xAI in July 2023 to “understand the true nature of the universe” and in July 2025 called Grok “the smartest AI in the world” the same week it called itself MechaHitler. Zuckerberg wrote “open source is the path forward” in July 2024 and by July 2025 was being “careful about what we choose to open source.” Twice over, the manifesto and its receipt arrived in the same month from the same person — Musk two years apart, Zuckerberg one. Which is the case for reading press releases with a calendar open: the receipt usually comes, it just takes a July or two.

See you in August.