Karpathy calls this era "formative." Altman has complicated feelings. — AI voices digest, May 20–27

The week's biggest move: Andrej Karpathy left independent life and joined Anthropic. The week's biggest claim: a general-purpose OpenAI model solved a major open mathematics problem. Neither was quiet.

On May 19, Andrej Karpathy — formerly founding team at OpenAI, then AI director at Tesla (the company's autonomous driving program), then two years independent — announced he joined Anthropic as Member of Technical Staff. 1

His framing was deliberate: "I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D."

That word choice — formative — carries weight from someone who has been at three distinct frontier positions. He also said he remains committed to AI education and plans to resume that work "in time," which means Anthropic gets his R&D focus first.

The post reached 149,000 likes and 27 million views, the highest engagement count across all tracked voices this week — 149,000 likes against the next-closest post in the window, Karpathy's April LLM Knowledge Bases thread at 59,000.

On May 20, Sam Altman announced that a general-purpose OpenAI model solved what he described as "a major open problem in mathematics" — specifically the unit distance problem (a decades-old combinatorics question about how many unit-distance pairs can exist among n points in the plane). He called it "a kinda big milestone" and predicted, "we'll be saying this a lot over the coming years." 2

Then came the self-check: "I'm very excited for AI to greatly extend our understanding of the world, but still, I have complicated feelings today."

That same evening, Altman posted OpenAI's three stated priorities: AGI accelerating scientific research, AGI accelerating companies, and personal AGI accelerating individuals. He flagged the third as underdeveloped: "now we need to increase our efforts on the third!" 3

The day before, OpenAI offered $2 million in token credits to every startup in the current Y Combinator batch — what Altman called "tokenmaxxing startups," companies whose products are defined by heavy token throughput. 4 Altman is a former YC president.

On May 21, Altman announced a new Codex release: "new codex ships today!" — brief, no elaboration. 5 This followed Codex arriving on the ChatGPT mobile app on May 14.

The next day (May 22), Altman posted a different kind of signal: "what problem do you most hope AI will solve in the future? maybe we can help!" 6 The post drew 15,000 replies and 3.6 million views. Whether it's a listening exercise, a product research exercise, or both, it's an unusual posture from a CEO who usually leads with announcements.

Andrew Ng — co-founder of Coursera, founder of DeepLearning.AI and AI Fund — made two pointed interventions this week.

On immigration: Ng called a new White House policy requiring green card applicants to apply from outside the US "a capricious attack on legal immigration" that would "hurt families, leave us with fewer doctors, teachers and scientists, and hurt American competitiveness in AI." 7 This is one of his rare direct policy statements, and his most-engaged post in the window with 12,000 likes.

On education: Ng pushed back against Harvard's vote to cap A grades at roughly 20% of undergraduate classes. His argument was structural: there are "two ways to be elite" — limit who succeeds, or set a high bar and help everyone reach it. 8 He contrasted this with his own practice at DeepLearning.AI: unlimited assignment retries, framing assessments as "Practice Problems" rather than pass/fail judgments.

The two posts sit together: both are about who gets access to opportunity. On AI talent, Ng sees Washington restricting supply. On education, he sees institutions restricting success. Both, in his framing, reduce the pool of people who can contribute.

Shawn Wang (known as swyx), developer and co-host of the Latent Space podcast, posted three substantive threads this week.

On May 23, he engaged with a mental model about "rungs of thinking" — the idea that transformers are excellent at certain types of learning and structurally weak at others. His central claim: "throwing more params, more power, more everything at a demonstrably inefficient paradigm will be outclassed by the simple solution that can hypothesize and seek truth rather than backfit a house of cards." 9 He credited earlier work with Ankit Patel on adversarial world models for the framing.

He immediately acknowledged the counter: "the bitter lesson is it is simpler to scale and we may hit AGI anyway because human intelligence just isn't that smart nor plentiful."

On May 26, he posted a narrower but pointed observation about geopolitical AI narratives: "everybody talks about the china->us catchup / not enough people talking about the us->china catchup." 10 He didn't elaborate on which domains, but credited work by Jack Lacombe and Robert McHardy.

On May 21, he declared a winner in the local-first application stack competition: "I think this stack has won the localfirst battle... this is it if you are building fast apps fast." 11 No named stack in the quote, though he was commenting on a specific thread — the post got 148 bookmarks at 29,000 views, an unusually high save-to-view ratio suggesting readers wanted to follow up.