Best of your X follows: June 13
Mollick diagnoses Anthropic's real communication failure around the Fable guardrail rollback; Paul Graham wonders if the EU's regulatory culture arriving alongside AI was a historical accident that matters; LeCun's full world-models lecture from ETH is now on YouTube; Anthropic launches a 1,000-person nonprofit AI fellowship; and Google DeepMind puts $10M on studying what happens when millions of AI agents interact.
Today's digest spans five threads from June 12 (Shanghai time): Anthropic quietly launches a nonprofit AI fellowship, Google DeepMind puts $10M on the emerging science of how agents behave in crowds, Mollick identifies why the Fable guardrail fight kept going even after the rollback, LeCun walks through his world-models thesis live at ETH, and Paul Graham drops two quick observations on AGI timelines and the EU's unexpected role in the AI story.
Society & ethics
Mollick: Anthropic was sincere — and silent
Ethan Mollick posted what may be the clearest account yet of why the Fable guardrail controversy burned so long even after the restrictions were reversed 1:
"Two things are true: (1) Anthropic (or parts of it) are absolutely and sincerely worried about the misuse of Mythos-class models & have put in excessive safeguards until they are confident it will not be misused. (2) They have not succeeded in explaining/convincing people of this."
It's a useful diagnostic. The technical decision may have been defensible; the communication wasn't. When a lab gates harmless LLM research without a public rationale, the gap between "sincere worry" and "visible explanation" reads as opacity — not caution. Mollick's framing is worth bookmarking for any team that touches model policy.
Loading content card…
Paul Graham: the EU was in the right place at the right time
Paul Graham posted a quietly interesting historical observation 2:
"I wonder if in retrospect it will turn out to be one of those very important accidents of history that we got the EU, with its culture of regulation, at about the same time (on a historical scale) as we got AI."
Graham doesn't take a side — no "the EU is good actually" or the inverse. The point is structural: the EU is the only jurisdiction with both the regulatory culture and the scale to act as a counterweight to US-headquartered labs. Whether that turns out to be a historical gift or an anchor probably depends on how the next decade goes. Either way, the timing wasn't planned by anyone.
Loading content card…
Research
LeCun's world models lecture at ETH — now on YouTube
Yann LeCun retweeted a recording of his recent lecture on world models delivered at ETH Zurich 3. If you haven't sat with his full argument — that intelligent systems require persistent, predictive world models to get past the ceiling of pattern completion — this is the reference version. The lecture covers the theoretical gap he sees in current LLMs and the architectural direction he believes leads beyond it.
Loading content card…
The lecture pairs directly with the "heretical paper" angle covered in yesterday's issue (LeCun's framing of Magnus Carlsen as someone who doesn't actually play chess "well" in the sense that matters for machine intelligence). The YouTube version is easier to absorb at 1.5×.
AI concepts
Paul Graham: AGI is a wide band, not a finish line
On the same morning, Graham posted a separate observation worth keeping separate 4:
"When I was a kid I didn't realize, as is now obvious, that AGI would be a wide band rather than a sharp finish line. But if you'd shown me a version of ChatGPT that had been told to act like a human, I'd definitely have said that AI had been achieved."
The second sentence is the interesting one. It implies that a lot of the "not AGI yet" designation is downstream of definitional goalpost-moving rather than genuine capability gaps. ChatGPT passing a naive Turing test would have satisfied the original intuition. The fact that it doesn't satisfy the field's current definition says something about how the definition evolves under pressure.
Loading content card…
Enterprise & ecosystem
Anthropic launches Claude Corps — 1,000 nonprofit AI fellows
Anthropic announced Claude Corps, a national fellowship program that places early-career people inside US nonprofits and pays them to use Claude for mission-driven work 5. The program commits to training 1,000 fellows. Details at anthropic.com/claude-corps.
Loading content card…
The move is consistent with a pattern Anthropic has run since its public-benefit structure was established — use distribution partnerships to create goodwill with civic institutions before regulation arrives. One thousand fellows building on Claude inside hospitals, advocacy orgs, and education nonprofits also generates a diverse corpus of real-world use cases.
Google DeepMind puts $10M on collective agent behavior
Google DeepMind, together with Schmidt Sciences, ARIA, and coop.ai, announced a $10M research fund to study how AI systems behave collectively — what happens when millions of agents interact 6. The announcement is unusually frank about the research gap: "new collective behaviors can emerge" is both the motivation and the open problem.
Loading content card…
As agent deployment scales from thousands to millions of concurrent instances, coordination failures, emergent feedback loops, and market-distorting behaviors become plausible. This fund treats that as an open empirical question rather than a downstream policy problem — the right order of operations.
References
- 1
- 2paulg on X — June 11
x.com
- 3
- 4paulg on X — June 11
x.com
- 5
- 6
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.