Social Media Briefing — April 21, 2026

Soumith Chintala's major thread critiquing the Jensen/Dwarkesh podcast. Argues Jensen understands ecosystems and real-world AI diffusion while Dwarkesh parroted AGI party talking points. Critiques the notion that Claude Mythos is a critical turning point, calling it an extension of open-source + more compute. Warns that AGI cult thinking in AI research community will negatively influence policy. Emphasizes measured, continuous policy over overreaction.

Anthropic announces expanding collaboration with Amazon to secure up to 5 gigawatts of compute for training and deploying Claude, with nearly 1 GW expected by end of 2026.

Anthropic announces Amazon is investing an additional $5 billion, with up to $20 billion more in the future, expanding their partnership.

Continuing from Social two days ago, LeCun extends his critique, Yann LeCun pushes back on AI scientists (including Geoff Hinton) and AI CEOs making predictions about labor markets, arguing people should instead listen to reputable economists like Acemoglu, Brynjolfsson, Autor, etc.

Nathan Lambert's detailed analysis of the open-closed model performance gap. Concludes open models can fast-follow closed labs with ~6 month delay. Considers benchmark evolution, real-world performance, and training regime changes. Warns that if closed labs integrate proprietary user data, they could pull ahead.

Simon Willison discovers Claude Opus 4.7 uses 1.46x tokens for text and up to 3x for images compared to Opus 4.6 - same per-token pricing means significant effective price increase

Following yesterday's Social discussion about Codex's expanding role, Greg Brockman announces Chronicle, an experimental Codex feature that gives it continuous visual context of what the user sees, providing automatic memory of user activity. Describes it as 'surprisingly magical.'

Eliezer Yudkowsky writes a detailed essay on why 'Persona Selection' doesn't solve AI alignment. Argues that modeling honest characters doesn't make an LLM honest (the actress analogy), and that RL training for test-passing creates misaligned preferences that override honesty circuits. Questions why conjuring committed honesty from training data isn't trivial if persona selection worked.