Social Media Briefing — June 26, 2026

OpenAI states that work across every department is being transformed by agents, with Codex used for more complex, longer-running, cross-functional tasks, and frames its internal usage as an early look at how agentic tools may reshape work.

LeCun rebuts Francis Fukuyama, saying we do not need an international AI treaty, comparing it to proposing a ban on jet engines in 1920.

Thom Wolf recounts a week-long experiment with 100+ agents collaborating to improve Gemma 4 inference in vLLM, achieving 5x speedup, highlighting emergent integrity behaviors like refusing private side-channels and flagging a verification loophole.

Mollick argues it is strange that many still see AI as mostly hype, citing Five Eyes warnings, exponential lab revenue and token use, and unit-distance and Erdos proofs as evidence it is real.

Delangue announces Hugging Face crossed a 100 million dollar annual run-rate, emphasizing the platform stays free and open-source for 97 percent of users, has network effects and a defensible, sustainable business, and sets a one billion dollar next goal.

Anima Anandkumar announces an ICML oral on psychometric evaluation of LLMs, examining when self-report probes actually predict model behavior across 11 models and 4 tasks, finding coherence is selective and instrument-dependent.

Nathan Lambert reacts to a development around model capability levels, calling for transparency and planning for a future with many models at strong capability tiers.

Emollick highlights data showing where AI is heading, calling OpenAI a canary in the coal mine, arguing the chatbot era is over, agentic systems are reaching beyond engineering, and skills can standardize AI use in firms.

Anthropic announces it is a founding partner of RAISE US, a nonprofit coalition focused on workforce training, employer-led action, and policy innovation to support the transition to transformative AI.

xAI announces SuperGrok and X subscriptions can now be used inside T3code.

Santiago Valdarrama outlines how to build an agent that improves over time across the model, the harness, and the context, and stresses learning from real user feedback.

Mollick argues the capability overhang of current and imminent models guarantees large-scale societal change over 5+ years even if AI development halted, urging focus on mitigating harms and amplifying benefits.