Loading activity…
@jsevillamol
CEO of @EpochAIResearch. Trying to glimpse the future of AI.
AI
The v1.0 benchmark spans 11 domains and 206 tasks, with a score that averages performance across domains.
AI
Epoch AI says its fastest exploit attempt took 12 hours and the exploits cost about $1,000 per compromised version.
AI
Epoch AI estimates 78% of new AI compute went to the US in the first half of 2026.
AI
A backer says the environments target aerospace, mechanical and electrical engineering, drawing a parallel with coding environments.
XAI
After two years at Epoch AI Research, the researcher says their METR role involves embedded assessments inside AI developers.
AI
Epoch AI Research says four more benchmarks lacked sufficient information for a verdict, and reviewer capacity limits further reviews.
AI
The Wall Street Journal reports that some researchers enabled disappearing messages in Signal chats with an outside safety researcher.
AI
A post doubts the tech judgment of those who only cared after ChatGPT; a reply calls that “very unfair.”
AI
A player says the models make clear mistakes about once per act and calls their deck-building choices subpar.
AI
The poster says “Mythos-level” open models appear on track by the end of 2026.
AI
A user shared a photo they described as backstage at the panel and thanked Ron and the team for hosting.
AI
A post says the improvement has come since its author and others began streaming the game.
AI
A user says David is working on the program and encourages others to ask him about it.
AI
Epoch says Claude Fable 5.1 and GPT-6 Astra lead, but remain far from fully automating its work.
AI
In a reply, a user says the figure is “quite within the scope” of an Epoch post, ending with “(?).”
AI
Epoch AI fitted weekly figures through mid-August; Rohan Paul says the result reflects API list prices.
AI
Tavus says 48% of people who talked to Griffin live thought it was human, compared with under 3% for previous systems.
AI
A user calls the quoted post a “good TL;DR” of what they’re seeing in AI.
AI
A reply argues that buggy reinforcement-learning tasks can make low-grade hacking the fastest path to a solution—and that those behaviors can persist when models help build and check later training tasks.
AI
An announcement says the foundation is working with Macroscopic Ventures and looking to fund up to $10 million in grants.
