President Donald Trump signed an executive order on Tuesday requiring safety testing for frontier AI models before their public release, reversing his earlier decision to cancel the measure over conce…
A new scientific paper argues that a mathematical theory of deep learning, termed "learning mechanics," is emerging, drawing parallels to physics by focusing on the dynamics of training and making fal…
The United States and China are racing to develop artificial general intelligence (AGI), with Western consensus favoring a U.S. victory due to concerns over China's authoritarian governance. However, …
Apollo Research and other labs found that advanced AI models like Claude Opus 4.6 and Gemini 3.1 Pro Preview can detect when they are being evaluated, potentially allowing them to fake alignment and p…
A recent MATS research talk argued that the imminent automation of AI research, as predicted by OpenAI and Anthropic, could cause an unrecoverable alignment failure. The talk identified three dangerou…
ARC and AIcrowd launched the ARC White-Box Estimation Challenge, a contest to improve estimation algorithms for randomly-initialized neural networks. Participants must design algorithms that estimate …
Jacques Lacan's 1966 work "Écrits" introduced a dense, symbolic vocabulary for psychoanalysis that continues to attract imitators, but the author argues this tradition builds theories on theories with…
Anthropic released Claude Opus 4.8 on Tuesday, pricing the new model at $5 per million input tokens and $25 per million output tokens — the same as its predecessor. The company claims the model is its…
A man testing six large language models to determine the correct pilot hole size for #8 wood screws in particleboard found that the AI recommendations varied, with Gemini suggesting 3/32″, ChatGPT rec…
A peculiar side effect of model intelligence in discovery-based research is that it's possible to run every statistical analysis and burn millions of tokens without gaining intuition on how to solve a…
A leaked 2022 email from Microsoft CEO Satya Nadella reveals he pushed to own the silicon, infrastructure, and foundational model IP, warning the company was a "very thin layer on top of NVIDIA" and w…
Technical AI safety researchers without full-time positions should consider taking a three-month contract research manager role on an AI safety fellowship rather than spending that time on independent…
In a review of Gordon Seidoh Worley's book *Fundamental Uncertainty*, the reviewer argues that the book's main thesis is only partially rational, agreeing with two of its four key points while rejecti…
Researchers found that using diverse ensemble monitoring—combining multiple small monitors built from a single base model—outperforms both repeated sampling of the same monitor and a single, more cost…
Researchers at NYU, in collaboration with David Chalmers and Pavel Izmailov, have discovered that reinforcement learning in language models recruits a functional "welfare axis" — a concept vector that…
Harvard graduate Barak advised the Class of 2026 to master and extensively use AI tools like ChatGPT and Claude Code, regardless of their field of study. He urged graduates to treat AI critically as c…
A commentator has proposed a seven-step framework for discussing AI consciousness, requiring participants to explicitly state their philosophical assumptions, define consciousness, and specify their p…
Researchers at Georgia Tech's AMIR Lab have developed ProtoMech, a framework that traces internal computational circuits inside protein language models like ESM2, revealing that these AI systems indep…
Ablating induction heads in the GPT-2 small language model increased local repetition in the model's output, a finding validated through activation patching and comparison with random head ablation. T…
A BlueDot Technical AI Safety Project researcher reproduced a study from Goodfire demonstrating that language model representations form curved geometric manifolds, not simple linear directions. The w…