Google DeepMind researchers investigate why filtering supervised fine-tuning (SFT) data fails to remove safety-relevant properties from language models, proposing a method to identify the source of th…
A global pause on advanced AI development is unlikely to happen before superintelligent AI (ASI) emerges, argues a commentator, because ASI risk is abstract and hard to demonstrate, unlike nuclear wea…
A new study finds that stronger AI models can imitate weaker predecessors' mistakes only in narrow cases, such as GPT-5.4 mimicking GPT-4o on math questions, but generally fail to impersonate specific…
The IETF and other standards bodies are accelerating efforts to establish agent identity standards, addressing challenges like static authorization grants for dynamic needs. Anthropic supports Workloa…
A researcher found that a clean implementation of abliteration on Qwen3.5-27B costs only about 1.4 TruthfulQA points, far less than the 5.5+ points lost by HuiHui AI's crude method, confirming that mo…
AI safety advocates should work at companies deploying AI, not just at frontier labs, to mitigate real-world harms. The author argues that safety is a relationship between a system and its deployment …
Researchers developed custom Triton kernels that accelerate JumpReLU Sparse Autoencoder inference by 2–14× on real SAEs, exploiting activation sparsity to skip zero entries during matrix multiplicatio…
Shaping AI outcomes is like poker: high variance, small edges, and low skill expression, making it impossible to judge progress by results. They advocate for a diversified portfolio of bets rather tha…
Anthropic is treating the possibility of AI welfare seriously, testing its Claude models for signs of morally relevant internal states like negative self-image, but critics argue the tests may conflat…
A researcher trained a cheap Gemma 2B judge to detect misalignment in AI agents, but testing against Anthropic's AuditBench showed the judge failed to reduce audit costs or reliably distinguish misali…
A new argument against trying hard in longtermist areas, particularly AI safety, suggests that intense effort distorts epistemics and increases the risk of negative impact due to high uncertainty. The…
Continual learning in LLM agents could cause goal and value changes after deployment through loss of developer control, value systematization, and memetic spread, while also eliminating the last-mover…
Ted Chiang argues in The Atlantic that large language models are not conscious, using an intuition pump comparing chatbot personas to fictional characters. The author of this response agrees with Chia…
The term 'AGI' has become nearly useless as AI systems now perform real economic work, falling inside a fuzzy conceptual cloud where they surpass some definitions of AGI while falling short of others,…
Google DeepMind researchers found that supervised fine-tuning (SFT), not reinforcement learning, drives most safety properties in Gemini models. Comparing SFT-only versions of Gemini 3.1 Pro and Gemin…
Alex Bores is running for Congress in NY-12, with voting beginning today. The author suggests that instead of solely supporting Bores, AI safety advocates should ally with anti-data-center movements t…
A proposal for an AML-like system to monitor AI infrastructure using open-source intelligence aims to detect hidden training runs and rogue data centers, potentially serving as a verification mechanis…
A new METR paper finds that frontier AI models like GPT-5.5 can complete tasks that take humans about three minutes without any chain-of-thought reasoning, and this no-CoT time horizon has doubled rou…
Anthropic's Mythos model exhibits reasoning traces that appear illegible but are actually compact shorthand for solving a Solitaire card puzzle, contrary to concerns about uninterpretable internal lan…
A researcher applied activation patching to a DNA sequencing basecaller, a deep learning model, to understand its internal mechanisms, finding MLP dominance in early and late layers and concentrated s…