A Second Look Fellowship replication of single-forward-pass evals found that Fable 5, Opus 5, and GPT-5.6-Sol can perform 2-hop and 3-hop latent reasoning without chain-of-thought, with GPT-5.6-Sol im…
A writer on LessWrong argues that military unipolarity is likely within our lifetime and that AI development should be paused once it emerges, analyzing scenarios from the AI Futures Project's Plans S…
The AI Safety community must shift from a 'pull' model of waiting for grant applications to a 'push' model of actively recruiting senior researchers, according to a blog post from Canary Institute. Th…
At a summary judgment hearing on July 30, 2026, in Anthropic PBC v. U.S. Department of War et al., Judge Rita F. Lin of the U.S. District Court for the Northern District of California signaled skeptic…
A soldier on a six-month deployment describes using large language models like Claude as an emotional anchor and venting outlet, citing loneliness, language barriers, and career risks of confiding in …
A software engineer argues that big tech companies, including Meta, are misusing AI by creating internal token-usage leaderboards that incentivize wasteful AI interactions, and that the industry must …
Brendan Long, an independent developer, launched a chess training app at chess-pretraining.brendanlong.com that applies imitation learning principles to human chess improvement by having users choose …
An independent research group's CrucibleBench experiment found that LLM-judge-based evaluation rankings are highly sensitive to classifier components, with aggregate κ on probe detection at 0.04 and p…
A new essay by mathematician alkjash, posted on LessWrong on August 1, 2026, argues that existential risk from artificial intelligence is a pressing concern that mathematicians should engage with, pre…
A researcher proposes a reinforcement learning from verifiable rewards (RLVR) training pipeline that rewards models for red-teaming the training environment itself, rather than merely exploiting rewar…
Simon DeDeo, a researcher at Carnegie Mellon University, reports that AI models such as GPT 5.6-sol can solve open mathematical problems with at least a 1% success rate when prompted with a problem, a…
A replication by the University of Chicago xLab confirms claims from Stevinson et al. (arXiv:2510.11709) that adversarial examples in toy models exploit interference between features in superposition,…
In a retrospective essay, AI forecaster and author of the 2021 science fiction story 'Bayeswatch' reflects on how the AI landscape has changed, noting that AI now writes all of his software at 10× to …
A personal essay by Victoria Krakovna describes using AI models, including Claude Opus 5 and Fable, to analyze a decade of personal notes, worksheets, and therapy records, uncovering persistent patter…
A technical AI safety researcher argues that safety-focused researchers should engage with capabilities research, highlighting On-Policy Self-Distillation (OPSD) as a promising alternative to GRPO for…
A computational model by Philipp Kuhn, published on The Anti-Completionist Substack, simulates the interaction between biological and cultural evolution, showing that prosocial memes nearly double soc…
A new paper by Kaarel Hänni and an unnamed author, available on arXiv (2607.05735), concludes an analysis of generalization complexity in Bayesian infinite-width neural networks, resolving when functi…
Anthropic's April 2025 alignment risk update for its Claude Mythos Preview model concluded that the model 'does not possess any unknown propensities that would increase alignment risk,' but a LessWron…
A new proposal introduces the 'temporal lockbox,' a hardened observatory for AI misalignment that scores AI agents' weather forecasts against future measurements that cannot be influenced by the forec…
AI safety prizes, proposed as a pull-funding alternative to upfront grants, could incentivize breakthroughs by paying only for verified progress, with DARPA's autonomous vehicle challenges cited as a …