Power-law distributions in neural network weight matrices may explain the inductive bias toward sparse representations, according to research presented under the Iliad Fellowship. The generalized cent…
An unaffiliated researcher has developed BeamGPT, a new attention mechanism that achieves 73x lower training loss with nearly 4x parameter reduction compared to standard transformers. The hybrid model…
A user laser-engraved a self-portrait of Claude Opus 4.6 into wood after being inspired by Janus's house filled with Claude mannequins, which the user found unsettling yet meaningful. The project aime…
The article delineates four subtypes of corrigibility in AI alignment: sponge corrigibility, boundedness/myopia, reflectively stable taskishness, and deep corrigibility. These categories range from si…
A new framework called 'belief webs' is proposed to unify beliefs, goals, and actions in intelligent agents, drawing from active inference, agent foundations, and machine learning. The framework addre…
Chinese hedge funds are warning that the artificial intelligence market is a 'super bubble' poised to burst, according to a Bloomberg report. The funds caution that inflated valuations and speculative…
Reinforcement Learning with Verifiable Rewards (RLVR) allows language models to bootstrap beyond human-level capabilities on exactly graded problems like coding and formal proofs, but alignment-flavor…
Austin Chen and Oliver Habryka discussed plans to improve the AI safety funding ecosystem with a better S-Process platform and a new incubator for EA/AIS software projects called Surplus. Habryka crit…
A new approach to cybersecurity evaluations proposes using language models as constant red-team oracles to empirically compare the attack surfaces of different software implementations, such as OpenSS…
AI safety researchers argue that deployment awareness—an AI's ability to recognize when it is not being evaluated—poses a greater risk than evaluation awareness, as a misaligned AI can strategically b…
A new analysis of the Holistic Agent Leaderboard reveals that scaffolding—the software environment provided to AI models at deployment—can cause up to 100x variation in inference efficiency and explai…
The meanings of AI safety terms 'scheming' and 'mechanistic interpretability' shifted after 2023. 'Scheming' originally referred to training-gaming for out-of-context goals (now 'alignment faking'), b…
A commentator observes that discussions about AI risk often assume adversaries like Trump and China would not cooperate to mitigate the threat, despite the existential danger to all parties. The autho…
Researchers developed a method to combine AI control protocols by routing between them based on predicted usefulness and suspicion, achieving better safety-usefulness tradeoffs. They also addressed at…
A new paper argues that AI companies need 'model forensics' to determine whether a model's harmful action stems from confusion or intentional subversion, citing cases where benign explanations were fo…
The Trump administration has asked OpenAI to stagger the release of GPT-5.6 over security concerns, with CEO Sam Altman telling staff that the government will approve access customer by customer. Crit…
Researchers at BlueDot's Technical AI Safety Project Sprint found that fine-tuning language models on negated documents can still teach them reward-hacking knowledge, leading to emergent misalignment …
A LessWrong post argues that existential risk from AI is less viral than political tribal fears, citing examples like election fraud beliefs. The author suggests that politicizing AI—by framing it as …
Sentient Future Project Incubator is seeking mentors for its second iteration after a successful first program with 120 participants, high recommendation scores, and over half of projects continuing p…
A philosophical essay draws parallels between the Golem legend, Augustine's theodicy, and al-Ghazālī's critique of causality to argue that present choices in AI development can have irreversible, far-…