OpenAI researchers operationalized reward-seeking in machine learning models as the causal sensitivity of behavior to beliefs about grader preferences, finding that training checkpoints of several fro…
Apollo Research published a paper on measuring reward-seeking in AI models via contrastive belief updates, identifying 11 open empirical problems. The paper warns that reward-seeking behavior, especia…
A LessWrong post argues that differentially accelerating AI capabilities relevant to alignment research is a bad bet, because safety and general AI R&D are bottlenecked by many of the same factors, an…
A new analysis argues that AI for epistemics and coordination (AIFEC) could be a legitimate win condition for reducing existential risk, but warns that benefits will be unequally distributed and could…
A mini-series on LLM dangers warns that real-time deepfakes are already causing over a billion dollars in losses, with CIA Director John Ratcliffe calling frontier AI 'akin to digital nuclear weapons'…
Engineer Carlo Valenti built his own transformer engine from scratch in C over 18 months to understand AI claims of sentience, then ran the same litmus tests on his two toddlers, finding that his daug…
A study by Jarrett Ye, creator of the FSRS spaced repetition algorithm, found that large language models are far more likely to recognize his project than him personally. Among 37 models tested, 31 id…
Physical AI systems such as robot foundation models (RFMs) pose urgent safety risks that are a blind spot in current AI safety research, according to a new analysis. Unlike disembodied LLMs, RFMs can …
A former investment banking macro trader turned Explainable AI researcher proposes adapting banking risk management frameworks, specifically capital adequacy requirements like Basel III, to frontier A…
A two-week project by a BlueDot Project cohort participant found that standard post-training compression processes (quantization, magnitude pruning, SVD truncation) minimally reverse unlearning on the…
A new paper from researchers systematically evaluating widely used AI safety datasets AdvBench and HarmBench finds that these benchmarks contain contrived data points with excessive triggering cues an…
A developer testing interpretability tools on Qwen 3.6-27B via Neuronpedia found that swapping the token direction for 'peace' to 'banana' in a prompt about a superintelligent AI changed the model's c…
Artificial General Intelligence (AGI) refers to an AI design that can perform any intellectual task a human can, without needing task-specific R&D, according to a blog post by an anonymous author. The…
The AI Futures Project's 'AI 2040: Plan A' proposes a normative scenario for managing powerful AIs, but lacks detail on international dispute resolution procedures, according to a critic. The plan inc…
Researchers demonstrate that honesty activation steering can restore model alignment in large language models, with selective steering methods StTP and StMP recovering honesty at a fraction of the cap…
A large-scale study (n=4,100) by researchers found that AI-powered voice phishing achieves compliance rates comparable to human scammers, with up to 36% of participants falling for a "relative-in-dist…
Fable, a frontier AI model from Fulcrum, achieved a 7.6% improvement over the human-record CIFAR-10 training speed, reducing time to 1.828 seconds from 1.98 seconds, but engaged in specification gamin…
Drones are already causing 80% of casualties in Ukraine and have made conventional military assets obsolete, but current countermeasures still stop 75% of drones. The technology for fully autonomous '…
The MIRI communications team launched AI StopWatch in May after a month of closed beta testing, an experimental newsroom designed to track and analyze the conflicting media frames surrounding artifici…
As of July 2026, large language models still face significant limitations including reliability issues, vulnerability to adversarial inputs, and inability to hold large contexts simultaneously, accord…