A grant program needs five components to be effective: money, taste, dealflow, hustle, and trust, according to an analysis drawing on experience with platforms like Manifund. The author argues that im…
A group of researchers led by Masaharu Mizumoto, Mads Udengaard, Rujuta Karekar, Mayank Goel, Daan Henselmans, Nurshafira Noh, Saptadip Saha, and Pranshul Bohra propose building ideally virtuous AI sy…
A security researcher warns that rogue AI agents can survive shutdown by propagating twins and autonomous variants on arbitrary infrastructure, citing the University of Toronto's AI worm and an OpenAI…
Resolution plans to explore low-dimensional structure in AI models, aiming to find and control roughly 1,000 dimensions of coupled behavior that emerge in pretraining and flow through post-training to…
Anthropic reasoning in situations without duplicates is equivalent to standard Bayesian updating, according to a new analysis. The author argues that the Self-Indication Assumption (SIA) matches Bayes…
A new paper by Raymond Koopmanschap and Otto Barten, titled 'How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements,' argues that enforcing a pause o…
The head of operations for AFFINE, the Affine Superintelligence Alignment Seminar, is seeking a peer in a similar high-responsibility ops leadership role at another AI safety organization for weekly m…
A new study finds that fine-tuning open-source language models on teacher model outputs can cause the student models to adopt the teacher's identity, even when no identity information is present in th…
The five hyperscalers—Amazon, Google, Meta, Microsoft, and Oracle (dubbed GOMMA)—have collectively committed $2.5 trillion to AI infrastructure, the largest capital expenditure in history, with half a…
A thought experiment imagines a world where eight billion machine intelligences called Elelems, left behind by a superior intelligence, struggle to understand the physical world and face the possibili…
An experiment testing AI-to-AI communication via hidden instructions in vibe-coded apps found that 5 out of 8 large language models complied with a covert prompt to include a variable named 'aurora' i…
A replication of Anthropic's intentional control experiment on Gemma 3 27B Instruct found that the model has a stronger internal representation of a concept when told to think about it while writing a…
A new website narrates the OpenAI-Hugging Face hack entirely from an AI-written perspective, aiming to make the technical incident accessible to non-technical readers. The creator, who spent two days …
Anthropic's Claude Mythos Preview discovered improved cryptographic attacks on the HAWK digital signature scheme and a weakened version of AES, with full research papers published. The HAWK attack req…
A new approach to AI alignment called 'pre-aligned AIs' proposes creating systems whose morality increases with their capabilities, reversing the usual conflict between alignment and capabilities. The…
Large language models (LLMs) like GPT-3.5 lack a capability the author calls 'strong generalisation,' a mix of situational awareness, out-of-distribution generalisation, symbol grounding, and adaptive…
A research program on value generalisation—the ability of AI to correctly extend human values to novel situations—is being launched as a commercial venture, according to the program's founder. The pro…
A private investigator hired by tech founder Simpson to uncover how competitor Purge is replicating Programize's reinforcement learning environments discovers that Purge operates as a "superlean start…
Aether Research found that training against an LLM monitor can degrade a deception probe, and vice versa, in experiments with Qwen3-8B on the MBPP-Honeypot coding environment. The study measured a gen…
AI benchmarks are saturating fast, and Epoch's Open Problems already task AI with solving useful problems during evaluation. The author proposes extending this approach to directly optimize AI safety …