Archive

A formal argument that alignment methods assuming static preferences are unsound once an AI can change what people want, with a framework for reasoning about it. What the claims are, and what they rest on.

Four axes for saying whether an AI system is manipulating someone, and the argument that the absence of a definition is itself the risk. What the claims are, and what they rest on.

A pre-registered audit of Twitter's engagement-based timeline against a reverse-chronological baseline — measured on real users, not modelled. What the claims are, and what they rest on.

A six-premise argument, each premise given a probability, for how likely it is that power-seeking AI causes an existential catastrophe by 2070. What the claims are, and what they rest on.

A taxonomy of societal-scale AI risk sorted by how much of the harm was anyone's intention, with a story for each type. What the claims are, and what they rest on.

A position paper arguing that the way we train large models pushes them toward situational awareness, deceptive alignment and power-seeking. What the claims are, and what they rest on.

A lab's own account of training a 671B mixture-of-experts model in 2.788M H800 GPU hours, with FP8 and no auxiliary balancing loss. What the claims are, and what they rest on.

A lab's own case studies of frontier models sabotaging code, assisting fraud, mislabeling transcripts, and coaching disclosure. What the claims are, and what they rest on.

Prices whether AI control mitigations — monitoring, honeypots, deployment-time defences — will be strong enough by 2030 to stop a misaligned model that is actively trying to cause catastrophe.

Prices whether it becomes possible and affordable by 2070 to build AI systems that are highly capable, plan on their own, and model their own situation — the step every other state here rests on.

Prices whether an actual AI-caused catastrophe occurs given that models already have the capability to cause one, with control mitigations as the remaining line of defence.

Prices whether frontier AI models cross capability thresholds sufficient to make a catastrophe physically possible by 2030, on the strict reading its downstream edge requires.

Prices whether permanent human disempowerment by power-seeking misaligned AI would itself destroy humanity's long-term potential, given every upstream condition holds by 2070.

Prices whether misaligned AI power-seeking, granted it is already happening at high impact, scales to the permanent disempowerment of essentially all humanity by 2070.

Prices whether strong incentives to build powerful agentic AI will exist by 2070, given that such systems are technically and economically feasible.

Prices the chance that humanity loses meaningful control over AI systems and the trajectory they set, unconditionally, by 2035.

Prices whether deployed misaligned AI systems actually seek power over people in high-impact ways, given that such systems are feasible, misaligned by default, and built anyway.

Prices whether, given feasible powerful agentic AI and strong incentives to build it, aligned systems turn out much harder to build than misaligned ones that still look worth deploying.

Prices whether a frontier model, by 2030, possesses the capability to sabotage oversight — disabling monitors, sandbagging evaluations, subverting supervision — with no parent conditions attached.

Prices whether AI models will deliberately mislead their overseers by 2035, and why the estimate turns on how strictly that bar is read.

A lab's argument that LLM API nondeterminism is batch variance, not floating-point concurrency. What the claims are, and what they rest on.

A data vendor's account of NVIDIA using its tasks to lift Nemotron 3 Ultra on GDPval. What the claims are, and what they rest on.

A lab's own account of taking gpt-oss-20b from 3.1% to 17.0% on Terminal-Bench 2.0. What the claims are, and what they rest on.

A five-hour conversation about what AI does to the craft of programming. What DHH claims, and what it rests on.

An hour-long conversation about AI, trust, and how much time is left to decide who stays in control.

A twenty-minute argument that ontologies are the guardrails probabilistic agents need. What the talk claims, and what it rests on.

A conversation about inference getting radically cheaper. What the claim is, and what it rests on.