bayes

Quoting the account’s own earlier “most bottlenecks are fake” post, it names progress on alignment, safety and integration as a real bottleneck: capabilities are outrunning our ability to bound what happens at every scale, from model to civilization. A diagnosis, not a probability of catastrophe. Full post text inspected via the X API.

Argues that building computers that solve problems without creating a new species is possible in principle. Today’s systems are more lifelike, and on the current course we will probably get something lifelike, but calling that inevitable is a failure of imagination and memetic contagion. Explicitly not a judgment of whether a successor species is good or likely. Full post text inspected via the X API.

Endorses open, bilateral research on verification technology to build trust for crises, while doubting its effects and noting it could be harmful if widely deployed. The account doubts full international cooperation because the incentive to race is profound under most risk models, and doubts everyone opts out without a coordination-forcing disaster, which might even fuel racing; marginal gains are still worth pursuing. The 0.1%/0.05% figures are a hypothetical about leaders’ choices, not the account’s estimate. Quoted post belongs to someone else; full text inspected via the X API.

Says crazy dangerous general models are not needed for most of the benefits people want now: industrial takeoff and a thousand years of medical progress are probably possible without “RL’ing a demon into existence”. A claim about possibility, not a forecast that labs will choose that path. Full post text inspected via the X API.

Quoting someone else’s question whether to fear the technology or a small number of people controlling it, the account answers that it fears both. Short post; it does not rank the two risks or give probabilities. Full post text inspected via the X API.

Quotes its own post asking why labs, if alignment, safety, control and monitoring matter (it says they do), are not open-sourcing that work without leaking much IP. Rejects the reply that alignment is simply what the market wants: if OpenAI invented the seatbelt and Anthropic the airbag, both should be public knowledge and deployed everywhere. A demand for openness about safety methods, not about all model weights. Full post text inspected via the X API.

Argues that doom stories imagine AI fitness rising to infinity while human fitness stays fixed, but culture and engineering raise humanity’s effective fitness. Greater intelligence does not computationally necessitate greater fitness. Says strong instrumental convergence is a real issue at scale but ultimately an engineering skill issue, and expects far better ways to program these systems. Confidence that the problem is solvable, not a claim that current systems are safe. Full post text inspected via the X API.

Asks whether alignment by default was always wrong, with models previously too weak to reveal the gap, or held in the pretraining era until RL circa 2026 began warping previously good minds. The account leans toward the latter and calls the question important. An inclination, not an established finding. Full post text inspected via the X API.

Argues that even with solved interpretability and alignment, checking whether a given output is good or bad may be hard, as it is for the most aligned humans. A verified value-aligned model therefore does not imply a value-aligned outcome. A conceptual limit, not a claim that alignment work is futile. Full post text inspected via the X API.