korrents

On the map

Tap a claim on the ring to put it at the centre.

← No single interpretability technique is reliable enough alone to…

17 connected korrents · 13 moments from 8 Feb 2022 to 3 Nov 2026.

Everything filed under AI alignment AI alignment Everything filed under LLMs LLMs Everything filed under AGI AGI Everything filed under AI agents AI agents Everything filed under reinforcement learning reinforcement learning Everything filed under metascience metascience Everything filed under design design Same subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subject Read this korrent: No single interpretability technique is reliable enough alone to justify a verdict like a model being unsafe; decorrelated techniques are needed. No single interpretability technique isreliable enough alone to justify a verdictlike a model being unsafe; decorrelatedtechniques are needed. Last stated 11 months ago 20 Dec 2025 NN Neel Nanda — holds since 2025-12-20 — tap for who they are Same subject: Auditing properties that do not shape a model's actions is probably doomed, and weight-based interpretability has not gone well. — tap to centre the map on it Auditing properties that do notshape a model's actions is probablydoomed, and weight-basedinterpretability has not gone well. Last stated 11 months ago 20 Dec 2025 NN Neel Nanda — holds since 2025-12-20 — tap for who they are Same subject: Safety evaluations should report over-refusal, since a model that always refuses would be trivially safe. — tap to centre the map on it Safety evaluations should reportover-refusal, since a model thatalways refuses would be triviallysafe. Last stated 11 months ago 27 Dec 2025 NN Neel Nanda — holds since 2025-12-27 — tap for who they are Same subject: The best interpretability deep dives combine several kinds of evidence, since no single technique is trustworthy enough on its own. — tap to centre the map on it The best interpretability deep divescombine several kinds of evidence,since no single technique istrustworthy enough on its own. Last stated 11 months ago 27 Dec 2025 NN Neel Nanda — holds since 2025-12-27 — tap for who they are Same subject: There is no route to AI alignment or safety that goes around reliability: a system that cannot reliably follow a known algorithm cannot be made safe. — tap to centre the map on it There is no route to AI alignment orsafety that goes around reliability:a system that cannot reliably followa known algorithm cannot be madesafe. Last stated a year ago 7 Jun 2025 GM Gary Marcus — holds since 2025-06-07 — tap for who they are Same subject: Judging one model worse than another from a one-shot, probabilistic output on a given day is not a good way to evaluate it. — tap to centre the map on it Judging one model worse than anotherfrom a one-shot, probabilisticoutput on a given day is not a goodway to evaluate it. Last stated 6 months ago 19 May 2026 TG Tommy Geoco — holds since 2026-05-19 — tap for who they are Same subject: A more capable model showing less reasoning is not malicious or intentional hiding; it simply needs less backtracking and self-correction. — tap to centre the map on it A more capable model showing lessreasoning is not malicious orintentional hiding; it simply needsless backtracking andself-correction. Last stated 2 months ago 3 Oct 2026 SR Sebastian Raschka — holds since 2026-10-03 — tap for who they are Same subject: Ambitious interpretability methods are often so painful to train that using them on frontier models is not worth the resources. — tap to centre the map on it Ambitious interpretability methodsare often so painful to train thatusing them on frontier models is notworth the resources. Last stated 11 months ago 20 Dec 2025 NN Neel Nanda — holds since 2025-12-20 — tap for who they are Same subject: No single way of modeling data and methods in code is universally correct; every approach is just a set of trade-offs. — tap to centre the map on it No single way of modeling data andmethods in code is universallycorrect; every approach is just aset of trade-offs. Last stated 5 years ago 8 Feb 2022 PH Phil Hagelberg — holds since 2022-02-08 — tap for who they are Same subject: A country outside the AI supply chain should just buy the index — which works only in the world where AI ends up commoditised rather than concentrated. — tap to centre the map on it A country outside the AI supplychain should just buy the index —which works only in the world whereAI ends up commoditised rather thanconcentrated. Last stated 6 months ago 4 Jun 2026 AI Alex Imas — holds since 2026-06-04 — tap for who they are Same subject: A poor country should prioritise owning a piece of AI over retraining its workers, but it should not bet everything on that. — tap to centre the map on it A poor country should prioritiseowning a piece of AI over retrainingits workers, but it should not beteverything on that. Last stated 6 months ago 4 Jun 2026 PT Phil Trammell — holds since 2026-06-04 — tap for who they are Same subject: A program that generates rich enough tests of intelligence would be as hard to build as AGI itself. — tap to centre the map on it A program that generates rich enoughtests of intelligence would be ashard to build as AGI itself. Last stated 2 years ago 6 Nov 2024 FC François Chollet — holds since 2024-11-06 — tap for who they are Same subject: A format reward should only be added when the answer is already correct. — tap to centre the map on it A format reward should only be addedwhen the answer is already correct. Last stated a month ago 10 Oct 2026 SR Sebastian Raschka — holds since 2026-10-10 — tap for who they are Same subject: A model should not be called a reasoning model unless it has been trained with reinforcement learning with verifiable rewards. — tap to centre the map on it A model should not be called areasoning model unless it has beentrained with reinforcement learningwith verifiable rewards. Last stated 2 months ago 3 Oct 2026 SR Sebastian Raschka — holds since 2026-10-03 — tap for who they are Same subject: A physical AI company needs three AIs, not one — the agent, the simulator and the critic — turning deployment into a flywheel. — tap to centre the map on it A physical AI company needs threeAIs, not one — the agent, thesimulator and the critic — turningdeployment into a flywheel. Last stated 4 months ago 3 Aug 2026 DD Dmitri Dolgov — holds since 2026-08-03 — tap for who they are Same subject: A codebase is the most important context an AI agent has, and letting many different models add to it without consistent patterns leads to trouble. — tap to centre the map on it A codebase is the most importantcontext an AI agent has, and lettingmany different models add to itwithout consistent patterns leads totrouble. Last stated 3 weeks ago 3 Nov 2026 DB Derek Briggs — holds since 2026-11-03 — tap for who they are Same subject: A collection of AI agents does not automatically form a functioning organization any more than a group of smart people does. — tap to centre the map on it A collection of AI agents does notautomatically form a functioningorganization any more than a groupof smart people does. Last stated 6 months ago 21 May 2026 RK Rohit Krishnan — holds since 2026-05-21 — tap for who they are Same subject: A company cannot be led by machines, because nobody would have recourse against them. — tap to centre the map on it A company cannot be led by machines,because nobody would have recourseagainst them. Last stated 2 months ago 15 Sept 2026 TL Tobias Lütke — holds since 2026-09-15 — tap for who they are
same subject or similar wordinga cloud: claims about one subject, named for itbar: when it was last stated, on a scale from 2015 to today — full is todaya face: someone who holds the claim — tap it for who they are

At the centre No single interpretability technique is reliable enough alone to justify a verdict like a model being unsafe; decorrelated techniques are needed. Last stated 20 Dec 2025 · 11 months ago Holds Neel Nanda Read this korrent →