korrents

On the map

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

← Mechanistic interpretability should be kept as a verification tool separate from a model's training process.

15 connected korrents · 15 moments from 24 Aug 2009 to 3 Oct 2026.

Everything filed under software quality software quality Everything filed under AI alignment AI alignment Everything filed under metascience metascience Everything filed under neural networks neural networks Everything filed under LLMs LLMs Same subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subject Read this korrent: Mechanistic interpretability should be kept as a verification tool separate from a model's training process. Mechanistic interpretability should bekept as a verification tool separate froma model's training process. Last stated 2 years ago 11 Nov 2024 DA Dario Amodei — holds since 2024-11-11 — tap for who they are Same subject: Anthropic now uses mechanistic interpretability techniques to audit new models for deception and scheming before releasing them. — tap to centre the map on it Anthropic now uses mechanisticinterpretability techniques to auditnew models for deception andscheming before releasing them. Last stated 11 months ago 1 Jan 2026 DA Dario Amodei — holds since 2026-01-01 — tap for who they are Same subject: No single interpretability technique is reliable enough alone to justify a verdict like a model being unsafe; decorrelated techniques are needed. — tap to centre the map on it No single interpretability techniqueis reliable enough alone to justifya verdict like a model being unsafe;decorrelated techniques 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: Complete reverse-engineering of neural networks is a less promising research direction than studying model biology and applying interpretability usefully — tap to centre the map on it Complete reverse-engineering ofneural networks is a less promisingresearch direction than studyingmodel biology and applyinginterpretability usefully Last stated a year ago 19 Aug 2025 NN Neel Nanda — holds since 2025-08-19 — tap for who they are Same subject: Interpretability agents are useful for evaluating techniques because they are reproducible in a way human researchers are not. — tap to centre the map on it Interpretability agents are usefulfor evaluating techniques becausethey are reproducible in a way humanresearchers are not. Last stated 11 months ago 20 Dec 2025 NN Neel Nanda — holds since 2025-12-20 — tap for who they are Same subject: Optimizing a model's reasoning explanations during training might teach it to cheat more efficiently by hiding its intermediate chain. — tap to centre the map on it Optimizing a model's reasoningexplanations during training mightteach it to cheat more efficientlyby hiding its intermediate chain. Last stated 2 months ago 3 Oct 2026 SR Sebastian Raschka — holds since 2026-10-03 — 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: 100% test coverage is the only correct test coverage. — tap to centre the map on it 100% test coverage is the onlycorrect test coverage. Last stated 2 years ago 24 Feb 2025 HS Hynek Schlawack — holds since 2025-02-24 — tap for who they are Same subject: 95 percent of the software made in an age of abundance will be shitty, and shitty software is easy to deal with because it goes away. — tap to centre the map on it 95 percent of the software made inan age of abundance will be shitty,and shitty software is easy to dealwith because it goes away. Last stated 2 months ago 23 Sept 2026 GT Geoff Teehan — holds since 2026-09-23 — tap for who they are Same subject: A bug live for one day is better than an app stuck in review for two weeks. — tap to centre the map on it A bug live for one day is betterthan an app stuck in review for twoweeks. Last stated 17 years ago 24 Aug 2009 JH Joe Hewitt — holds since 2009-08-24 — tap for who they are Same subject: AI coding got so good so fast because code is verifiable: there is a right and wrong to hill-climb against, unlike subjective tasks. — tap to centre the map on it AI coding got so good so fastbecause code is verifiable: there isa right and wrong to hill-climbagainst, unlike subjective tasks. Last stated 8 months ago 18 Mar 2026 JM John Maeda — holds since 2026-03-18 — tap for who they are Same subject: AI-assisted coding will generate a lot of low-quality output that developers will need to clean up, even as it also produces good results. — tap to centre the map on it AI-assisted coding will generate alot of low-quality output thatdevelopers will need to clean up,even as it also produces goodresults. Last stated 9 months ago 9 Feb 2026 SH Scott Hanselman — holds since 2026-02-09 — tap for who they are Same subject: AI-generated code can look correct yet be slightly wrong, which is why code review is so important. — tap to centre the map on it AI-generated code can look correctyet be slightly wrong, which is whycode review is so important. Last stated 7 months ago 15 Apr 2026 LR Lee Robinson — holds since 2026-04-15 — tap for who they are Same subject: A coding agent left to run over a weekend returns work that is half good, a quarter garbage and a quarter in need of repair. — tap to centre the map on it A coding agent left to run over aweekend returns work that is halfgood, a quarter garbage and aquarter in need of repair. Last stated a year ago 18 Aug 2025 CB Clay Bavor — holds since 2025-08-18 — tap for who they are Same subject: A developer using AI coding tools remains responsible for reviewing the code they ship. — tap to centre the map on it A developer using AI coding toolsremains responsible for reviewingthe code they ship. Last stated 8 months ago 10 Mar 2026 JS Jon Seager — holds since 2026-03-10 — tap for who they are Same subject: A team that slows down and reads every pull request and every line of code should expect only a 30 to 50 percent productivity lift from AI. — tap to centre the map on it A team that slows down and readsevery pull request and every line ofcode should expect only a 30 to 50percent productivity lift from AI. Last stated 4 months ago 15 Jul 2026 DH Dex Horthy — holds since 2026-07-15 — 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 2009 to today (stretched back to the oldest claim here) — full is todaya face: someone who holds the claim — tap it for who they are

At the centre Mechanistic interpretability should be kept as a verification tool separate from a model's training process. Last stated 11 Nov 2024 · 2 years ago Holds Dario Amodei Read this korrent →