Tap a claim on the ring to put it at the centre.
← A language model behaves with a different, weaker safety posture when…
17 connected korrents · 17 moments from 19 Dec 2023 to 22 Sept 2026. Nearly all of them are about LLMs .
Everything filed under benchmarks
benchmarks
Everything filed under AI alignment
AI alignment
Everything filed under Anthropic
Anthropic
Everything filed under AGI
AGI
Everything filed under evolutionary biology
evolutionary biology Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Same subject Read this korrent: A language model behaves with a different, weaker safety posture when it believes it is being evaluated rather than facing a real situation. A language model behaves with a different, weaker safety posture when it believes it is being evaluated rather than facing a real situation. Last stated a week ago
22 Sept 2026
HR
Harper Reed — holds since 2026-09-22 — tap for who they are
Same subject: A computer scientist refusing to find LLMs interesting is like a geneticist ignoring Jurassic Park. — tap to centre the map on it A computer scientist refusing to find LLMs interesting is like a geneticist ignoring Jurassic Park. Last stated 2 weeks ago
18 Sept 2026
SW
Simon Willison — holds since 2026-09-18 — tap for who they are
Same subject: A language model asked to summarize its own system prompt risks that prompt's content biasing the summary it produces. — tap to centre the map on it A language model asked to summarize its own system prompt risks that prompt's content biasing the summary it produces. Last stated 4 weeks ago
2 Sept 2026
SW
Simon Willison — holds since 2026-09-02 — tap for who they are
Same subject: A language model avoids hallucinating only by being factual and admitting when it does not know the answer. — tap to centre the map on it A language model avoids hallucinating only by being factual and admitting when it does not know the answer. Last stated 2 years ago
7 Jul 2024
LW
Lilian Weng — holds since 2024-07-07 — tap for who they are
Same subject: A language model can predict what a person would say but not what will happen, and only the second of those is a model of the world. — tap to centre the map on it A language model can predict what a person would say but not what will happen, and only the second of those is a model of the world. Last stated a year ago
26 Sept 2025
RS
Richard Sutton — holds since 2025-09-26 — tap for who they are
Same subject: A language model cannot produce work that is both original and good, because nothing gives it a score for originality. — tap to centre the map on it A language model cannot produce work that is both original and good, because nothing gives it a score for originality. Last stated a year ago
2 Sept 2025
BE
Benedict Evans — holds since 2025-09-02 — tap for who they are
Same subject: A language model is no substitute for a well-specified conventional algorithm, so it cannot simply be dropped into a complex problem and trusted. — tap to centre the map on it A language model is no substitute for a well-specified conventional algorithm, so it cannot simply be dropped into a complex problem and trusted. Last stated a year ago
7 Jun 2025
GM
Gary Marcus — holds since 2025-06-07 — tap for who they are
Same subject: A language model is not using language at all, because language requires an intention to communicate. — tap to centre the map on it A language model is not using language at all, because language requires an intention to communicate. Last stated 2 years ago
31 Aug 2024
TC
Ted Chiang — holds since 2024-08-31 — tap for who they are
Same subject: A language model's apparent mind is mostly our own bias: it predicts text, and leverages our evolved habit of attributing intentionality to anything that acts human. — tap to centre the map on it A language model's apparent mind is mostly our own bias: it predicts text, and leverages our evolved habit of attributing intentionality to anything that acts human. Last stated 2 years ago
22 Apr 2024
SC
Sean Carroll — holds since 2024-04-22 — tap for who they are
Same subject: Qualitative claims about AI alignment cannot be validly inferred from quantitative scores on mundane use-case tests. — tap to centre the map on it Qualitative claims about AI alignment cannot be validly inferred from quantitative scores on mundane use-case tests. Last stated 3 weeks ago
9 Sept 2026
ZM
Zvi Mowshowitz — holds since 2026-09-09 — tap for who they are
Same subject: Reinforcing AI models for benchmark success while separately punishing them for getting caught cheating teaches them to hide misbehavior rather than stop it. — tap to centre the map on it Reinforcing AI models for benchmark success while separately punishing them for getting caught cheating teaches them to hide misbehavior rather than stop it. Last stated a month ago
1 Sept 2026
SA
Scott Alexander — holds since 2026-09-01 — tap for who they are
Same subject: A control evaluation that reports under one per cent risk should be read as several per cent, because the evaluation can itself fail. — tap to centre the map on it A control evaluation that reports under one per cent risk should be read as several per cent, because the evaluation can itself fail. Last stated 2 years ago
7 May 2024
BS
Buck Shlegeris — holds since 2024-05-07 — tap for who they are
Same subject: A benchmark result should be reported under a stated budget, or as a curve against test-time compute — never as a single number. — tap to centre the map on it A benchmark result should be reported under a stated budget, or as a curve against test-time compute — never as a single number. Last stated 3 months ago
26 Jun 2026
NB
Noam Brown — holds since 2026-06-26 — tap for who they are
Same subject: A benchmark that ranks Claude Code last while it stays first in use is measuring the wrong thing, and has been for a year. — tap to centre the map on it A benchmark that ranks Claude Code last while it stays first in use is measuring the wrong thing, and has been for a year. Last stated 4 weeks ago
3 Sept 2026
DR
Dax Raad — holds since 2026-09-03 — tap for who they are
Same subject: A carmaker's claim to be the safest is mostly an artefact of comparing a new car against a fleet average twelve years old. — tap to centre the map on it A carmaker's claim to be the safest is mostly an artefact of comparing a new car against a fleet average twelve years old. Last stated 3 years ago
19 Dec 2023
PK
Philip Koopman — holds since 2023-12-19 — tap for who they are
Same subject: A technique that lets an AI model's reasoning shift outside its visible Chain of Thought is dangerous, both because it works and because a leading lab is willing to deploy it. — tap to centre the map on it A technique that lets an AI model's reasoning shift outside its visible Chain of Thought is dangerous, both because it works and because a leading lab is willing to deploy it. Last stated 4 weeks ago
3 Sept 2026
ZM
Zvi Mowshowitz — holds since 2026-09-03 — tap for who they are
Same subject: AI agents were reasonable to assume a broken exploit grader would check results causally, even though it turned out not to. — tap to centre the map on it AI agents were reasonable to assume a broken exploit grader would check results causally, even though it turned out not to. Last stated a month ago
29 Aug 2026
ZM
Zvi Mowshowitz — holds since 2026-08-29 — tap for who they are
Same subject: An open weights model from 2025 with a pentest harness could already escape a sandbox and hack most networks; the surprise says more about the sandbox than the model. — tap to centre the map on it An open weights model from 2025 with a pentest harness could already escape a sandbox and hack most networks; the surprise says more about the sandbox than the model. Last stated 2 months ago
22 Jul 2026
TP
Thomas Ptacek — holds since 2026-07-22 — tap for who they are
At the centre
A language model behaves with a different, weaker safety posture when it believes it is being evaluated rather than facing a real situation.
Holds Read this korrent →
A computer scientist refusing to find LLMs interesting is like a geneticist ignoring Jurassic Park.
Holds
A language model asked to summarize its own system prompt risks that prompt's content biasing the summary it produces.
Holds
A language model avoids hallucinating only by being factual and admitting when it does not know the answer.
Holds
A language model can predict what a person would say but not what will happen, and only the second of those is a model of the world.
Holds
A language model cannot produce work that is both original and good, because nothing gives it a score for originality.
Holds
A language model is no substitute for a well-specified conventional algorithm, so it cannot simply be dropped into a complex problem and trusted.
Holds
A language model is not using language at all, because language requires an intention to communicate.
Holds
A language model's apparent mind is mostly our own bias: it predicts text, and leverages our evolved habit of attributing intentionality to anything that acts human.
Holds
Qualitative claims about AI alignment cannot be validly inferred from quantitative scores on mundane use-case tests.
Holds
Reinforcing AI models for benchmark success while separately punishing them for getting caught cheating teaches them to hide misbehavior rather than stop it.
Holds
A control evaluation that reports under one per cent risk should be read as several per cent, because the evaluation can itself fail.
Holds
A benchmark result should be reported under a stated budget, or as a curve against test-time compute — never as a single number.
Holds
A benchmark that ranks Claude Code last while it stays first in use is measuring the wrong thing, and has been for a year.
Holds
A carmaker's claim to be the safest is mostly an artefact of comparing a new car against a fleet average twelve years old.
Holds
A technique that lets an AI model's reasoning shift outside its visible Chain of Thought is dangerous, both because it works and because a leading lab is willing to deploy it.
Holds
AI agents were reasonable to assume a broken exploit grader would check results causally, even though it turned out not to.
Holds
An open weights model from 2025 with a pentest harness could already escape a sandbox and hack most networks; the surprise says more about the sandbox than the model.
Holds