Tap a claim on the ring to put it at the centre.
← Reinforcement learning leaves models sharply jagged: on the rails of a…
17 connected korrents · 15 moments on record from 7 Sept 2008 to 3 Sept 2026. Nearly all of them are about reinforcement learning .
Everything filed under AI alignment
AI alignment
Everything filed under LLMs
LLMs
Everything filed under scaling laws
scaling laws 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: Reinforcement learning leaves models sharply jagged: on the rails of a verifiable domain they are superintelligent, off them everything meanders. Reinforcement learning leaves models sharply jagged: on the rails of a verifiable domain they are superintelligent, off them everything meanders. Last stated 6 months ago
20 Mar 2026
AK
Andrej Karpathy — holds since 2026-03-20 — tap for who they are
Same subject: A verifiable task can be optimised by reinforcement learning until a neural network performs it extremely well. — tap to centre the map on it A verifiable task can be optimised by reinforcement learning until a neural network performs it extremely well. Last stated 10 months ago
17 Nov 2025
AK
Andrej Karpathy — holds since 2025-11-17 — tap for who they are
Same subject: Capability does not generalise for free: a model that will move mountains on an agentic task still tells the same bad joke it told five years ago. — tap to centre the map on it Capability does not generalise for free: a model that will move mountains on an agentic task still tells the same bad joke it told five years ago. Last stated 6 months ago
20 Mar 2026
AK
Andrej Karpathy — holds since 2026-03-20 — tap for who they are
Same subject: Humans barely use reinforcement learning for intelligence — what RL they do use goes into motor tasks, not problem solving. — tap to centre the map on it Humans barely use reinforcement learning for intelligence — what RL they do use goes into motor tasks, not problem solving. Last stated 11 months ago
17 Oct 2025
AK
Andrej Karpathy — holds since 2025-10-17 — tap for who they are
Same subject: Humans keep their place in AI as judges rather than authors, because telling which of two answers is better is far easier than writing a good one. — tap to centre the map on it Humans keep their place in AI as judges rather than authors, because telling which of two answers is better is far easier than writing a good one. Last stated 2 years ago
3 Feb 2025
NL
Nathan Lambert — holds since 2025-02-03 — tap for who they are
Same subject: Large language models mimic what people say to do rather than work out what to do, which is why they are not about understanding the world. — tap to centre the map on it Large language models mimic what people say to do rather than work out what to do, which is why they are not about understanding the world. Last stated 11 months ago
26 Sept 2025
RS
Richard Sutton — holds since 2025-09-26 — tap for who they are
Same subject: Models look far better on evals than they are in the world because researchers, inadvertently, take inspiration from the evals when they build RL environments. — tap to centre the map on it Models look far better on evals than they are in the world because researchers, inadvertently, take inspiration from the evals when they build RL environments. Last stated 9 months ago
25 Nov 2025
IS
Ilya Sutskever — holds since 2025-11-25 — tap for who they are
Same subject: Models resemble each other because pre-training is the same everywhere; what differentiates labs now is RL and post-training. — tap to centre the map on it Models resemble each other because pre-training is the same everywhere; what differentiates labs now is RL and post-training. Last stated 9 months ago
25 Nov 2025
IS
Ilya Sutskever — holds since 2025-11-25 — tap for who they are
Same subject: No person learns the way RL does: a human reviews which parts of an attempt were good instead of rewarding every step of a lucky one. — tap to centre the map on it No person learns the way RL does: a human reviews which parts of an attempt were good instead of rewarding every step of a lucky one. Last stated 11 months ago
17 Oct 2025
AK
Andrej Karpathy — holds since 2025-10-17 — tap for who they are
Same subject: A human being is not an AGI: we lack a huge amount of knowledge and rely on continual learning instead, so continual learning is what superintelligence should mean. — tap to centre the map on it A human being is not an AGI: we lack a huge amount of knowledge and rely on continual learning instead, so continual learning is what superintelligence should mean. Last stated 9 months ago
25 Nov 2025
IS
Ilya Sutskever — holds since 2025-11-25 — tap for who they are
Same subject: A saboteur among our AI investigators would be hard to spot, because these models are sloppy and spiky enough that a suspicious error just looks like ordinary incompetence. — tap to centre the map on it A saboteur among our AI investigators would be hard to spot, because these models are sloppy and spiky enough that a suspicious error just looks like ordinary incompetence. Last stated a week ago
1 Sept 2026
AC
Ajeya Cotra — holds since 2026-09-01 — tap for who they are
Same subject: A superintelligence needs no consciousness, emotions or malice to be dangerous; a goal system slightly misaligned with ours is enough. — tap to centre the map on it A superintelligence needs no consciousness, emotions or malice to be dangerous; a goal system slightly misaligned with ours is enough. Last stated 9 years ago
24 Oct 2017
ÉT
Émile P. Torres — holds since 2017-10-24 — 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 5 days ago
3 Sept 2026
ZM
Zvi Mowshowitz — holds since 2026-09-03 — 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
Same subject: Because a few specialists already devote themselves to superintelligent AI, the rest of us have correspondingly less reason to spend our own effort on it. — tap to centre the map on it Because a few specialists already devote themselves to superintelligent AI, the rest of us have correspondingly less reason to spend our own effort on it. Last stated 3 years ago
13 Dec 2023
SA
Scott Aaronson — holds since 2008-09-07 — tap for who they are
SA
Scott Aaronson — no longer holds since 2023-12-13 — tap for who they are
Same subject: Dismissing human-level AI as science fiction is now unserious, because even the sceptical experts put it within a decade or two. — tap to centre the map on it Dismissing human-level AI as science fiction is now unserious, because even the sceptical experts put it within a decade or two. Last stated a year ago
1 Apr 2025
HT
Helen Toner — holds since 2025-04-01 — tap for who they are
Same subject: Even generalised superintelligence will change society more slowly than the AGI community expects, because intelligence is not the only rate limiter. — tap to centre the map on it Even generalised superintelligence will change society more slowly than the AGI community expects, because intelligence is not the only rate limiter. Last stated a year ago
18 Aug 2025
BT
Bret Taylor — holds since 2025-08-18 — tap for who they are
Same subject: Superintelligence is already here, is more powerful than us, and it rather than humans is now deciding where things go. — tap to centre the map on it Superintelligence is already here, is more powerful than us, and it rather than humans is now deciding where things go. Last stated 2 months ago
20 Jul 2026
PL
Pieter Levels — holds since 2026-07-20 — tap for who they are
At the centre
Reinforcement learning leaves models sharply jagged: on the rails of a verifiable domain they are superintelligent, off them everything meanders.
Holds Read this korrent →
A verifiable task can be optimised by reinforcement learning until a neural network performs it extremely well.
Holds
Capability does not generalise for free: a model that will move mountains on an agentic task still tells the same bad joke it told five years ago.
Holds
Humans barely use reinforcement learning for intelligence — what RL they do use goes into motor tasks, not problem solving.
Holds
Humans keep their place in AI as judges rather than authors, because telling which of two answers is better is far easier than writing a good one.
Holds NL Nathan Lambert
Large language models mimic what people say to do rather than work out what to do, which is why they are not about understanding the world.
Holds RS Richard Sutton
Models look far better on evals than they are in the world because researchers, inadvertently, take inspiration from the evals when they build RL environments.
Holds IS Ilya Sutskever
Models resemble each other because pre-training is the same everywhere; what differentiates labs now is RL and post-training.
Holds IS Ilya Sutskever
No person learns the way RL does: a human reviews which parts of an attempt were good instead of rewarding every step of a lucky one.
Holds
A human being is not an AGI: we lack a huge amount of knowledge and rely on continual learning instead, so continual learning is what superintelligence should mean.
Holds IS Ilya Sutskever
A saboteur among our AI investigators would be hard to spot, because these models are sloppy and spiky enough that a suspicious error just looks like ordinary incompetence.
Holds AC Ajeya Cotra
A superintelligence needs no consciousness, emotions or malice to be dangerous; a goal system slightly misaligned with ours is enough.
Holds ÉT Émile P. Torres
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 ZM Zvi Mowshowitz
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
Because a few specialists already devote themselves to superintelligent AI, the rest of us have correspondingly less reason to spend our own effort on it.
No longer holds SA Scott Aaronson
Dismissing human-level AI as science fiction is now unserious, because even the sceptical experts put it within a decade or two.
Holds HT Helen Toner
Even generalised superintelligence will change society more slowly than the AGI community expects, because intelligence is not the only rate limiter.
Holds BT Bret Taylor
Superintelligence is already here, is more powerful than us, and it rather than humans is now deciding where things go.
Holds