korrents

On the map

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

← The KL term in GRPO is not essential; whether to use it is a…

17 connected korrents · 13 moments from 1 Oct 2023 to 10 Oct 2026.

Everything filed under reinforcement learning reinforcement learning Everything filed under AI alignment AI alignment Everything filed under robotics robotics Everything filed under LLMs LLMs Everything filed under software performance software performance Same subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subject Read this korrent: The KL term in GRPO is not essential; whether to use it is a hyperparameter to decide case by case, and on math models may do better without it. The KL term in GRPO is not essential;whether to use it is a hyperparameter todecide case by case, and on math modelsmay do better without it. Last stated 2 months ago 3 Oct 2026 SR Sebastian Raschka — holds since 2026-10-03 — tap for who they are Same subject: A KL loss term is very important in domains with a lot of reward hacking but usually not necessary for math models. — tap to centre the map on it A KL loss term is very important indomains with a lot of reward hackingbut usually not necessary for mathmodels. Last stated a month ago 10 Oct 2026 SR Sebastian Raschka — holds since 2026-10-10 — tap for who they are Same subject: When a KL loss term is not needed, it is just overhead, since it requires keeping an extra reference model in memory. — tap to centre the map on it When a KL loss term is not needed,it is just overhead, since itrequires keeping an extra referencemodel in memory. Last stated a month ago 10 Oct 2026 SR Sebastian Raschka — holds since 2026-10-10 — tap for who they are Same subject: Clipped policy ratios are technically essential in GRPO. — tap to centre the map on it Clipped policy ratios aretechnically essential in GRPO. Last stated a month ago 10 Oct 2026 SR Sebastian Raschka — holds since 2026-10-10 — tap for who they are Same subject: Using a model as its own activation oracle introduces correlated risk if the model is misaligned. — tap to centre the map on it Using a model as its own activationoracle introduces correlated risk ifthe model is misaligned. 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 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: Quoting a number for P(doom) is a ridiculous notion, because it implies a precision that nobody actually has. — tap to centre the map on it Quoting a number for P(doom) is aridiculous notion, because itimplies a precision that nobodyactually has. Last stated a year ago 23 Jul 2025 DH Demis Hassabis — holds since 2025-07-23 — 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 substitutefor a well-specified conventionalalgorithm, so it cannot simply bedropped into a complex problem andtrusted. Last stated a year ago 7 Jun 2025 GM Gary Marcus — holds since 2025-06-07 — tap for who they are Same subject: An answer computed by a Python program deserves more trust than a language model's mental arithmetic. — tap to centre the map on it An answer computed by a Pythonprogram deserves more trust than alanguage model's mental arithmetic. Last stated 2 years ago 5 Feb 2025 AK Andrej Karpathy — holds since 2025-02-05 — 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: One pre-trained robot model now matches or beats the specialists that were fine-tuned with reinforcement learning for the very tasks they were built for. — tap to centre the map on it One pre-trained robot model nowmatches or beats the specialiststhat were fine-tuned withreinforcement learning for the verytasks they were built for. Last stated 3 months ago 12 Aug 2026 CF Chelsea Finn — holds since 2026-08-12 — tap for who they are Same subject: Reading RL transcripts after a post-training run to find problems in training environments seems a good general practice. — tap to centre the map on it Reading RL transcripts after apost-training run to find problemsin training environments seems agood general practice. Last stated 11 months ago 27 Dec 2025 NN Neel Nanda — holds since 2025-12-27 — tap for who they are Same subject: RL environments exist to make a model generalise, not to teach it each skill one at a time — exactly as pre-training does. — tap to centre the map on it RL environments exist to make amodel generalise, not to teach iteach skill one at a time — exactlyas pre-training does. Last stated 9 months ago 13 Feb 2026 DA Dario Amodei — holds since 2026-02-13 — tap for who they are Same subject: Reinforcement learning only works once a model already knows something, which is why robots must be pre-trained by imitation first. — tap to centre the map on it Reinforcement learning only worksonce a model already knowssomething, which is why robots mustbe pre-trained by imitation first. Last stated a year ago 12 Sept 2025 SL Sergey Levine — holds since 2025-09-12 — tap for who they are Same subject: A highly dynamic economy that cheaply mass-produces robots may increase rather than decrease the risk of Malthusian resource dilemmas. — tap to centre the map on it A highly dynamic economy thatcheaply mass-produces robots mayincrease rather than decrease therisk of Malthusian resourcedilemmas. Last stated 3 years ago 1 Oct 2023 TC Tyler Cowen — holds since 2023-10-01 — tap for who they are Same subject: A humanoid robot is the wrong tool on a factory floor, because a machine specialised in the task beats it. — tap to centre the map on it A humanoid robot is the wrong toolon a factory floor, because amachine specialised in the taskbeats it. Last stated 2 years ago 6 Apr 2025 MH Molson Hart — holds since 2025-04-06 — 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 The KL term in GRPO is not essential; whether to use it is a hyperparameter to decide case by case, and on math models may do better without it. Last stated 3 Oct 2026 · 2 months ago Holds Sebastian Raschka Read this korrent →