1層で十分?単一トランスフォーマー層がフルパラメータRL訓練に匹敵
本論文では、強化学習(RL)を用いた大規模言語モデルの訓練において、単一のトランスフォーマー層のみを更新する手法が、全パラメータを更新する従来の手法と同等の性能を達成できることを示す。実験結果から、特定の条件下では1層の更新で十分であり、計算コストを大幅に削減できる可能性が明らかになった。
本論文では、強化学習(RL)を用いた大規模言語モデルの訓練において、単一のトランスフォーマー層のみを更新する手法が、全パラメータを更新する従来の手法と同等の性能を達成できることを示す。実験結果から、特定の条件下では1層の更新で十分であり、計算コストを大幅に削減できる可能性が明らかになった。
Max Weinbach says he had early access to OpenAI's new model GPT-5.6 Sol, calling it his favorite model by far. He highlights that it never gives up and will keep reasoning until it's done. OpenAI announced that GPT-5.6 Sol, along with Terra and Luna, will launch publicly on Thursday, with preview access expanding globally now.
The US government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for all customers, citing a potential jailbreak technique that involved asking the model to review a codebase for vulnerabilities—a capability Anthropic says is available in other public models. Access was abruptly cut off on June 12.
Andrej Karpathy announces the release of Claude Fable 5, the same underlying model as Mythos but with added safeguards. He calls it a major step forward, particularly for long problem-solving sessions on difficult tasks, and describes it as state-of-the-art on nearly all benchmarks with exceptional performance in software engineering, research, and vision.
Roman Storm warns that the legal theory in his case could set a precedent making open-source developers liable for how others use their code, potentially criminalizing the mere publication of privacy, messaging, or crypto tools. He notes that developer Michael Lewellen cannot publish lawful code due to prosecution fears, and argues this chilling effect extends beyond any single case.
Meta's engineering culture is deteriorating under Mark Zuckerberg and Scale AI CEO Alexandr Wang, who have introduced keyboard tracking, reassignments to data labeling, and AI-centric performance metrics. Critics argue this incentivizes performative AI use, drives away experienced engineers, and contributed to a major Instagram hijacking incident caused by AI-written and AI-reviewed code.