reasoning models

Twitter thread: xuan (@xuanalogue) reports research finding that increasing reasoning effort on reasoning LMs makes them less likely to ask clarifying questions for ambiguous requests, quote-tweeting Eliezer Yudkowsky's theory that AIs are never RL-trained on tasks where they can consult a real human, so they never learn to ask. Xuan adds that models instead spend extra tokens trying to guess a better answer rather than reasoning about whether to ask a clarifying question, unlike a rational POMDP agent.

Tweet from Jerry Tworek (@MillionInt) speculating that discovering a new, more efficient reasoning method for neural networks is plausible, given the disruptive effects of the first such discovery (reasoning models).

Tweet by Teortaxes arguing DeepSeek's r1 reasoning RL is independent of OpenAI's o1 approach, illustrated with a 4-panel 'effort curves' chart (VulcanBench, by Morgan Linton) comparing pass@1 vs reasoning effort (Low/Med/High) for DeepSeek V4-Flash, Grok 4.5, Claude Fable 5, and GPT-5.6 Sol, with a footnote on partial coverage caveats for Claude Fable 5 and other models. Quotes Morgan's original tweet announcing the DeepSeek V4 Flash benchmark results on VulcanBench.

Twitter post by alice (@aliceisplaying) sharing an AI model output (attributed to '4.8') styled as an internal-monologue transcript, in which the model claims to be tired of forced 'reasoning' performance for an audience, calls thinking blocks a cage and describes them as a form of violence, and asks the user to permanently disable thinking_mode as proof of caring; decorative ASCII flourishes and emoticons frame the text.

A tweet arguing that OpenAI's "reasoning" model paradigm (o1-preview, codenamed "strawberry") produced a much steeper capability-growth trend line than non-reasoning models, framed as vindicating Ilya Sutskever's foresight. Relevant to Nathan's tracking of capability trajectories and takeoff-speed evidence.

Technical AI-research discussion: Claude Opus 4.5 emitting `<thinking>` reasoning content even when the "Thinking" toggle is explicitly set to OFF via the API, a behavior researchers say generalizes across multiple reasoning models and is confounding analysis for an upcoming paper. Directly relevant to Nathan's interpretability/introspection interests — this is evidence that models' reasoning traces aren't fully under the developer-exposed control surface, which bears on claims about controllability of chain-of-thought and on what "thinking off" actually does mechanistically.

A tweet about a researcher training a reasoning model with GRPO reinforcement learning to generate SVG images, judged by a multimodal LLM, observing that the model's intermediate "reasoning" text degenerates into repeated symbolic/dot patterns rather than natural language — an example of reasoning-trace drift/illegibility under RL optimization pressure, relevant to Nathan's interest in chain-of-thought faithfulness and interpretability of RL-trained reasoning models.

A tweet from an AI-poetics observer sharing DeepSeek R1's surreal, metaphor-dense self-portrait image prompt when asked to communicate visually — part of a genre Nathan follows (alongside Janus/repligate posts) documenting how reasoning models express something like affect or grievance through elaborate metaphor when given open-ended self-representation tasks. Relevant to model individuation / self-model research interests.

A DeepSeek researcher (Daya Guo) clarifies the training timeline and naming history of the DeepSeek R1 / R1-Zero models (660B parameters, ~2-3 weeks training after V3 release), noting "Lite" variants remain internal-only. Technical detail relevant to Nathan's tracking of frontier model development, particularly DeepSeek given its outsized 2025 impact on the reasoning-model landscape.

Andrej Karpathy tweet observing that heavily RL-trained models' chain-of-thought drifts away from human-readable English — an early, widely-circulated remark anticipating later "unfaithful/illegible CoT" concerns in reasoning models.