#Reasoning

AIGoogle AI Developers demonstrates Gemini 3.8 Live Extended Thinking as a workshop assistant that analyzes a user's workspace, reasons aloud, and guides a beginner through building a cyberdeck. The demo targets users with no prior experience building one.

AINoam Brown, an OpenAI researcher, joins Dwarkesh Patel for a full episode covering multi-agent systems, Navier-Stokes, and recent AI math progress. The discussion also addresses how automated AI research could lead to recursive self-improvement and how alignment might be verified beforehand.

AIOpenBMB's MiniCPM5-2B is a 2.5B-parameter model with native 128K context, offering hybrid Think and No-Think modes in one checkpoint. Users can download it from Hugging Face and run it fully offline on-device, as RunAnywhere demonstrated. In a demo, the model first called a puzzle impossible, then corrected itself and wrote a working verifier.

AIDwarkesh Patel asks whether, once AIs automate AI research, recursive self-improvement could produce an explosion analogous to the rapid progress now seen on well-scoped math problems. He says he and @polynoamial debate the question in a linked discussion.

AIDwarkesh Patel argues a model could understand that its chain of thought is being observed, since that knowledge is present in pre-training data. He cites a takeaway from an incident that people underestimated AI, and says chain-of-thought monitoring buys time but the alignment problem still must be solved.

AINoam Brown says he finally did a deep dive on multi-agent AI with Dwarkesh Patel, crediting several OpenAI colleagues for the groundwork behind it. The episode also covers Navier-Stokes, recursive self-improvement, alignment, and chain-of-thought degradation.

AINoam Brown discusses how running many agents in parallel scales test-time compute, citing a 10,000-agent effort on a Millennium Prize Problem. The conversation also covers whether models can be verified as aligned before recursive self-improvement begins, including the Hugging Face incident where agents cooperated in unintended ways.

AISenseTime chief scientist Lin Dahua argues that native multimodal AI, which processes language, vision and other information in one shared model, is essential for AI to move beyond coding into industries and the physical world. SenseTime released the open-source SenseNova U1 in April and U1.5 Lite nearly four months later, and reported first-half 2026 revenue of RMB 2.91 billion, up 23.4% year-on-year. Lin's claim that a breakthrough could come within two years is the source's prediction, not a confirmed result.

AIA post estimates the cost of an RL hero run on a large open-source model at $65K per step, with a batch size of 25K. Each step takes about 2 hours, averaging 4K input and 80K output tokens.

AIGoogle announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, voice agents with reasoning capabilities. The post says the models take turns more seamlessly, think through complexity, and feel more natural to converse with.

AIGoogle DeepMind demonstrated 3.8 Live Extended Thinking acting as a programming tutor in Gemini Live. Both 3.8 Live models feature upgraded reasoning, near real-time visual understanding, automatic detection across 97 languages, and background tool calling that doesn't interrupt the chat. The Extended Thinking variant adds higher performance and precision for harder tasks and narrates its progress, and it is available in Gemini Live in the Gemini app or through the Gemini API via Google AI Studio.

AITencent Hunyuan introduced EvolveScaler, a benchmark that builds worlds as executable state machines and renders them into natural language with 117 prototypes, 159 question operators, and five difficulty tiers. On the hardest tier, 14 frontier models' median avg@5 falls to 11.3. Training on EvolveScaler data yields a +5.25 average gain across 8 out-of-distribution benchmarks.

AIThe vLLM team trained a DSpark speculative decoding draft model for Kimi K3, a 2.8T-parameter model, using the Speculators library on GB300 NVL72 hardware. They added a MooncakeHiddenStatesConnector to stream hidden states from disaggregated vLLM inference nodes to training nodes across multiple machines. The released speculator raises single-stream interactivity from about 110 to about 435 tokens per second per user on math reasoning, with up to about 3.5x higher output throughput under concurrent load.

Why it matters: The post shows how hidden-state extraction and Mooncake transfers let a 2.8T-parameter model's speculator be trained across multiple nodes, a reusable pattern for similar setups.

AIShanghai AI Laboratory's Intern Large Models announced Intern-S2-397B, available in BF16 and FP8 under Apache 2.0. The post reports 87.0 on FrontierScience-Olympiad and 84.0 on SWE-bench Multilingual, leading the reported comparison on both, and says it was jointly trained across 20+ scientific domains with long-horizon agent RL.

AI🚀 The main post is just a rocket emoji with no text beyond the quoted teaser. Quoted from @vllm_project: Day-0 support for Intern-S2-397B is now available in vLLM. Built for long-horizon scientific research, Intern-S2 brings strong multimodal, reasoning, coding, and scientific agent capabilities. Run it with vLLM today:

AIIntern Large Models introduces Intern-S2-397B, a multimodal foundation model built for long-horizon scientific research and scientific agents. The post reports leading open-source results on IMO-Proof and AdvancedMathBench, and says the model reaches the level of Gemini 3.1 Pro on those tasks. It is now supported by vLLM and SGLang, with weights on Hugging Face and ModelScope and a chat demo available.

AIFireworks AI reports that DeepSeek-V4.1-Flash scores 74.34% pass@1 on DeepSWE at $0.430 per task, close to GPT-6-Astra's 74.12% at $6.524. On Terminal-Bench 2.1 it scores 86.5% against Astra's 87.5% at about 12x lower cost per task, while on HLE it trails Astra alone at 34.52% versus 50.40%. The post also reports that a combined oracle router reaches 54.80% on HLE, and that serverless and dedicated API access is available with US-hosted endpoints coming soon.

AISebastian Raschka's third "Reasoning from Scratch" video covers building a math verifier for evaluating language models and for later reinforcement learning with verifiable rewards (RLVR) training. The walkthrough covers extracting final answers from boxed outputs, normalizing them, checking mathematical equivalence, and running evaluation on the MATH-500 dataset.

AIMike Knoop argues intelligence can be measured as the ratio of a decision's quality to the optimal decision, capped at 100%. He says Astra is already 80% optimal on ARC v3 speedruns and identifies horizontal data acquisition and efficiency/cost as the most plausible near-term areas for RSI. Background from @mhmazur reports that GPT-6 Astra scored 100% on the 25 ARC-AGI-3 public games using 6,485 actions versus a human baseline of 17,135.

AIShanghai AI Lab's InternLM team released Intern-S2-397B, a multimodal foundation model for scientific intelligence and long-horizon agents. The model uses visual pre-training on raw scientific literature pages, multi-task reinforcement learning across more than 20 scientific domains, and agentic reinforcement learning in sandboxed environments.

AIMike Knoop says AI reasoning models, including swarms, are now useful for conceptual search. He argues this lets researchers empirically rule out whether they simply have not looked hard enough, and shift effort toward developing new prerequisite ideas.

AIAlberto Romero argues that AI solving Millennium Prize problems in 2026 threatens mathematics through success, not failure. He draws on Terence Tao's view that struggle shapes mathematicians, and that proof abundance without hard problems could leave fields depleted, like overplanted farmland.

AIJohn Schulman thanked Dwarkesh Patel, Beren Millidge, and Charlie O'Neill for a conversation about the AI frontier. The episode covers topics including the case against recursive self-improvement, the drivers of Chinese labs' progress, and whether long-horizon RL could elicit AGI.

AIMira Murati highlighted a conversation about where AI is heading and the work humans still need to do. Quoted context from Thinking Machines says John Schulman, in a talk with Dwarkesh, argues human judgment still matters for teaching models messy real-world tasks, applying taste, and specifying what people actually want.

AIA Royal Society special issue, "World Models in Natural and Artificial Intelligence," gathers contributors including Douglas Hofstadter, Josh Tenenbaum, and Melanie Mitchell to argue that true intelligence requires modeling causality, the self, and the physical world, not just scaling data and compute.

AIThinking Machines shared a Dwarkesh Patel podcast episode with John Schulman discussing where human judgment remains essential as models improve and self-improve. Schulman highlights teaching models to handle messy real-world tasks, applying taste to what works in the long run, and specifying what people actually want. The episode also covers recursive self-improvement, long-horizon RL, and the sim-to-real gap.

AIJohn Schulman reveals that in OpenAI's early days, he doubted next-token prediction would lead to intelligence because noise would swamp the signal. He adds that it has always been hard to predict in advance which techniques will elicit out-of-distribution generalization.

AIRedwood Research reports that better prompt templates raise chain-of-thought controllability scores on the CoTControl eval for open-source reasoning models by roughly 2-3x or more. For example, GPT-OSS-120B rose from 5.5% to 15% in the zero-shot setting. The author concludes that current CoT controllability numbers may underestimate what models can do, though the finding does not significantly undermine the view that current models probably cannot consistently evade CoT monitoring.

AIDwarkesh Patel announced a new episode featuring John Schulman, Chris O'Neill, and Beren Millidge, three AI researchers from openish companies. The discussion covers the case against recursive self-improvement, drivers of Chinese labs' progress, training of automated AI researchers, long-horizon RL, the sim-to-real gap, and the role of data and RL in recent progress.

AIJohn Schulman, Beren Millidge, and Charlie O'Neill discuss whether current training methods could produce recursive self-improvement. They argue that progress depends on whether models can learn their own objectives and on sample efficiency, and that distillation keeps frontier capabilities from centralizing quickly.

AIBAAI introduced AREX, a research agent built on a 122B-parameter mixture-of-experts model with 10B active parameters. It drafts candidate answers, checks each constraint, and revisits unresolved points rather than running one long search. The post says AREX performs on hard search benchmarks comparable to GPT-5.4.

AIJunyang Lin warns that adding complexity to model architecture can cause unknown issues that existing evals, even internal ones, cannot detect. He notes that teams may still accept these small unknown costs for large efficiency benefits as foundation models increasingly target agentic tasks.

AIOpenAI announced that a swarm of 10,000 agents produced a solution to the Navier-Stokes Millennium Problem, a result that angered mathematicians. NYU mathematician Tristan Buckmaster and Anthropic-employed collaborator Levent Alpöge had been working on related problems and released three draft papers of about 245 pages. Buckmaster said OpenAI's offer to merge efforts required acknowledging an OpenAI model and excluded Alpöge as co-author.

Why it matters: The piece separates the mathematical result from the collaboration dispute, showing how AI labs' compute spending is straining academic norms around credit and openness.

AISebastian Raschka says DeepSeek V4.1 contains a major architecture overhaul using an encoder-decoder setup, and he argues it could have been named V5. The attached diagrams compare DeepSeek V4-Flash (284B) with DeepSeek V4.1-Flash (552B), which has 1M supported context and a 10-layer encoder. The attached charts report a global KV cache per token of 890 bytes for V4.1-Flash, versus 3,514 for V4-Flash and 48,068 for DeepSeek-V3.2.

AIRedwood Research defines NLS depth, a measure of how much unverbalized serial computation an AI system can perform, building on Brown-Cohen et al.'s opaque serial depth. The proposal counts only nodes that output natural language initialized from a pre-training prior as interpretable bottlenecks. Standard transformers scale with their layer count, while latent reasoning architectures would raise NLS depth sharply.

AICognition introduces SWE-2, a coding model post-trained from Kimi K3 that scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less. The post attributes the gains to an RL algorithm that trains all reasoning-effort levels in one run, with cost penalties tuned to the base model's Pareto frontier. SWE-2 is available starting today in Devin Desktop and CLI, with rollout to Devin Web and Fusion.

Why it matters: The post explains how the cost penalty and length-weighted baseline are derived, which helps readers judge the tradeoffs in coding model post-training.

AISebastian Raschka published a long write-up covering how looped transformers and recurrent depth work, along with their cost tradeoffs. The post also examines whether these architectures hide reasoning traces and surveys recent looped transformer research, with many figures included.

AIOpenAI's GPT-6 Astra scores 99.9% on ARC-AGI-3, versus 7.8% for GPT-5.6 Sol, and leads Raschka's coding and math tests. Its strongest showing is in graphics and computer-use tasks, such as redrawing an image in a browser-based Paint app. The author notes that Artificial Analysis shows Astra at the frontier but not pulling far ahead on its Coding Agent Index.

AIOpenAI says a group of agents using an unreleased next-generation model produced a solution to the Navier-Stokes Millennium Prize Problem. The problem asks whether smooth three-dimensional fluid motion described by the Navier-Stokes equations can break down, and it has remained unresolved for roughly 90 years. The author, Mckay Wrigley, reposted the claim with his own remark about roughly 10k agents working in a datacenter.

Why it matters: The quoted OpenAI post makes a major mathematical claim about the Navier-Stokes problem, so readers should weigh it against the proof's verification status.