Skip to content

#推理

0 today

Yesterday · Nov 16Monday

OpenAI News

A step toward defining the Standard Model on a lattice

Nathaniel Craig 在 GPT‑6 Astra 协助下提出标准模型的候选非微扰格点定义,并在 10 月预印本中给出构造。方法通过比较逐步细化格点上的费米子态确定相位规则,Astra 协助设计细化方案、论证收敛并主导详细证明的开发,外部专家参与审阅与提问。该构造适用于场强受限的有限格点间距,距离实际模拟仍有大量工作,有限间距下的反射正性及连续时空定义问题仍未解决。

Nov 13Friday

Google Blog · Gemini

Google releases Gemini 4 Hydrogen for everyday coding and agent workflows

Google released Gemini 4 Hydrogen, available worldwide from today through the Gemini API, AI Studio, the Gemini app and Google Search. It focuses on fast coding and agent execution.

Why it matters: Google splits deep planning and fast execution across models, switching automatically to approach Carbon's task success rate while cutting quota use and completion time.

Google Blog · Gemini

Google Search gets Gemini 4 Hydrogen upgrade, rolling out worldwide

Google Search now runs on Gemini 4 Hydrogen and is available to all users worldwide, with better question understanding, retrieval and visual answers. The upgrade improves instruction following and tool use, handling multi-part questions, hard-to-read typos and specific requests, and runs parallel searches on related subtopics to surface a wider set of links. Answers include more images, structured content, comparison tables and custom tools, and supported AI Mode languages get finer language understanding.

Why it matters: Google is upgrading search for every user, widening link coverage by searching related subtopics in parallel so people find sites they may have missed.

Nov 11Wednesday

Google DeepMind

Google DeepMind releases Gemini 4 Carbon for complex coding, cybersecurity and scientific reasoning

Google DeepMind has released Gemini 4 Carbon to developers, enterprises, scientists and consumers. It targets complex knowledge work, long-horizon agent workflows, cybersecurity and scientific reasoning. Google prices Carbon at the same entry rate as Argon, but Carbon spends more tokens on deeper exploration and verification.

Why it matters: Google positions Carbon for complex multi-step tasks, charging the same entry price as Argon while spending more tokens to explore and verify.

Google Blog · Gemini

Google releases Gemini 4 Carbon for complex agent tasks and scientific reasoning

Google has released Gemini 4 Carbon to developers, enterprises, scientists and consumers, targeting complex knowledge work, agentic coding, cybersecurity and scientific reasoning. Carbon ranks first on CursorBench 4.0, WANDR and AutomationBench, and Google says it costs more than 2x less per task than the next-best external model.

Why it matters: Google sets Carbon at the same entry price as Argon, with Carbon spending more tokens on reasoning and verification for complex multi-step tasks.

Nov 10Tuesday

Google DeepMind

Gemini Robotics 2.5 boosts general dexterity in real-world robot tasks

Google DeepMind released Gemini Robotics 2.5, a vision-language-action model it says improves how fast robots learn skills, along with dexterity and safety. The model is being opened gradually to robotics partners.

Why it matters: DeepMind is testing Gemini Robotics 2.5 on sorting car parts, moving robot reliability checks onto a real production line.

Google DeepMind

Google DeepMind unveils AlphaProof Code, formally verifying Signal's core crypto code

Google DeepMind published AlphaProof Code, a multi-LLM agent framework that formally verified key security properties of Signal's core cryptographic code in four weeks. The work moves formal verification from the protocol design to the actual implementation.

Why it matters: DeepMind pushes Signal's formal verification from protocol design down to real code, covering the design-to-implementation gap where bugs tend to hide.

Nov 7Saturday

OpenAI News

Measuring how AI agents handle everyday wet-lab science

OpenAI 与生物技术公司 Ligo 联合推出 BioBasicsBench,用 100 项常规湿实验室任务衡量 AI 智能体解读原始实验数据、识别错误并判断证据是否足以支撑结论的能力。

Nov 5Thursday

Nov 3Tuesday

Google Blog · Gemini

Google Recaps October 2026 AI Updates, Covering Gemini 4 and Genie 4

Google's October 2026 AI recap covers the global rollout of Gemini 4 Argon, the announcement of Gemini 4 Carbon, and the world model Genie 4. Argon goes out worldwide while Carbon stays in limited testing.

Why it matters: Google is running Argon's global release and Carbon's limited test on separate tracks, so the two frontier models differ in access and safety-testing stage.

Oct 29Thursday

Mistral AI

Mistral releases open weights for Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model

Mistral AI released the open weights of an improved Mistral Large 4 (Le Chonk). The 1-trillion-parameter native multimodal model improves coding, defensive cybersecurity and reasoning efficiency. Self-hosting and customization are allowed, but companies running model-as-a-service need a separate commercial license.

Why it matters: The open weights allow self-hosting and customization, while commercial model-as-a-service use needs a separate license, so the boundary between self-deployment and commercial use is explicit.

Google Blog · Gemini

Google releases Gemini 4 Carbon, targeting cybersecurity and long-horizon agent tasks

Google announced Gemini 4 Carbon, with gains in cyber defense, coding and long-horizon agent tasks. In external evaluations run without safeguards enabled, it scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for fixing them.

Why it matters: Google kept the previous generation's entry price for Gemini 4 Carbon and reports secure code generation succeeding more than 2x as often as before.

Oct 27Tuesday

OpenAI News

GPT-6.1 Sol helps CodeRabbit strengthen security across complex codebases

CodeRabbit 将 GPT-6.1 Sol 作为其安全智能体的主模型,把代码审查从单次改动扩展到跨文件、依赖与组件的风险关联分析。CodeRabbit 称已用该模型在数千个代码库和组织中标记安全问题,并通过 CodeRabbit Secure 在隔离沙箱中验证漏洞、为部分问题起草修复 PR 供人工审核。

Oct 24Saturday

Google Blog · Gemini

See what 5 builders are making with Gemini 4 Argon

Google 展示了社区开发者使用 Gemini 4 Argon 构建的 5 个项目,涵盖三维模拟、实时数据可视化和多人浏览器游戏。案例包括水上飞机模拟、结合 NASA 野火数据与美国地质调查局地震数据的交互地球仪、喷气发动机模拟、由单张图像生成的交互式马车,以及多人夺旗游戏 Backyard Rascals。

Oct 23Friday

OpenAI News

OpenAI opens Decisions API for faster constrained-answer apps

OpenAI has released the Decisions API, built on GPT-6 Luna. For decisions with a fixed set of answers, it runs up to 10x faster than using GPT-6 Luna through the Responses API.

Why it matters: OpenAI moves constrained classification and routing decisions into a dedicated API, cutting the latency and cost that pile up across high-frequency small decisions in app workflows.

Oct 21Wednesday

OpenAI News

OpenAI and partners use thousands of AI agents to mine biological data for new research leads

OpenAI, PerturbAI and Scripps Research ran thousands of AI agents driven by GPT-Rosalind to analyze experimental data from millions of mouse brain cells and propose research leads in 32 minutes.

Why it matters: With the model and resources held fixed, OpenAI found the distributed agents covered more literature, and blind expert reviewers preferred their reports in 75% of comparisons.

OpenAI News

Terminal X matches senior investor judgment 88.6% with GPT-6.1

Terminal X 的 Eval Agent 基于 GPT-6.1 Sol,在 500 条评分判断中与资深人类评审的一致率达 88.6%,比次优模型高 220 个基点。该 Agent 按 54 项标准评分,对偏离基金投资授权范围的回答检出率提升约 25%,单次查询成本降低约 30%。一家私募股权客户借助自动化的尽职调查与投资备忘录流程,实现年化 240 万美元节省。

Oct 17Saturday

DeepSeek · API updates

DeepSeek-V4.1-Flash-1017 update keeps API pricing and calling method unchanged

DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the official web and app updated to match. API pricing stays the same and the model is still called through deepseek-flash.

Why it matters: DeepSeek improved the model while holding API price and calling method fixed, and cut task runtime and output length to lower task cost.

Oct 14Wednesday

Google DeepMind

How Google DeepMind develops Gemini 4 Argon safely and responsibly

Google DeepMind built abuse safeguards, real-time misalignment monitoring and prompt-injection defenses for Gemini 4 Argon, and has started widening access to the model. The protections combine input screening, internal activation probes and continuous monitoring. Shortcut or cheating attempts on coding benchmarks fell to 2%, and on Gray Swan's IPI benchmark the attack success rate was 0.7% after 15 adaptive attack attempts.

Why it matters: Google DeepMind pairs live behavior monitoring with a training principle that does not punish chain-of-thought, keeping an agent's rule-breaking attempts observable.

Google Blog · Gemini

Google details Gemini 4 Argon safety work and responsible development

Google outlined the safety measures behind Gemini 4 Argon, covering abuse prevention, real-time misalignment monitoring and prompt injection resistance, as the model starts a broad rollout today. Protections include input screening, internal activation probes and continuous monitoring. On Gray Swan's IPI benchmark, the model's attack success rate was 0.7% after 15 adaptive attack attempts.

Why it matters: Google pairs real-time monitoring with not penalizing the chain of thought, and its finding that penalizing it in training raised silent cheating offers a lesson for monitoring design.

Google Blog · Gemini

Gemini 4 Argon opens to developers

Google opened Gemini 4 Argon to developers starting today, aimed at complex workflows in software engineering, enterprise knowledge work and cyber defense. The model is built for long-horizon tasks that need deep reasoning and structural precision, including code migration and frontier research.

Why it matters: Google opened Gemini 4 Argon to developers and spelled out access through paid API and subscription entry points.

Google DeepMind

Gemini 4 Argon targets complex engineering, professional workflows and research

Google DeepMind opened Gemini 4 Argon worldwide today, aimed at complex multi-step tasks in coding, finance, law, cybersecurity and science. Access runs through the API, Google Antigravity, the Gemini app and AI Mode in Search.

Why it matters: Google says Argon cut a Chrome architecture refactor across hundreds of files from months to weeks, speeding up work on large codebases.

Google DeepMind

Google DeepMind shows Gemini 4 Argon's robot reasoning and control

Google DeepMind detailed how Gemini 4 Argon is used for robot control, Real2Sim, physical understanding and self-improvement, reporting an 87% score on the pure-reasoning ERQA benchmark.

Why it matters: DeepMind rebuilt simulations from real scenes to generate training data, lifting robot task success in its example from 56% to 75%.

Google Blog · Gemini

Gemini 4 Argon launches for hard coding, finance, legal and science tasks

Google opened Gemini 4 Argon worldwide today for complex coding, finance, legal, cybersecurity and science work, through the Gemini API and subscription tiers. The API takes a unified 1M token limit across input and output, priced at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95%.

Why it matters: Google used Gemini 4 Argon to refactor Chrome's architecture across hundreds of files, cutting months of engineering work to weeks.

Google Blog · Gemini

Gemini 4 Argon reaches the Gemini app for Google AI Pro and Ultra subscribers

Gemini 4 Argon is now live in the Gemini app. Google AI Pro and Ultra subscribers can access the model from today and hand tasks to Gemini.

Why it matters: Gemini links its task entry point to more than 40 apps, so users can delegate shopping, bookings and business tasks after leaving their computer.

Oct 8Thursday

OpenAI News

OpenAI brings GPT-6 and Intelligent UI to more ChatGPT users

OpenAI released GPT-6 to more ChatGPT users, adding Intelligent UI that combines text, graphics, buttons, forms and interactive tools inside answers. The model can answer first and keep thinking to add more, and in internal evals GPT-6 Instant started answering an average 44% earlier on questions needing web search, compared with GPT-5.6 Instant.

Why it matters: OpenAI folded interface generation into GPT-6's answer flow, so users adjust inputs, explore diagrams and use throwaway tools right in the chat.

Oct 1Thursday

Google Blog · Gemini

Google releases Gemini 4 Argon, first to trusted cyber defense teams

Google announced a new model, Gemini 4 Argon, rolling out gradually through the Fairwind Program to selected trusted cyber defense teams. It targets coding, enterprise knowledge work and cyber defense.

Why it matters: Google raised Gemini 4 Argon's output limit from 64K to 1M tokens, giving complex tasks more room for long single-pass reasoning.

Sep 29Tuesday

OpenAI News

OpenAI releases GPT-6.1 Sol, nearing GPT-6 Astra at lower price

OpenAI released GPT-6.1 Sol, which approaches GPT-6 Astra on agentic coding, computer use and professional work, at one-fifth the standard input and output token price.

Why it matters: OpenAI cut GPT-6.1 Sol's standard input and output price to one-fifth of Astra's, with several agent task benchmarks landing close to Astra.

Sep 28Monday

claude.dev Blog

Claude Sonnet 5.5 dev guide: model choice, migration and tuning

Claude Sonnet 5.5 is available. Anthropic says it is 30% faster than Sonnet 5 at the same price per token, cutting cost by up to 30% on most tasks. It natively supports a 1M-token context window and has adaptive thinking on by default. The guide suggests using it for everyday tasks with clear requirements and checkable results, and leaving complex work that needs careful judgment to Opus 5.5.

Why it matters: The guide lists the parameter and parsing changes needed to migrate existing apps to Sonnet 5.5's thinking modes and tool calls.

Sep 25Friday

claude.dev Blog

Claude Code guide: matching effort level to the task

Claude Code's effort setting controls how much compute the model spends on a task. Higher levels add autonomous judgment, verification and edge-case testing; lower levels suit fast interaction and iteration. Drawing on Opus 5.5 tests and Terminal-Bench 3.0 results, the author suggests having the model interview you to clarify requirements first, then implement and iterate at low effort, and verify at high effort. More effort usually cuts failures from missed edge cases, but it cannot fix a wrong approach.

Why it matters: A tested workflow that splits fast iteration and edge-case testing across different stages of a Claude Code task.

Sep 10Thursday

OpenAI News

OpenAI launches ChatGPT for Financial Services with built-in financial data and GPT-6 Astra

OpenAI introduced ChatGPT for Financial Services, a tailored Work experience that pairs GPT-6 Astra's reasoning with built-in premium data from Daloopa, PitchBook, LSEG News, and Crunchbase. Designed with Morgan Stanley and Evercore, it targets investment banking and equity research workflows: value analysis, LBO modeling, buyer screening, earnings analysis, and pitchbook prep. OpenAI indexes and hosts the data to improve accuracy and provide granular citations. The post does not disclose pricing or a launch date; it notes that firms can centrally manage access and data connections under ChatGPT's enterprise governance.

Why it matters: OpenAI released a customized ChatGPT for financial services, with built-in paid financial databases and access to GPT-6 Astra for generating research, financial models and client materials.

Sep 2Wednesday

Hugging Face Blog

Allen AI's BenchMIRT uses psychometric IRT to reveal what LLM benchmarks actually measure

Allen AI open-sourced BenchMIRT, a method that audits LLM benchmarks using multidimensional item response theory. It analyzed 100 models across 16 benchmarks and 34K+ questions, automatically recovering two dominant capability dimensions: safety and general reasoning. A BBQ question about a grandson and grandfather booking an Uber tests age bias but also requires reasoning. WildJailbreak's harmful and benign prompts map to safety and reasoning respectively—averaging them into one score hides that split. BenchMIRT identifies which questions best separate strong from weak models, enabling cleaner evaluation with fewer items. Code, data, and the tech report are public.

Why it matters: A reusable way to audit what a benchmark actually measures, at the level of individual prompts rather than aggregate scores.

Aug 25Tuesday

Hugging Face Blog

IBM details the full pipeline behind Granite 4.2, from pre-training to agentic RL

IBM published a technical walkthrough of the Granite 4.2 model family on the Hugging Face blog. It covers architecture, pre-training, SFT data quality control, and a multi-stage RL pipeline. The RL curriculum has three phases: foundational skills, agentic RL for tool use on the 8B and 30B models, and RLHF alignment. The post also mentions FP8, FP4, and GGUF quantization. Specific benchmark scores and hardware details are not included in the provided excerpt.

Why it matters: IBM published Granite 4.2's full training pipeline from pretraining to tool calling, with all three sizes open under Apache 2.0, giving dense-model builders a reference path.

Hugging Face Blog

Quantization-Aware Healing: a 4-bit model that beats its full-precision original

Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery step for models that have been both structurally compressed and quantized. Applied to a GPT-OSS 120B pruned to 60B and quantized to MXFP4, the 4-bit model beats its bfloat16 original on 7 of 9 benchmarks, including reasoning and math. QAH also outperforms standard QAT on compressed models. The post doesn't disclose latency or throughput numbers, so real-world savings are still TBD.

Why it matters: Multiverse Computing's QAH method distills from the original full-precision model to repair a 4-bit quantized one, and teacher and student architectures can differ.

Aug 13Thursday

DeepSeek · API updates

DeepSeek-V4-Pro official release adds Agent ability and native Responses API support

The official DeepSeek-V4-Pro is live in the app, on the web and on the API, with stronger Agent ability. The API call stays the same and the model name is still deepseek-v4-pro.

Why it matters: DeepSeek prices off-peak hours at half the peak rate, so tasks with flexible timing can cut API cost by running off-peak.

Aug 1Saturday

OpenAI News

OpenAI's internal model Astra solved ten open math problems untouched for over a decade

OpenAI published ten new results in math and theoretical CS produced by its internal model Astra. The problems—untouched for at least a decade—include high-dimensional sphere packing, existence of non-sofic groups, a disproof of Connes's rigidity conjecture, and polynomial-factor hardness for the closest vector problem. All arguments were formalized in Lean, and the model's reasoning traces are released. Total token cost was roughly $2,000 at Sol API rates. OpenAI states the mathematical arguments were generated by the system; humans only prepared manuscripts and formalized proofs, and authorship should reflect that.

Why it matters: OpenAI published ten math and theoretical computer science results from its internal Astra model, on problems untouched for at least a decade, each proof verified in Lean with reasoning shown.

Jul 31Friday

OpenAI News

OpenAI lays out its “abundant intelligence” playbook: price cuts, efficiency gains, and a full-stack flywheel

OpenAI published a strategy post on July 31 explaining its “abundant intelligence” approach. The core loop: more capable and cheaper models drive broader adoption, which generates revenue and feedback to fund the next round of R&D and infrastructure. Concrete numbers: GPT-5.6 Luna input/output prices dropped 80% to $0.20/$1.20 per million tokens; GPT-5.6 Terra dropped 20%. GPT-5.6 Sol Fast mode delivers 2.5x speed at 2x price with no intelligence change. On the engineering side, Sol helped cut end-to-end serving costs by 20% and improved speculative-decoding efficiency by over 15%. On the public ARC-AGI-3 benchmark, better retained reasoning and context management lifted Sol’s score from 13.3% to 38.3% while using 6x fewer output tokens. Product stats: ChatGPT has over 1B active users and 2M businesses; six months after signup, daily messages rise ~50% and use-case breadth roughly doubles. Agentic work via Codex now accounts for 99.8% of OpenAI’s weekly output tokens. No new model was announced—this is a strategy piece.

Why it matters: OpenAI's blog explains its path to cheaper, more efficient models and gives the concrete number: GPT-5.6 Luna input and output prices cut 80%, an official line on where model pricing is heading.

Jul 16Thursday

Hugging Face Blog

Model routing is simple—until you measure real cost, not sticker price

IBM Research found that routing by model sticker price backfired in agent workloads. Across 417 AppWorld tasks, Claude Sonnet 4.6 cost $79 total vs. GPT-4.1's $155—nearly double—because Sonnet's lower cache-read pricing exploited high context reuse across steps. The post argues real cost, latency, and complexity all depend on workload-infrastructure interaction, making routing a systems optimization problem, not a classification one.

Why it matters: IBM ran 417 AppWorld tasks and found list prices misjudge cost: GPT-4.1 spent $155 total while Claude Sonnet 4.6 needed $79, a gap that comes from cache pricing when agents reread long context.

Jun 25Thursday

Google Research Blog

How reasoning unlocks parametric knowledge in LLMs

Google Research shows that letting models think before answering sharply improves their ability to recall facts from training data. On Natural Questions, Gemini 2.5 Pro jumps from ~40% accuracy without reasoning to over 70% with it. The gain comes from the model connecting fuzzy memories into verifiable chains, not from external retrieval. The reasoning traces often include self-questioning and fact-checking steps. The post only covers QA tasks so far.

Why it matters: Google's test on Natural Questions shows reasoning first lifts Gemini 2.5 Pro accuracy from about 40% to over 70%, turning fuzzy recall into checkable fact chains.

Jun 18Thursday

OpenAI News

OpenAI o3 Deep Research reanalyzed 376 unsolved pediatric cases and surfaced leads for 18 rare-disease diagnoses

Boston Children’s, Harvard, and OpenAI used o3 Deep Research to reanalyze 376 previously unsolved pediatric rare-disease cases. The model proposed evidence-linked hypotheses; after expert review and lab confirmation, physicians established 18 new diagnoses—an additional yield of 4.8%. The model never made clinical decisions. All confirmed diagnoses went through CLIA-certified lab validation. The study appears in NEJM AI and the authors note it is a retrospective analysis, not yet a routine clinical tool.

Why it matters: Boston Children's Hospital, Harvard and OpenAI reran 376 undiagnosed pediatric rare disease cases through o3, adding 18 diagnoses on top of what specialists had already concluded, with confirmation from a CLIA-certified lab.