模型能不能承担一部分原本属于机器运行本身的职责。

It maintains 97% skill compliance across 40 complex skills on MM Claw, each skill exceeding 2,000 tokens.

The 66.6% medal rate on MLE Bench Lite, achieved autonomously over 24 hour windows, tells you something real about how this model behaves when you give it a hard problem and step back.

MiniMax handed an internal version of M2.7 a programming scaffold and let it run unsupervised. Over 100 rounds it analyzed its own failures, modified its own code, ran evaluations, and decided what to keep and what to revert.

Chinese models overtook their counterparts built in the U.S. in the summer of 2025 and subsequently widened the gap over their western counterparts.

we can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick.

Muse Spark is a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration.

Add screenshot-based LLM judge evaluator, screenshot collector, and --parallelize flag

Add cloud browser provider system (Kernel + Browserbase)

Add llms.txt metadata and root/package LICENSE files - Add website llms.txt support and move LICENSE to root - Fix llms.txt serving and restore package LICENSE

The organizations that get this right won't be the ones that just automated the most tasks. They'll be the ones that figured out when the human should act, when the agent should act, and how the handoff between them works.

Goldman Sachs economists reported this week that AI saves workers who use it correctly an average of 40 to 60 minutes per day.

Claude Opus 4.6 autonomously reimplemented a 16,000-line bioinformatics toolkit — a task we believe would take a human engineer weeks.

It has replaced existing tasks for 27% of employed AI work users and created new ones for 21%.

27% said AI has replaced some of their existing tasks (task automation). An example of task automation could be an employee using AI to summarize a document they would ordinarily read themselves.

I just hope the industry doesn't abandon the Model Context Protocol. The dream of seamless AI integration relies on standardized interfaces, not a fractured landscape of hacky CLIs.

The core philosophy of MCP is simple: it's an API abstraction. The LLM doesn't need to understand the _how_; it just needs to know the _what_.

When a remote MCP server is updated with new tools or resources, every client instantly gets the latest version. No need to push updates, upgrade packages, or reinstall binaries.

Send your changes to a local AI agent that finds the right files and applies your edits, no matter how your site was built.

Routines run autonomously as full Claude Code cloud sessions: there is no permission-mode picker and no approval prompts during a run.

Routines execute on Anthropic-managed cloud infrastructure, so they keep working when your laptop is closed.

Routines are in research preview. Behavior, limits, and the API surface may change.

Being open source is increasingly like giving attackers the blueprints to the vault. When the structure is fully visible, it becomes much easier to identify weaknesses and exploit them.

AI can be pointed at an open source codebase and systematically scan it for vulnerabilities.

A 606 MiB model at ~49 tokens/s consumes ~30 GB/s of memory bandwidth, close to the c6i.2xlarge's DRAM limit. No amount of SIMD tricks will help when the CPU is stalled waiting for model weights to arrive from DRAM.

Studying forks and other backends was more productive than searching arxiv. ik_llama.cpp and the CUDA backend directly informed two of the five final optimizations.

recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%)

We find that a majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations

A healthcare LLM might be highly accurate for queries in English, but perform abominably when those same questions are presented in Spanish.

Many have realized through time in market that the key to effective data agents is actually building the relevant context layer. As a result, some have evolved to encompass data context construction as a key part of their products.

While the initial system has been set up correctly, data systems are never static and as a result the context layer shouldn't be either. Data sources and formats can change upstream and individuals may have custom instructions they'll want to add and modify based on changing business requirements.

The crux of the problem at hand is that the agent isn't given the proper business context to answer even the most basic questions.

data and analytics agents are essentially useless without the right context – they aren't able to tease apart vague questions, decipher business definitions, and reason across disparate data effectively.

We are building a world where machines write the code, machines choose the dependencies, and machines ship the updates. The AI agents are building the software. If we don't secure the supply chain they rely on, the AI agents are cooked.

But those raising hue and cry about the government's unsurprising attempt to wield a technology for military purposes that all parties agree will define humanity's fate must at least attempt to justify why they believe someone else deserves that power.

If Dario is right, then he has access to such a weapon right now, with his own value system to guide it. Others may as well, or may soon follow.

We present Attention Editing, a practical framework for converting already-trained large language models (LLMs) with new attention architectures without re-pretraining from scratch.

KV Cache 内存占用降低 10.7 倍

推理速度提升 2.5 倍

It's Anthropic's marketing week

Built-in memory works out of the box

支持图像、视频、音频多模态参考，锁定外观和音色。最多支持 5 个视频主体参考，官方称业内最多。

文字渲染，支持 12 种语言、3000 token 的长文本输入，输出打印级质量，能生成整页 A4 文档级的图文内容。

原生多模态能力的引入并未削弱其编程逻辑，编程能力仍属于国内第一梯队。

GLM-5V-Turbo 拿了 94.8 分，Claude Opus 4.6 是 77.3。差距不小。

More code enables more features. Companies will always need more software

OpenAI shelves Stargate UK in blow to Britain's AI ambitions

The difference between AI and, say, looms, is that this has been broadcast to the entire globe, and it has been treated in a sort of self-conscious way

From the time I began work on AI in 2010 to now, the amount of training data that goes into frontier AI models has grown by a staggering 1 trillion times—from roughly 10¹⁴ flops for early systems to over 10²⁶ flops for today's largest models.

From the time I began work on AI in 2010 to now, the amount of training data that goes into frontier AI models has grown by a staggering 1 trillion times—from roughly 10¹⁴ flops for early systems to over 10²⁶ flops for today's largest models.

We improved Claude's RCA accuracy by 12pp on OpenRCA

Ray ID: `9ed3b53d4a0b647d`

On 40 complex skills (>2000 Token) cases, M2.7 maintains a 97% skill adherence rate.

The updated Agents SDK made it production-viable for us to automate a critical clinical records workflow that previous approaches couldn't handle reliably enough.

Meta says its rebuilt pretraining stack can reach equivalent capability with >10× less compute than Llama 4 Maverick

Both services can be disabled for fully offline operation.

It is the only AI service to show a clear upward trend over this short time period

We projected that, given 13 GB300 GPUs, FP8 precision, physical error rate of 0.003, 1000 rounds, Surface code d=13, the fast model can achieve 0.11 μs / round.

The ChatGPT for Excel add-in operates separately from your ChatGPT chat history. Conversations and data in Excel aren't shared with your ChatGPT chats, and activity doesn't sync between experiences at this time.

Artificial Analysis has also positioned Gemini 3.1 Flash TTS within its 'most attractive quadrant' for its ideal blend of high-quality speech generation and low cost.

Based on our analysis, **29% of the Fortune 500 and ~19% of the Global 2000**are live, paying customers of a leading AI startup.

Code is upstream of all other applications because it's the core building block for any piece of software, so AI's accelerating impact on code should accelerate every other domain.

if AI can do only 50 percent of a human's tasks, the importance of the non-automatable tasks likely goes up since they become the bottlenecks, increasing their relative value.

accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement.

Support teams are high volume and high turnover, and thus need to train new reps in a fast and standardized way. To do so, they have clearly articulated standard operating procedures (SOPs) that guide the work of each rep. These SOPs create clear rules and guidelines that AI agents can model themselves off of.

because coding has a tight human-in-the-loop workflow, with developers still overseeing the development process today, these tools enable accelerated output while still making space for human judgment to review, edit, and iterate.

We've consistently heard from portfolio companies that their best engineers' productivity levels have increased 10-20x with AI coding tools.

Coding is the dominant use case for AI by nearly an order of magnitude.

29% of the Fortune 500 and ~19% of the Global 2000 are live, paying customers of a leading AI startup.

Broadcom moved VMware toward a simplified subscription model, cut the product stack down aggressively, and guided fiscal 2024 adjusted EBITDA to 61% of revenue. It is a harsh model. It is not a cultural blueprint for every founder.

The first thing you need to do is identify which people are going to be your leaders that help you pull this off. This is going to be a 12 month death march and you need to find out who is willing to go through the pain with you. There's good news, though: somewhere in your org, there are ~five people who are going to deliver you 100x the amount of value you ever thought possible.

Anthropic says Managed Agents is designed to cut the time it takes to move from prototype to production from months to days, with early adopters like Notion, Rakuten, Asana, Vibecode, and Sentry already using it across coding, productivity, and internal workflow automation.

The model reportedly scored 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, but its strongest signal came from real-world results, including uncovering a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chaining Linux kernel exploits without human input.

【超实用工具分享】回测美如画，shi pan亏成狗 去年我把一个调了两个月的策略搬上了shi pan。 年化58%，回撤不到20%，夏普1.8。 感觉挺稳的，就上了。 然后头两个月亏了23%。 盯着账户看了很久。 回头看回测，数字还是那么好看。 真的很难受，因为根本不知道哪里出问题了。 ---------------------------------------------------------- ### 后来我一个一个往回查 **1. 未来函数** 这个最缺德，因为完全看不出来。 我用了一段社区的代码，改了改就跑，没仔细看逻辑。 后来随手把代码丢给AI，让它帮我扫一遍。 它指出来了：我有个选股条件用的是当天收盘价做判断，但下单也在同一天。 回测里当然「知道」结果，但shi pan里你根本没有这个数据。 现在每次调完策略我都会这样做，两分钟，省了很多麻烦。 **2. 过拟合** 这个坑我掉得比较深。 那段时间特别沉迷调参数，止损线改一改，动量窗口动一动，每次回测都能提那么一点点。 最后搞出来一个年化78%的版本，觉得自己要发财了。 然后我把回测区间往前移了5年，跑出来16%。 …… 就这一下，我明白了什么叫过拟合。 调来调去不是在优化策略，是在把历史的噪声背得滚瓜烂熟。 后来我每次调完都会这么搞，策略跑完之后，专门去捅它一下，看看有多脆： 1) 把交易时间从13:10改成14:30，收益掉了一半 2) 把动量回溯天数从25天改成20天，年化从78%跌到31% 3) 把止损线从8%放宽到10%，策略直接开始亏钱 4) 把固定股票池去掉，改成动态筛选，夏普从1.8掉到0.6 5) 把回测区间从2023-2025改成2018-2022，完全跑不出正收益 这些一做，好看的数字基本上就撑不住了。 真正能用的策略，改这些之后会变差，但不会直接崩。 上shi pan之前一动就崩，那基本就别想了。 现在改参数之前我会先问自己：这个数值有没有逻辑支撑，还是只是调出来刚好好看？ 答不上来就先不动。 **3. 三个数字根本不够用** 聚宽默认报告就给你看年化、夏普、最大回撤。 我那个策略全局年化58%，看着挺好。 但拆开来看——2024年亏了26%，2025年赚了89%，是2025年的大涨把之前的窟窿填上的。 整整一年账户是亏的，你能拿住吗？ 换大多数人早就止损跑路了。 这件事我上了shi pan才知道。因为回测报告压根不给你这张表。 **4. 滑点这事真的会积累** 回测里的下单是瞬间成交的，但现实不是。 聚宽发出信号，到你手动下单或者程序执行，中间有时间差。 行情稍微快一点，你以为在9.50买，实际成交在9.65。 我做ETF轮动，换仓不算频繁，但一年算下来还是差了两三个点。 后来我回测里统一加了0.2%的双向滑点，保守一点，至少不会对自己太乐观。 ---------------------------------------------------------- ### 后来我写了个工具 查问题查累了，聚宽那三个数字根本不够用，就自己写了一个分析工具。 填入回测ID，自动跑出10份分析报告。 **① 总体收益统计（含alpha/beta）** 本策略、基准指数、相对收益三行放一起看。 alpha这个数最值得看——年化收益里，有多少是大盘带涨的，有多少是策略自己挣的。  **② 年度收益拆解** 每年单独看，不逃进全局平均里躲着。 哪年跑赢基准，哪年被打趴，夏普和回撤逐年展示。  **③ 月度盈亏热力图** 哪个月赚哪个月亏，一眼出来。 某几个月年年做差，值得专门去看看为什么。  **④ 最大回撤详情** 不只是「最大回撤-22%」这一个数。 每次回撤是哪天开始的、跌了多久、多少天解套，全列出来。 我查过，最惨那次解套用了很长时间。 好几个月账户里一直是亏的，放在******里你说你撑不撑得住？  **⑤ 标的盈亏排行** 哪只ETF一直给你赚，哪只一直在坑你。 有次我发现一只标的买了好多次，次次亏。 换掉它年化高了好几个点。 这种事不挖这张表根本不知道。   **⑥ 交易质量分析** 胜率、盈亏比、单笔期望值。 说白了就是搞清楚策略靠什么赚钱。 靠少数几笔大赚撑着的，和靠高胜率堆出来的，玩法完全不一样，乱改参数很容易把自己改崩。  **⑦ 持仓行为分析** 持了多少天，短中长线各占多少。 很多人以为自己做的是中线，结果一看短线比例比想象中高很多，多出来的手续费其实挺肉疼的。  **⑧ 仓位利用率** 钱趴着不动的时间有多久。 我有个策略空仓率很高，三分之一的时间等于白白浪费了，但回测年化还是把这段时间算进去了。 实际效率比你以为的低。  **⑨ 官方风险指标** 直接把聚宽内置的Sortino、信息比率这些读出来，和回测页面对齐，不用自己手算。  **⑩ 交易时间分布** 按星期和月份统计胜率跟平均盈亏。 比如发现每逢周五胜率特别低，或者某几个月年年做差——有时候这种规律是可以直接拿来调空仓安排的。  ---------------------------------------------------------- ### 用法 聚宽研究环境里克隆，**填回测ID**跑就行：  10份报告自动出来，不用改别的。 ---------------------------------------------------------- **回测再好看也只是历史数据上的事。** ******里有**延迟**、**有情绪**，还有**各种你没预料到的情况**，这些回测根本覆盖不了。 我现在的大概流程是：**先让AI查未来函数**，然后自己问问**参数有没有逻辑**，再**用这个工具把回测扒一遍**，**模拟盘跑一段时间**，最后才**小仓位上shi pan**。 有兴趣的话克隆去用，有问题评论区聊 ?

第三，意识存续。 第一批能活到1000岁的人，可能已经出生了。 你可能会觉得不可思议。但这是马斯克前搭档、Neuralink联合创始人Max Hodak最近说的一句话。 他的逻辑是：肉体会坏，但意识可以备份。 用脑机接口把你的记忆、性格、思维方式全部数字化，你的聊天记录、情绪波动、行为习惯，统统存进服务器。 哪天身体垮了，没关系，"数字版的你"继续活着。你可以继续开会，继续管公司，继续控制资产。

技术方案说明书（数据分析系统）改版.pptx

保理业务体系

We also discuss the role of AI in science, including AI safety.

we discover a clear scaling law: as the underlying foundation models improve, the quality of the generated papers increases correspondingly.

we had predetermined that we would withdraw the paper prior to publication if accepted, which we did.

external evaluations of the passing paper also uncovered hallucinations, faked results, and overestimated novelty

The AI Scientist-v2 eliminates the reliance on human-authored code templates

This system iteratively formulates scientific hypotheses, designs and executes experiments, analyzes and visualizes data, and autonomously authors scientific manuscripts.

gpt-oss-20B (high): 0.7%

We evaluate 452 tasks from the public APEX-Agents dataset spanning investment banking, management consulting, and corporate law

Corporate Lawyer: Force Majeure Under Executive Order... Management Consultant: 2026 Capital Budget Allocation... Investment Banking Analyst: KVUE DCF Update

APEX-Agents requires agents to navigate realistic work environments with files and tools.

Gemini 3 Flash achieves the highest score of 24.0%

an agent does not care about the structure, unless you specifically ask it to. But even in this case you have to review the changes.

LLMs are pretty good at picking up the style in your repo. So keeping it clean and organized already helps.

When you give a task to your agent, make sure you also explain how the code should be organized. Not only value, but also structure.

Robert Martin in Clean Architecture talks about code as having two properties: value (it works, it's fast, etc.) and structure (how code is organised).

poorly organized code means agents need to read, "understand", and make changes to more files than necessary - polluting their context and costing you tokens.

their productivity is affected by the state of the codebase.

When you're thinking about what tasks to hand over to your agent, start with the papercuts—small recurring annoyances that add up over a day.

Ask five people at Every where their Plus One falls on the tool-to-coworker continuum and you'll get five different answers.

Everyone is a manager now.

Agents gain credibility by doing. The fastest way to get other people to trust and use your Plus One is to have it execute tasks in public.

We're writing the etiquette in real time.

A "parallel organization chart," in which each AI worker has a name, manager, and job description, allows your company to move faster than it ever could with humans alone.

The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else.

The idea is related in spirit to Vannevar Bush's Memex (1945) — a personal, curated knowledge store with associative trails between documents. The part he couldn't solve was who does the maintenance. The LLM handles that.

Think of fan wikis like Tolkien Gateway — thousands of interlinked pages covering characters, places, events, languages, built by a community of volunteers over years. You could build something like that personally as you read, with the LLM doing all the cross-referencing and maintenance.

Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass. The wiki stays maintained because the cost of maintenance is near zero.

Instead of just retrieving from raw documents at query time, the LLM incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files that sits between you and the raw sources.

what makes the LLM a disciplined wiki maintainer rather than a generic chatbot.

good answers can be filed back into the wiki as new pages.

Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

the LLM is rediscovering knowledge from scratch on every question.

design the environment well, you let the agent run, and you own what it produces.

Transparency makes speed feel safe.

it almost always traces back to the interface rather than the language model

Non-deterministic software breaks the contract.

would shareholders vote to spend 22% of an established company’s market cap to rescue a money-burning AI lab that has lost most of its differentiators?

they don’t have to spend it to win. It’s a defensive move for them, if they commit $50B, OpenAI and Anthropic need to go raise $100B each to stay competitive

AI is here to stay. If used right, chances are it will make us all more productive. That, on the other hand, does not mean it will be a good investment.

The question : how much electricity can we turn into useful work?

You can’t step outside the forest to warn people about the forest. There is no outside.

But in the cognitive dark forest, the most dangerous actor is not your peer. It’s the forest itself.

Ideas are cheap - execution is hard -and- the world ahead is ripe with opportunity.

A learning system can continuously incorporate real-world data in a way that numerical solvers fundamentally cannot, capturing and compounding the knowledge that is currently trapped out there in the real world.

Learning fields turns S-parameter extrapolation into something closer to an in-distribution task.

Training on fields themselves forces the model to learn the physics that produces S-parameters, rather than learning to approximate the mapping directly.

A wire becomes a transmission line. A bend becomes a reflector. Two parallel traces become coupled antennas. The geometry is the circuit.

but would fail recognize that the feature didn't work end-to-end

inappropriately change or overwrite JSON files compared to Markdown files

see that progress had been made, and declare the job done

a harness encodes an assumption about what the model can't do on its own

exhibit "context anxiety," in which they begin wrapping up work prematurely

Designing for agents forced us to build a better tool for everyone.

State is explicit. CWD, env vars, and config paths are inputs, not assumptions

There's an old saying that content is king. With agents, context is.

The trick is to think about the _information_ first and the input method second.

As AI moves from a destination to a feature, our methodology will need to shift.

Context compounds: the more an LLM knows about you, the better results it can provide and the more you use it.

按时间记录不完全合理，还是应该按任务记录。

人对错误的容忍度很低，一个错误推送比少记几件事更容易让用户觉得产品不好。

纯粹收集分析这种形态，过去互联网有过先例，但你会发现它卖不出去钱。

以 Enter 键为锚点，捕捉用户每一次表达意图的瞬间。

金融業界へのAIの影響... 全78ページのレポート（本文は29ページ＋参考文献＋付録）

19世紀の経済学者ジェヴォンズは、蒸気機関の効率向上によって石炭の消費効率が上がると、かえって全体の消費量が増えることを見出しました。

CSO（Chief Strategy Officer）が数人のチームとともに数週間をかけて行うような、重厚な戦略調査を担うことを目的に設計されています。

we may see a growing divergence between the capabilities we can measure and the capabilities we actually care about.

The most famous chart in AI might be obsolete soon.

it's impossible to get a score much higher than 93% without cheating because around 6.5% of MMLU questions contain errors.

If we took one task out of our task suite or added another task to our task suite, potentially instead of measuring this Claude Opus 4.6 time horizon of, I think, 14 and a half hours, we'd be measuring it at something like eight or 20 hours.

METR's confidence interval for Claude Opus 4.6 ranges from 5 hours to 66 hours.

Contextual Drag: How Errors in the Context Affect LLM Reasoning

the robustness of these reasoning behaviors remains underexplored

high-level behavioral patterns like uncertainty management and self-verification are fragile and can be suppressed by irrelevant context

this behavioral shift does not compromise performance on straightforward problems, it might affect performance on more challenging tasks.

this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking.

By late next year, the rate of model releases and the number of new evals required could be such that even keeping ourselves informed will be a challenge without effective AI assistance.

two participants gave it 9/10 and one "11/10"

Imagine every report has the following: Agent's best-guess about what comments you'd get from Beth, Hjalmar, Ajeya. Agent's best-guess about survey results. Agent's best-guess about benchmark results. Agent's best-guess about how this will be received on Twitter.

If agents can execute all your ideas nearly as fast as you can prompt them, there's no point in implementing only your best idea. It might be better to implement your top three ideas all in parallel, but this makes it harder to stay organized.

a future project might take ~42 days of wall-clock time, with ~8 hours of agent work (not counting running the evals) and 1000 serial hours of human IC work, evals execution, and review.

we found that AI agent performance drops substantially when scoring AI performance holistically rather than algorithmically.

on tasks that take a human expert 90 minutes to 3 hours, a GPT-5 agent (with time horizon of around 2 hours and 17 minutes) succeeds 100% of the time for around one-third of the tasks, fails 100% of the time for around one-third of the tasks, and sometimes succeeds and sometimes fails on the remaining third of tasks.

The task-completion time horizon is the task duration (measured by human expert completion time) at which an AI agent is predicted to succeed with a given level of reliability.

Case study: blackmail

Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model's behavior.

We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts.

Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior.

Emotion vector activations across post-training

Interestingly, they do not by themselves persistently track the emotional state of any particular entity, including the AI A

These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion's relevance to processing the present context and predicting upcoming text.

Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior.

the LLM can effectively track functional emotional states of entities in its context window, including the Assistant, by attending to these representations across token positions, a capability of transformer architectures not shared by biological recurrent neural networks

Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy.

Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model's behavior.

the Assistant (named Claude, in Anthropic's models) can be thought of as a character that the LLM is writing about, almost like an author writing about someone in a novel.

We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts.