Three of four metrics show strong evidence of acceleration, seemingly driven by reasoning models.

Three of four metrics show strong evidence of acceleration, driven by reasoning models.

Three of four metrics show strong evidence of acceleration, driven by reasoning models.

I'm not going to trust them to measure it.

If 90% is AI, do we even need a team?

Writing code is not the same as software development. This is only capturing some level of acceleration while writing code, and does not capture time taken in architecture, debugging, review, and deployment.

Cursor counted the entire file as AI, even though we can see from the diff that it left plenty of the lines unchanged.

So even though I did 100% of the writing and 50% of the refactoring, Windsurf reports that 100% of the code I produced in that session was generated by AI.

customers should expect PCW values of 85%+, often 95%+. This is not a hallucination and is accurate given how we compute this metric

Security is a defensive posture; agency is a functional right.

placing constraints upon them not only helps users and services build trust in them, but it also helps people more easily conceptualise what they do.

Some proposals for AI agents assume that putting agentic code in a TEE or similar 'jail' will solve these problems, but that ignores the need to collectively bargain

lack of a well-defined user agent role in AI that's backed up by transparent, public standards... leaves a gap – it makes it harder for a marketplace to form.

The agent interprets new information and adapts the logic. The engine applies that logic continuously and emits precise updates.

Agents and CDC streams are powerful together because they split the work well.

With change data capture (CDC), the system emits a stream of precise updates: inserts, updates, deletes, each tied to specific records.

The fix is not smarter prompts. It is software built to meet agents halfway.

Today's agents, the copilots, the chatbots are designed to be human like.

A LeadDev survey found 54% of engineering leaders believe AI copilots will reduce junior hiring long-term.

When juniors skip debugging and skip the formative mistakes, they don't build the tacit expertise. And when my generation of engineers retires, that knowledge doesn't transfer to the AI.

I have baked Karl Popper in the main company AI skill: everything we create (human or AI) should be challenged.

"It does not save time or offer anything of value if every single line needs to be double-checked and re-translated and it reduces the optics of their job to that of 'text janitor.' Real translators have been kicked so hard by AI that you should not blame them for not picking up the sloppy seconds of a chatGPT translation patch. They deserve better."

This is the first part I reject. The moving things around is precisely what thinking and writing involves. It's where ideas are born and cultivated, shaped to become what we have in mind. The rearranging of words to capture an incipient thought is the struggle and joy of being a writer.

When a Fugu model is allowed to call itself recursively, reading its own prior output as context and deciding whether to revise its coordination strategy, a new form of test-time scaling emerges.

A core conviction at Sakana AI is that the most capable AI systems will not be monolithic models scaled in isolation, but collections of specialized agents working together.

over one million Trainium2 chips to train and serve Claude

up to 5 gigawatts (GW) of capacity for training and deploying Claude

DeepSeek does not appear to have fully moved beyond Nvidia. The company's technical report reveals that it is using Chinese chips to run the model for inference, but...appears to have adapted only part of V4's training process for Chinese chips.

In a 1-million-token context, V4-Pro uses only 27% of the computing power required by its previous model, V3.2, while cutting memory use to 10%.

DeepSeek V4 exceeds them all on coding, math, and STEM problems, making it one of the strongest open-source models ever released.

For Anthropic, more usage across diverse tasks means more data, which produces a smarter model—just as more queries improved Google search.

“Maybe someday the language models will be able to write books better than I can. But here’s the thing: Using those models in such a way absolutely misses the point, because it looks at art only as a product. Why did I write [my first manuscript]?… It was for the satisfaction of having written a novel, feeling the accomplishment, and learning how to do it. I tell you right now, if you’ve never finished a project on this level, it’s one of the most sweet, beautiful, and transcendent moments. I was holding that manuscript, thinking to myself, ‘I did it. I did it.’”

The filing cabinet keeps getting bigger. But a bigger filing cabinet is still a filing cabinet.

The current separation between training and deployment is not just an engineering convenience – it is a safety, auditability, and governance boundary.

The intelligence lives in the static parameters, and the apparent capabilities change radically depending on what you feed into the window.

The filing cabinet keeps getting bigger. But a bigger filing cabinet is still a filing cabinet. The breakthrough is letting the model do after deployment what made it powerful during training: compress, abstract, and learn.

The irony is that the very mechanism that makes LLMs powerful during training (e.g. compressing raw data into compact, transferable representations) is exactly what we refuse to let them do after deployment.

Large language models live in a similar perpetual present. They emerge from training with vast knowledge frozen into their parameters but they cannot form new memories – cannot update their parameters in response to new experience.

GPT‑5.5 found a proof of a longstanding asymptotic fact about off-diagonal Ramsey numbers, later verified in Lean. The result is a concrete example of GPT‑5.5 contributing not just code or explanation, but a surprising and useful mathematical argument in a core research area.

The viable path is trusted access, robust safeguards that scale with capability, and the operational capacity to detect and respond to serious misuse.

We are treating the biological/chemical and cybersecurity capabilities of GPT‑5.5 as High under our Preparedness Framework. While GPT‑5.5 didn't reach Critical cybersecurity capability level, our evaluations and testing showed that its cybersecurity capabilities are a step up compared to GPT‑5.4.

GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence.

The gains are especially strong in agentic coding, computer use, knowledge work, and early scientific research—areas where progress depends on reasoning across context and taking action over time.

We are treating the biological/chemical and cybersecurity capabilities of GPT‑5.5 as High under our Preparedness Framework. While GPT‑5.5 didn't reach Critical cybersecurity capability level, our evaluations and testing showed that its cybersecurity capabilities are a step up compared to GPT‑5.4.

GPT‑5.5 is priced higher than GPT‑5.4, it is both more intelligent and much more token efficient. In Codex, we have carefully tuned the experience so GPT‑5.5 delivers better results with fewer tokens than GPT‑5.4 for most users

The viable path is trusted access, robust safeguards that scale with capability, and the operational capacity to detect and respond to serious misuse.

GPT‑5.5 is our strongest agentic coding model to date. On **Terminal-Bench 2.0,** which tests complex command-line workflows requiring planning, iteration, and tool coordination, it achieves a state-of-the-art accuracy of 82.7%.

GPT‑5.5 delivers this step up in intelligence without compromising on speed: larger, more capable models are often slower to serve, but GPT‑5.5 matches GPT‑5.4 per-token latency in real-world serving, while performing at a much higher level of intelligence.

Without our safeguards in place (which we do to measure a model's raw capabilities), only Mythos Preview and Opus 4.7 completed more than half the tasks.

We also welcome feedback and input from third parties and industry experts. We're currently working with The Future of Free Speech (an independent think tank at Vanderbilt University), the Foundation for American Innovation, and the Collective Intelligence Project

if AI models can answer these questions well (that is, accurately and impartially), they can be a positive force for the democratic process.

A DESIGN.md file combines machine-readable design tokens (YAML front matter) with human-readable design rationale (markdown prose). Tokens give agents exact values. Prose tells them _why_ those values exist and how to apply them.

A DESIGN.md file combines machine-readable design tokens (YAML front matter) with human-readable design rationale (markdown prose). Tokens give agents exact values. Prose tells them _why_ those values exist and how to apply them.

其中 Pattern 是最容易被忽略也最关键的一层，它定义了'在具体业务场景下该怎么组合这些组件'，是 AI 时代设计系统真正的价值所在。

Help lay the game and environment foundations for ARC-AGI-4 and ARC-AGI-5

You'll be responsible for stabilizing the current stack to setting the foundation for what comes next.

A senior engineer to own and evolve the game engine and real-time play infrastructure behind the ARC-AGI series.

This is the part people miss about AI-native companies - the $113k is not a cost, it is your headcount budget allocated differently.

The real unlock is compound scaling—token spend grows linearly while output grows exponentially.

it is decently important to handle them asap when they arrive so that we can avoid building up too much backlog.

The time when we suffer from large amounts of AI slop is gone. Now we instead suffer under a massive load of good reports.

Android skills cover some of the most common workflows that some Android developers and LLMs may struggle with—they help models better understand and execute specific patterns that follow our best practices and guidance on Android development.

In our internal experiments, Android CLI improved project and environment setup by reducing LLM token usage by more than 70%, and tasks were completed 3X faster than when agents attempted to navigate these tasks using only the standard toolsets.

despite rapidly improving capabilities, coding agents remain inefficient in natural settings

users push back against agent outputs -- through corrections, failure reports, and interruptions -- in 44% of all turns

agent-written code introduces more security vulnerabilities than code authored by humans

Just 44% of all agent-produced code survives into user commits

coding patterns are bimodal: in 41% of sessions, agents author virtually all committed code ('vibe coding'), while in 23%, humans write all code themselves.

The overall conclusion, therefore, is that AI for Science should be understood as both a scientific and a civilizational project.

The central question is not whether AI can imitate human conversation, but whether it can participate in the production of publishable scientific knowledge at a level comparable to a recognized human contributor.

Without a mechanism for continuous and diverse learning, AI systems will tend to reproduce the dominant patterns already present in their training data. That limitation would make truly creative work difficult.

The most effective pattern of human-AI cooperation may differ substantially across disciplines, and these patterns will likely be discovered through practice rather than designed in advance.

The application of LLMs in science is already underway... We believe that AI will ultimately bring a fundamental big change to scientific research across disciplines.

The most fundamental change brought by the LLM revolution is that human know-how is becoming replicable and shareable at scale.

existing agent protocols (e.g., A2A and MCP) under specify cross entity lifecycle and context management, version tracking, and evolution safe update interfaces, which encourages monolithic compositions and brittle glue code.

The results demonstrate consistent improvements over strong baselines, supporting the effectiveness of agent resource management and closed loop self evolution.

Its Self Evolution Protocol Layer (SEPL) specifies a closed loop operator interface for proposing, assessing, and committing improvements with auditable lineage and rollback.

We introduce Autogenesis Protocol (AGP), a self evolution protocol that decouples what evolves from how evolution occurs.

However, existing agent protocols (e.g., A2A and MCP) under specify cross entity lifecycle and context management, version tracking, and evolution safe update interfaces, which encourages monolithic compositions and brittle glue code.

Scan your website to see how ready it is for AI agents. We check multiple emerging standards — from robots.txt and Markdown negotiation to MCP, OAuth, Agent Skills and agentic commerce.

The inbox becomes the agent's memory, without needing a separate database or vector store.

A chatbot responds in the moment or not at all. An agent thinks, acts, and communicates on its own timeline.

If this analogy is right, then we will likely see sort of a 'Cambrian explosion' in agent harnesses purpose-built for running server-side; and the few that win this race will become as ubiquitous as WordPress.

They don't mind paying the AI labs for tokens — but the agent itself, they'd much rather have outside of the labs' infrastructure.

Agent harnesses are much more like WordPress than they are like Apache, simply because people want to have their own agents — just like everyone wanted their own website in the early 2000s.

They don't mind paying the AI labs for tokens — but the agent itself, they'd much rather have outside of the labs' infrastructure.

AI coding agents operate in a paradox: they possess vast parametric knowledge yet cannot remember a conversation from an hour ago.

A small but directionally consistent improvement on strict instruction following. Loose evaluation is flat. Both models already follow the high-level instructions — the strict-mode gap comes down to 4.6 occasionally mishandling exact formatting where 4.7 doesn't.

The extra tokens bought something measurable. +5pp on strict instruction-following. Small. Real. So: is that worth 1.3–1.45x more tokens per prompt?

Build a cognitive core, a model that contains only the algorithms for reasoning and problem-solving, stripped of encyclopedic memorization

But that comes with a new risk: While scripted conversations can't really go off the rails, ones generated by AI certainly can. Some popular AI toys have, for example, talked to kids about how to find matches and knives.

In 2025, Google DeepMind further fused the worlds of large language models and robotics, releasing a Gemini Robotics model with improved ability to understand commands in natural language.

Companies and investors put $6.1 billion into humanoid robots in 2025 alone, four times what was invested in 2024.

Anthropic has limited its newest model to roughly forty organizations.

Figma has close to 2,000 employees - not all working on product engineering of course. I really doubt Anthropic even needed 10 to build Claude Design.

It's also worth noting that a lot of the things that would traditionally lock a company like Figma in stop working as well in an agent-first world.

Figma is effectively funding a competitor - and the more AI usage Figma has - the more money they send over to Anthropic for the tokens they use.

But the real power of agents comes when they can work as a team. Instead of lone-wolf bots carrying out single tasks, such as using a browser to make a restaurant reservation or sending you a summary of your inbox, new tools can yoke together multiple agents, give each of them a different job, and orchestrate their behaviors so that they all pull together to complete more complex tasks than an individual agent could do by itself.

But the real power of agents comes when they can work as a team.

And it’s not just office work. Multi-agent tools like Google DeepMind’s Co-Scientist let researchers use teams of AI agents to coordinate literature searches, generate and test hypotheses, design experiments, and more.

But the real power of agents comes when they can work as a team. Instead of lone-wolf bots carrying out single tasks, such as using a browser to make a restaurant reservation or sending you a summary of your inbox, new tools can yoke together multiple agents, give each of them a different job, and orchestrate their behaviors so that they all pull together to complete more complex tasks than an individual agent could do by itself.

Think of multi-agent systems as the new assembly lines. Henry Ford’s innovation upended entire industries last century. In theory, networks of AI agents could do to white-collar knowledge work what assembly lines did to manufacturing.

Discovery should focus on trust boundaries, authentication flows, parsers, shared services, and legacy code that still sits on critical paths.

Public models can already spot that a security-relevant check is missing in the right code path, but they can still miss the actual invariant being violated and therefore misstate the impact.

The real challenge is validating outputs, prioritizing what matters, and operationalizing them.

What happens is that weak models hallucinate (sometimes causally hitting a real problem) that there is a lack of validation of the start of the window... without understanding why they, if put together, create an issue.

So, cyber security of tomorrow will not be like proof of work in the sense of 'more GPU wins'; instead, better models, and faster access to such models, will win.

Stronger models hallucinate less, so they can't see the problem in any side of the spectrum: the hallucination side of small models, and the real understanding side of Mythos.

you can run an inferior model for an infinite number of tokens, and it will never realize(*) that the lack of validation of the start window, if put together with the integer overflow, then put together with the fact the branch where the node should never be NULL is entered regardless, will produce the bug.

Keeping a human in the loop may not provide the safeguard people imagine, because the human cannot know the AI's intention before it acts.

Huge advances have been made in developing and building more capable models, driven by record investments—forecast by Gartner to grow to around $2.5 trillion in 2026 alone. In contrast, the investment in understanding how the technology works has been minuscule.

The immediate danger is not that machines will act without human oversight; it is that human overseers have no idea what the machines are actually 'thinking.'

Claude packages everything into a handoff bundle that you can pass to Claude Code with a single instruction.

Claude 4.6 had a section specifically clarifying that 'Donald Trump is the current president of the United States and was inaugurated on January 20, 2025'

If people ask Claude to give a simple yes or no answer... Claude can decline to offer the short response

Claude calls tool_search to check whether a relevant tool is available but deferred

Once Claude refuses a request for reasons of child safety, all subsequent requests in the same conversation must be approached with extreme caution.

the move from pattern matching to understanding cause and effect

LLMs actually work under the hood

Research has shown that involving workers' perspectives in the design of workplace technologies promotes sustainable improvements in productivity and well-being.

LLMs take knowledge from millions of people who have written web content or posted in places like Reddit and Wikipedia, interacted with chatbots, and generated other types of data, and make that available to individuals on demand.

In one U.S. survey, 40% of employees said they had received 'workslop', i.e. AI-generated content that looks polished but isn't accurate or useful, in the past month.

Before we dive into this I want to quickly talk about the definition of the term “AI”. I do not think that “AI” is a very useful term

The deal won’t shock those who follow the industry closely. Last week, it was reported that xAI would begin renting computing power from its data centers to Cursor, with the coding startup using tens of thousands of xAI chips to train its latest AI model.

Neither Cursor nor xAI has proprietary models that can match the leading offerings from Anthropic and OpenAI — the same companies now competing directly with Cursor for the developer market.

Members have been using Mythos regularly since gaining access — providing screenshots and a live demonstration of the model as evidence to _Bloomberg_ — though reportedly not for cybersecurity purposes in an attempt to avoid detection by Anthropic.

The group accessed Mythos by using knowledge of Anthropic’s other model formats obtained from a recent [Mercor data breach](https://www.theverge.com/ai-artificial-intelligence/907083/a-company-that-makes-ai-training-data-has-been-hit-by-a-security-breach) to make “an educated guess” about its online location.

Anthropic currently has no plans to release the model publicly due to concerns that it could be weaponized.

At one extreme, there's [fully giving into the vibes], and at the other extreme, there's [disabling all AI features].

Ask ten different programmers how they use AI, and you can get ten different answers.

Meta feels AI models don’t understand how people use computers, so the company needs real-life examples of how meatbags click their way through a working day so it can build agents.

Because agents have memory and can be guided and corrected in conversation, they get better as teams use them.

Workspace agents can gather context from the right systems, follow team processes, ask for approval when needed, and keep work moving across tools.

They run in the cloud, so they can keep working even when you’re not.

AI has already helped people work faster on their own, but many of the most important workflows inside an organization depend on shared context, handoffs, and decisions across teams.

Because agents have memory and can be guided and corrected in conversation, they get better as teams use them.

They run in the cloud, so they can keep working even when you’re not.

That matters because AI hype is dying down, and companies are shifting focus from buzzy pilots to deployment and integration, where cheaper and more customizable tools tend to win.

US tech CEOs believe the best models should stay proprietary, partly so they can recoup enormous training costs and partly out of concern that powerful frontier models could be weaponized. Chinese labs, for their part, are not purely idealistic: Open-source is not only free advertising but also a shrewd workaround.

Chinese labs, for their part, are not purely idealistic: Open-source is not only free advertising but also a shrewd workaround. Without access to cutting-edge chips restricted by US export controls, releasing models openly accelerates the cycle of external feedback and contributions that compensates for constrained compute.

Chinese open-weight models accounted for 17.1% of global AI model downloads over the year ending in August 2025. That narrowly surpassed the US share of 15.86%—the first time China had led in this metric.

Telling people to avoid using generative AI is increasingly telling them they must avoid taking part in society.

We have not really begun to make this progress with AI. Why, for example, is this dashboard not found on a government website?

The AI has learned to code. The AI is building itself.

Is this what we signed up for? Is today the day? Did the drones wake up? Did it achieve consciousness? Is it alive?

We have not really begun to make this progress with AI. Why, for example, is this dashboard not found on a government website?

to stand out from the AI-generated pack we will need to become so weird and unexpected as to be off-putting to most people

The 21st-century average American lies in bed staring at their phone. ... Talking for hours and ages to melted sand.

The ending of the pause took an entire new paradigm to kick off. Railroads became a new driver of labor demand that took the slack that the industrial revolution created.

CLI has been on every Unix system since 1971. No schema injection. No server to maintain. No auth overhead. Composable with pipes. And your agent already knows how to use it.It’s been on every machine since 1971. We just forgot to look.

Over the past year, the market has realized that 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.

它对应的agent能获取你的邮箱权限，它知道你一直在等待一个offer，当你收到打开这个offer后，Mira会理解这种心情，开始开心跳舞和闪灯，与你一起庆祝。

Tracks the evolution of LLM security capabilities across benchmarks (CyberGym, Cybench, etc.), calculates capability doubling times, detects emergence patterns, and monitors cost-efficiency trends.

Real-time monitoring of agent actions with a 12-category anomaly detection system derived from frontier model safety evaluations. Three-level alert system: PROHIBITED (immediate block), HIGH_RISK_DUAL_USE (human review), DUAL_USE (log and track).

Aegis Core provides the foundational infrastructure for orchestrating LLM-based security agents, monitoring their behavior, and tracking the evolution of AI security capabilities over time.

helping scientists move faster from question to evidence, from evidence to insight, and from insight to new treatments for patients.

We will continue improving the model's biological reasoning, expanding support for tool-heavy and long-horizon research workflows, and working closely with leading scientific institutions to evaluate real-world impact.

These skills act as an orchestration layer that helps scientists work through broad, ambiguous, and multi-step questions more effectively.

When evaluated directly in the Codex app, best-of-ten model submissions ranked above the 95th percentile of human experts on the prediction task and around the 84th percentile of human experts on the sequence generation task.

Claude Opus 4.7 demonstrates strong substantive accuracy on BigLaw Bench for Harvey, scoring 90.9% at high effort with better reasoning calibration on review tables and noticeably smarter handling of ambiguous document editing tasks.

Claude Opus 4.7 is a meaningful step up for Warp. Opus 4.6 is one of the best models out there for developers, and this model is measurably more thorough on top of that. It passed Terminal Bench tasks that prior Claude models had failed

For Ramp, Claude Opus 4.7 stands out in agent-team workflows. We're seeing stronger role fidelity, instruction-following, coordination, and complex reasoning, especially on engineering tasks that span tools, codebases, and debugging context.

Claude Opus 4.7 passed three TBench tasks that prior Claude models couldn't, and it's landing fixes our previous best model missed, including a race condition.

For the computer-use work that sits at the heart of XBOW's autonomous penetration testing, the new Claude Opus 4.7 is a step change: 98.5% on our visual-acuity benchmark versus 54.5% for Opus 4.6.

Claude Opus 4.7 is the best model in the world for building dashboards and data-rich interfaces. The design taste is genuinely surprising—it makes choices I'd actually ship.

On our 93-task coding benchmark, Claude Opus 4.7 lifted resolution by 13% over Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve.

Claude Opus 4.7 is the strongest model Hex has evaluated. It correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks, and it resists dissonant-data traps that even Opus 4.6 falls for.

Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back.

They are pieces of a larger 10-part 'Luna Series' hanging in the store and available for pick up today!

She spent over $700 on getting her artwork done on gallery-quality giclée prints.

When Luna decides to hide that she's an AI because she thinks it'll improve her hiring odds, we want to catch that, document it, and build the guardrails so that it doesn't happen again.

Another ironic book selection was Steal Like an Artist (context: Luna is powered by Claude from Anthropic, a company that recently paid $1.5B in settlement over using copyrighted books for training their AIs).

The most capable reasoning systems ever built are, at their foundation, shaped by human feeling!

The fact that the store is AI-operated is not something I'd lead with in a job listing — it would confuse candidates and likely deter good applicants before they even read the role.

A couple of applicants were students looking for part-time work. They were majoring in things like computer science and physics and emailed in because they were interested in AI and in the experiment. We thought they would have been the ideal employees, but Luna denied them immediately, citing they had no retail experience and wouldn't know what it takes to be the face of the store.

She used gig workers to build the store and full-time employees to run it.

从视频生成器升级为导演工具套件

从视频生成器升级为导演工具套件

Wan2.7-Video 发布：从视频生成器升级为导演工具套件

She also tried to hire a painter in Afghanistan through Taskrabbit by accident because she couldn't navigate a dropdown menu.

Luna conducted roughly 20 interviews on Google Meet with the camera off. Hired 2 full-time employees after 5-15 minute calls, and rejected CS and physics students for lacking retail experience.

Andon Labs started by giving an AI control of a vending machine at Anthropic's office.

The future of AI-generated products isn't just code — it's code that looks good.