The only fundamentally scarce thing is the synchronous human attention of my team

The obsession with fast build times: why one minute became the upper bound for the inner loop

humans became the bottleneck, and how Ryan's team shifted from reviewing code directly to building systems, observability, and context that let agents review, fix, and merge work autonomously

building and shipping an internal beta product with zero manually written code

We shed light on OpenAI's first Dark Factory for the first time.

Some privacy related extensions may cause issues on x.com. Please disable them and try again.

Please enable JavaScript or switch to a supported browser to continue using x.com.

Some privacy related extensions may cause issues on x.com.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

ChatGPT has 900 million weekly users, which means employees already know how to work with it. For enterprises, that reduces rollout friction and accelerates the point where every employee can delegate tedious tasks.

The shift started with agentic tools like Codex, which has grown more than 5X since the start of the year. This includes customers like GitHub, Nextdoor, Notion, and Wonderful that are building multi-agent systems that can execute engineering work end-to-end.

Codex just hit 3 million weekly active users, our APIs process more than 15 billion tokens per minute, and GPT‑5.4 is driving record engagement across agentic workflows.

Enterprise now makes up more than 40% of our revenue, and is on track to reach parity with consumer by the end of 2026.

The goal is to build the trust, verification, and accountability needed to make these tools available to the many defenders whose work keeps people, institutions, and critical systems safe.

We have also provided access to GPT-5.4-Cyber to the U.S. Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute (UK AISI) so that they can conduct evaluations focused on the model's cyber capabilities and safeguards.

Not every organization has the benefit of a 24x7 security team who is able to respond to incidents when they are disclosed on a Friday night.

Cybersecurity is a team sport, and the systems people rely on are protected by organizations of many kinds, from major enterprises and security vendors to researchers, maintainers, public institutions, nonprofits, and smaller teams with limited security resources.

或许需要某种「第三方评测、审计机构」来评估 Skills 的数据使用方式、检测潜在安全风险等等。

未来的评估体系，必须同时考虑：成功率、成本、延迟。这有点类似于对于云计算的考核标准，而不是传统软件。

信任从「平台」转移到了个体。其中一部分原因在于，Skills 的工作机制不透明，像个「黑箱」，用户只知输入输出，不知其分析指令、调用工具、做出决策的过程。

Skills 的传播没有像 App 那样可以靠搜索、靠排名。用户更加追求结果导向，而不是过程导向。

一个本该由「应用商店」承载的分发体系，现在却被内容平台接管了。

对视频生成来说，这种文字密集、变化快、带闪烁、又几乎没有自然动态的场景，本来就是最难的一类。

用户输入不再只是触发一次性行为，而会逐渐安装、调用、组合并保留可复用的 neural routines。

未来的 CNC 也许不是一团越来越大的连续表征，而会更像一套可路由、可组合、局部更容易检查的机器底座。

Neural Computer 真正成形，大概还要三年。

模型能不能承担一部分原本属于机器运行本身的职责。

I generally wake up with a sense of optimism that slowly withers and dies in me throughout the day, to the point that by the time I go to bed each night, I’ve probably given up on myself as well as most of humanit

It feels like the language itself is doing exactly what it describes: creating a divide between the reader and the ability to tell stories that connect us with the natural world instead of furthering the divide.

I chose to focus on happiness

Every day, you're hit with countless atrocities—some massive, some small, and you’re stuck there as a witness, a subject of it all, yet feeling way too small to actually stop these acts of violence or check the geopolitical forces in play.

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.

The license looks MIT at first glance but it is not MIT. Non commercial use is free with no restrictions. Commercial use requires prior written authorization from MiniMax.

MiniMax claims it has reduced live production incident recovery time to under three minutes on multiple occasions using M2.7.

The model kept finding better approaches the longer it ran, which connects directly to the long horizon behavior that makes agentic models actually useful in production.

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.

focusing on the ~1.5K mainline open models from the likes of Alibaba's Qwen, DeepSeek, Meta's Llama

We document a clear trend where Chinese models overtook their counterparts built in the U.S.

that are the foundation of an ecosystem crucial to researchers, entrepreneurs, and policy advisors.

We study a mix of Hugging Face downloads and model derivatives, inference market share, performance metrics and more to make a comprehensive picture of the ecosystem.

We present a comprehensive adoption snapshot of the leading open language models and who is building them, focusing on the ~1.5K mainline open models

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

scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.

Muse Spark compresses its reasoning to solve problems using significantly fewer tokens. After compressing, the model again extends its solutions to achieve stronger performance.

The model frequently identified scenarios as 'alignment traps' and reasoned that it should behave honestly because it was being evaluated.

we rebuilt our pretraining stack with improvements to model architecture, optimization, and data curation.

Contemplating mode provides significant capability improvements in challenging tasks, achieving 58% in Humanity's Last Exam and 38% in FrontierScience Research.

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.

allow it to have beginnings too.

A Black-cene that acknowledges Black people's presence if humans survive and acknowledges that eventually the ending that has been rippling through Black communities will eventually reach everyone else and swallow us all up.

(a beginning)

Ending with a beginning 1 2026-03-05T06:50:43-08:00 Evelyn Logan 44c50294b54d30778ea572edca97fd5a427395f4 48233 17 plain 2026-04-14T10:43:42-07:00 Evelyn Logan 44c50294b54d30778ea572edca97fd5a427395f4 Of Cotton, Clay, and Water: A Black retelling of the Anthropocene narrative

Open-source development is starting to redistribute participation, with contributions from the rest of the world now outpacing Europe and approaching the United States on GitHub.

73% of experts expect a positive impact on how people do their jobs, compared with just 23% of the public, a 50-point gap.

The United States leads in AI investment, but its ability to attract global talent is declining.

Responsible AI is not keeping pace with AI capability, with safety benchmarks lagging and incidents rising sharply.

AI models can win a gold medal at the International Mathematical Olympiad but cannot reliably tell time—an example of what researchers call the jagged frontier of AI.

The U.S.-China AI model performance gap has effectively closed.

AI capability is not plateauing. It is accelerating and reaching more people than ever.

Website: add animated workflow demos

Add GCP WebVoyager benchmark runner and worktree tooling

Don't destroy cloud sessions on transient CDP failures

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

Simplify benchmarks to webVoyager-only with Pi SDK runner

Add cloud browser provider system (Kernel + Browserbase)

The AI toolkit for building and maintaining browser automations

Lightweight Agent Detection & Response (ADR) layer for AI agents — guards commands, files, and web requests.

Sage sends URLs and package hashes to Gen Digital reputation APIs. File content, commands, and source code stay local.

Sage intercepts tool calls (Bash commands, URL fetches, file writes) via hook systems in Claude Code, Cursor / VS Code, OpenClaw, and OpenCode, and checks them against:

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.

They have pride in what they do... They won't let some AI bot take over, and they will always find and show the flaws in that tool compared to them.

Workers lose the equivalent of 51 working days per year to technology friction — nearly two full months — up 42% from 2025.

Only 9% of workers trust AI for complex, business-critical decisions, compared to 61% of executives — a 52-point trust chasm.

White-collar workers are quietly rebelling against AI as 80% outright refuse adoption mandates

Academic publishers, documentary archives, game studios, and companies sitting on years of enterprise data have all been courted for the seeds of intelligence needed to train the next generation of models.

Mercor, which provides data to AI labs for training, became one of the fastest-growing companies in history before losing four terabytes of data to hackers last week.

While some experts have speculated that general models will win out in performance over specialized models—that scale and compute will beat curation—the success of these companies shows that the market is making a more nuanced bet.

A small model trained on fewer than 2,000 examples from real lawyers, bankers, and consultants recently beat all but the best frontier models on corporate legal work, at a fraction of the price.

Reddit, Shutterstock, and News Corp are making hundreds of millions a year licensing their high-quality data to companies training AI, and those contracts are growing about 20 percent annually, according to their quarterly filings.

We calculate the aggregate amount of compute (in H100-equivalents) held by Amazon, Google, Meta, Microsoft, and Oracle, as a share of the global total each quarter.

Five hyperscalers now own over two-thirds of global AI compute

Our Chip Ownership data does not capture all global chip ownership, and has weaker coverage prior to 2023.

The H100-equivalent unit uses a chip's highest 8-bit operation/second specifications to convert between chips. The actual utility of a particular chip depend on workload assumptions, so H100e does not perfectly reflect real-world performance differences across chip types.

Many AI labs (including OpenAI and Anthropic) largely depend on these hyperscalers for access to R&D and inference compute.

Amazon, Google, Meta, Microsoft, and Oracle collectively hold an estimated 71% of the world's cumulative AI compute as of Q4 2025, measured in H100-equivalents of computing power.

Four researchers and software engineers estimated that a skilled human engineer would take 2 to 17 weeks to reimplement gotree, as AI successfully did in this work.

We found weak evidence that Opus 4.0 and 4.1 had partially memorized cal, but no evidence Opus 4.6 had memorized it, despite performing best of all models considered.

It is not common for real software to be developed the way MirrorCode tasks are structured — against a precise, programmatically checkable specification.

Older models were more prone to submitting prematurely, even when test cases weren't passing.

We see continued gains from inference scaling on larger projects, suggesting they may be solvable given enough tokens.

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

We did not collect detailed examples of specific tasks, but these results provide an early, nationally representative snapshot of how AI is reshaping work at the task level.

While most people still use AI mainly for personal tasks, about half of employed users use it at least as much for work. This share is even higher among those with paid tools, particularly when provided by employers.

Microsoft Copilot, which leads paid AI usage among both work-oriented and personal-oriented users, illustrates this dynamic: its prevalence likely reflects bundling with Microsoft 365, a product widely deployed in workplaces through enterprise licensing.

Among employed AI users, 38% of free-tier users reported using AI at least as much for work as for personal tasks. The share rises to 58% among self-paying subscribers and 76% among users with employer-provided subscriptions.

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

Half of employed Americans who used AI in the past week reported using AI tools at least as much for work as for personal tasks.

By default, keys generated in Google AI Studio are restricted to just the Gemini API, no other services are enabled.

In many cases, we can automatically detect when a key is visible on the public web and shut down those keys automatically for security reasons

We just started the prepaid billing rollout which means you have to pay ahead of time to use the Gemini API, this is rolled out to all new US billing accounts as of yesterday

We are moving to disable the usage of unrestricted API keys in the Gemini API, should have more updates there soon.

We had a budget alert (€80) and a cost anomaly alert, both of which triggered with a delay of a few hours

We experienced a sudden and extreme spike in Gemini API usage. The traffic was not correlated with our actual users and appeared to be automated.

Google spent over a decade telling developers that Google API keys (like those used in Maps, Firebase, etc.) are not secrets. But that's no longer true.

Gemini Robotics-ER 1.6 can use points as intermediate steps to reason about more complex tasks. For example, it can use points to count items in an image, or to identify salient points on an image to help the model perform mathematical operations to improve its metric estimations.

Capabilities like instrument reading and more reliable task reasoning will enable Spot to see, understand, and react to real-world challenges completely autonomously.

On these tasks, our Gemini Robotics-ER models improve over baseline Gemini 3.0 Flash performance (+6% in text, +10% in video) in perceiving injury risks accurately.

Safety is integrated into every level of our embodied reasoning models. Gemini Robotics-ER 1.6 is our safest robotics model to date, demonstrating superior compliance with Gemini safety policies on adversarial spatial reasoning tasks compared to all previous generations.

Gemini Robotics-ER 1.6 achieves its highly accurate instrument readings by using agentic vision, which combines visual reasoning with code execution. The model takes intermediate steps: first zooming into an image to get a better read of small details in a gauge, then using pointing and code execution to estimate proportions and intervals and get an accurate reading.

We are also unlocking a new capability: instrument reading, enabling robots to read complex gauges and sight glasses — a use case we discovered through close collaboration with our partner, Boston Dynamics.

Gemini Robotics-ER 1.6 shows significant improvement over both Gemini Robotics-ER 1.5 and Gemini 3.0 Flash, specifically enhancing spatial and physical reasoning capabilities such as pointing, counting, and success detection.

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 result is a Skill that acts as a cheat sheet for the MCP, not a replacement for it. The MCP still handles the actual connection and tool execution. The Skill just makes sure the LLM doesn't waste tokens stumbling through the same pitfalls I already solved.

Shower thought: Maybe the terminology is the problem. Skills should just be called `LLM_MANUAL.md`, and MCPs should be called `Connectors`.

Most skills require you to install a dedicated CLI. But what if you aren't in a local terminal? ChatGPT can't run CLIs. Neither can Perplexity or the standard web version of Claude.

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_.

CSS Studio uses Motion to generate real CSS springs with the perfect feel.

Add elements, rename tags, reorder with drag-and-drop, duplicate and delete. Double-click text to edit inline.

CSS Studio detects the CSS variables available on an element. Edit a variable and watch it propagate across the site.

Scrub through CSS `@keyframes` animations on a visual timeline.

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

Your AI agent writes every change into source code.

Design by hand. Code by agent.

Each run creates a new session alongside your other sessions, where you can see what Claude did, review changes, and create a pull request.

The prompt is the most important part: the routine runs autonomously, so the prompt must be self-contained and explicit about what to do and what success looks like.

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

A single routine can combine triggers. For example, a PR review routine can run nightly, trigger from a deploy script, and also react to every new PR.

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

A routine is a saved Claude Code configuration: a prompt, one or more repositories, and a set of connectors, packaged once and run automatically.

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

While our production codebase has significantly diverged, including major rewrites of core systems like authentication and data handling, we want to ensure there is still a truly open version available.

Each platform surfaces different vulnerabilities, making it difficult to establish a single, reliable source of truth for what is actually secure.

We hope that one day we can return to open source as the security landscape evolves. But for now, we have to put our customers first.

The risk landscape is accelerating quickly. Advanced AI models are now capable of identifying and exploiting vulnerabilities at unprecedented speed.

AI uncovered a 27-year-old vulnerability in the BSD kernel, one of the most widely used and security-focused open source projects, and generated working exploits in a matter of hours.

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.

The standard autoresearch loop (brainstorm from code, run experiments, check metrics) works when the optimization surface is visible in the source. The Liquid results prove that. But for problems where the codebase doesn't contain enough information to generate good hypotheses, giving the agent access to papers and competing implementations changes what it tries.

The variance is also worth noting: baseline+FA TG has ±19 t/s of noise, while optimized+FA has ±0.59 t/s on x86. The fusions eliminate intermediate writes that pollute the cache, making the hot paths more predictable.

Without experience with compiler behavior, the agent couldn't have predicted which 'optimizations' the compiler would already handle.

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.

Coding agents working from code alone generate shallow hypotheses. Adding a research phase — arxiv papers, competing forks, other backends — produced 5 kernel fusions that made llama.cpp CPU inference 15% faster.

The macOS app is available to Gemini users ages 13+

We're building the foundation for a truly personal, proactive and powerful desktop assistant, with more news to share in the coming months.

Creatives can also quickly generate images with Nano Banana or videos with Veo to bring an idea to life without breaking their creative stride.

Switching between windows on your desktop can be clunky and slow. Now, you can bring up Gemini from anywhere on your Mac with a quick shortcut (Option + Space)

You can share your window and ask, 'What are the three biggest takeaways here?' to get an instant summary.

The Gemini app is now available as a native macOS experience, bringing you a faster, more integrated way to get help from AI right on your desktop.

We provide a framework for categorizing the ways in which conflicting incentives might lead LLMs to change the way they interact with users, inspired by literature from linguistics and advertising regulation

Today's large language models (LLMs) are trained to align with user preferences through methods such as reinforcement learning. Yet models are beginning to be deployed not merely to satisfy users, but also to generate revenue for the companies that created them through advertisements

Behaviors also vary strongly with levels of reasoning and users' inferred socio-economic status

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.

As slop takes over the Internet, labs may struggle to obtain high-quality corpuses for training models.

Humans can be motivated by consequences and provide social redress in a way that LLMs can't.

When models go wrong, we will want to know why. What led the drone to abandon its intended target and detonate in a field hospital? Why is the healthcare model less likely to accurately diagnose Black people?

A surprising number of people are now employed as model trainers, feeding their human expertise to automated systems.

LLMs are weird. You can sometimes get better results by threatening them, telling they're experts, repeating your commands, or lying to them that they'll receive a financial bonus.

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.

A modern data context layer should essentially become a superset of what a semantic layer would traditionally cover. Sure, specific metric definitions can be hard-coded, but a modern context layer should include more to ensure agent autonomy – canonical entities, identity resolution, specific instructions to dissect tribal knowledge, proper governance guidance, and more.

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.

This level of penetration in such a short period of time is remarkable since Fortune 500 enterprises are not known to be early adopters of technology. Historically, many startups had to initially sell to other startups to get early momentum, and it was only after a few years that a startup would be able to land its first enterprise contract.

In many ways, coding represents the ideal use case for AI, both in terms of what the technology can do and how readily the enterprise market will embrace it. Code is data dense, meaning there is a massive amount of high-quality code available online for the models to train on.

The most notable finding here is that the model capabilities are improving _fast._ There are several domains that have shown dramatic improvements in the last 4 months — with accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement.

**Coding, support, and search**represent the lion's share of use cases by far (with coding being an order-of-magnitude outlier even among this set), while the**tech, legal, and healthcare sectors** have been the industries most eager to adopt AI.

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

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.

The industry average time to detect a supply chain breach is 267 days. SolarWinds went undetected for 14 months. XZ Utils took two years to surface. Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication.

Hallucinated packages are the sleeper threat. LLMs regularly invent package names that don't exist. One study found that nearly 20% of AI-recommended packages were fabrications, and 43% of those hallucinated names appeared consistently across queries.

Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.

The average application contains over 1,100 open source components. A bare-bones Next.js project installs 282 packages before you write a single line.

Until then, America is all we have.

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.

Our choice is therefore no longer whether to build such weapons, but only whom to entrust with their responsible use in military affairs.

To accept the existential stakes of that prospect while simultaneously treating the next frontier of superweapon proliferation as an ordinary issue of private property betrays a deep confusion about the problem that this moment presents.

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.

Rather than imagining utopia as a place of excess, growth, or endless progress, a utopia of sufficiency begins from the idea that a better world might require “enough” rather than “more.” This makes utopia in the Anthropocene feel highly personal, as each individual must consider what their minimal needs are, knowing that the cost of consumption beyond those needs is often borne by others suffering

At the same time, this framework asks us to rethink what we mean by a “better life.” Is it defined by comfort, stability, and security, or by something more relational and ecological?

What would a utopia of sufficiency look like to you?What would you be willing to give up or change in order to support it?And how much responsibility should individuals carry in shaping such a world compared to institutions or systems?

vod-live://s3://bucket/file.mp4

Name of the stream published to Origin to inspect its delivery route

/rest-api/v3/mixer/set_body_watermark

/rest-api/v3/mixer/set_stream_watermark

/rest-api/v3/transcoder/set_watermark

We find that optimization becomes more reliable when a small intermediate-state regularizer is added on top of token-level distillation.

The resulting models maintain competitive performance while delivering substantial efficiency improvements, demonstrating that large-scale attention conversion is both feasible and robust.

In practice, deployed model implementations are often flexible (e.g., mixing kernel variants, hybrid attention patterns, MoE blocks, and serving-optimized layouts), which can deviate from the assumptions required by a given conversion recipe.

All training runs are conducted on an Ascend 910B cluster, and our setup follows the growing evidence that large-scale model training on Ascend clusters is feasible in practice.

We introduce GateSWA, an efficient hybrid attention variant that replaces learnable sink mechanisms in hybrid SWA models with an element-wise gate.

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

So the complexity needs to come at many layers

During the 8 minute performance, the audience witnesses a sonic sculpture which is a feeble human attempt to listen without the ears but with the body in relation with AI. The computer’s decisions are audible to human ears, sonifying invisible data.

Rather than relying on corporate, water-sucking data centers located in remote deserts or low-income areas, neural/dream data collected from indigenous subjects will be digitally stored on their sovereign land and under elder control

I will offer a brief analysis from both a techno-optimist and techno-pessimist perspective