Implicit state is the Enemy

The funny part is that none of this made the CLI worse for humans.

Every prompt is a flag in disguise

Designing for agents forced us to build better tools for everyone.

Whether or not this specific bet pays off, the underlying argument that the next meaningful leap in AI capability requires moving beyond language modeling is increasingly hard to dismiss.

The clustering of capital and talent around this problem is itself a signal. The applications that most clearly benefit from world models are those where LLMs have struggled most.

AMI Labs is not building a product for immediate deployment. This is a fundamental research effort, likely measured in years before commercial applications emerge.

LLMs have no grounded understanding of the physical world. They model the statistical distribution of language about reality, not reality itself.

You have to have people that have the ability to rethink the workflow at a scale that AI can execute, versus at a scale that humans can execute.

in 2024, 47% of AI solutions were built internally and 53% were purchased; today, 76% of all AI is purchased rather than developed in-house.

The government has so far favoured a pro-innovation, sector-led approach, prioritising voluntary principles over hard regulation.

We introduce Iterative Reward Calibration, a methodology for designing per-turn rewards using empirical discriminative analysis of rollout data

the trained 4B model exceeding GPT-4.1 (49.4 percent) and GPT-4o (42.8 percent) despite being 50 times smaller

naively designed dense per-turn rewards degrade performance by up to 14 percentage points due to misalignment between reward discriminativeness and advantage direction

computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments

intermediate actions that appear locally acceptable but collectively lead to unauthorized actions

model alignment alone does not reliably guarantee the safety of autonomous agents

verifiers and observer models inside the action-memory loop reduce silent failure and information leakage while remaining vulnerable to misspecification.

role-differentiated proposer/executor/checker/adversary systems may reduce correlated error under asymmetric information and verification burden.

We introduce a minimal hierarchical partially observed control model with latent dynamics, structured episodic memory, observer-belief state, option-level actions, and delayed verifier signals.

squirrel ecology offers a sharp comparative case because arboreal locomotion, scatter-hoarding, and audience-sensitive caching couple all three demands in one organism.

Existing research often studies these demands separately: robotics emphasizes control, retrieval systems emphasize memory, and alignment or assurance work emphasizes checking and oversight.

Agentic AI is increasingly judged not by fluent output alone but by whether it can act, remember, and verify under partial observability, delay, and strategic observation.

we use the distance preference characterized by these centers to score keys according to their positions, and also leverage Q/K norms as an additional signal for importance estimation

TriAttention enables OpenClaw deployment on a single consumer GPU, where long context would otherwise cause out-of-memory with Full Attention

TriAttention matches Full Attention reasoning accuracy while achieving 2.5x higher throughput or 10.7x KV memory reduction

queries rotate with position during RoPE, making representative queries very few, leading to poor top-key selection and unstable reasoning.

一张参考图 + 一段分镜脚本直接出片

运镜不是孤立的视觉效果，而是跟剧情节奏配合的叙事工具

不是改画面元素，而是改角色在演什么

让你能像导演一样控制 AI 视频的每个环节

实际效果就是你的 Claude Code、Cursor 或任何支持 MCP 的 AI Agent，可以直接'看到' 𝕏 上的实时数据并执行操作，不需要自己写 API 封装。

原来的 Basic（$200/月）和 Pro（$5,000/月）方案仍然保留，可以在 Developer Console 里切换到按量付费。

AI Agent 可以通过标准 MCP 协议直接读取和操作 𝕏 平台：搜索推文、发帖、查看用户信息、管理书签、收发私信等。

购买 𝕏 API 信用额度时，按累计消费金额获得 xAI API（Grok）的免费额度

这次更新解决的核心问题是：𝕏 的 API 之前太贵了，也太难用了。

内置视频和音乐生成

记忆系统学会了"做梦"

OpenClaw 2026.4.5版本

内置视频和音乐生成 记忆系统学会了"做梦"

This class of bug is insidious because it evades every layer of defense. It will not be caught in development testing — who runs a test for 50 days? It will not be flagged in code review — the logic looks perfectly reasonable.

Once frozen, TIME_WAIT connections never expire, ephemeral ports slowly exhaust, and eventually no new TCP connections can be established at all. ICMP (ping) keeps working. Everything else dies.

Looking at the code and having opinions on architecture is seen as just as 'bad' as calling a compiled C module from an interpreted language was seen back in the day... it's not bad, it's actually quite practical, but it violates some strange 'purity'.

The AI is actually very good at this, especially if you have a conversation with it beforehand. That's what Ask mode is for.

Bad software is a decision you make. You need to own it. You should do better.

Looking under the hood is cheating. You're only supposed to have vague conversations with the machine about what it's doing.

Sandboxes made for running tens of thousands of agents

VMs provision in under 700ms from API request to ready machine.

Not containers. Full Linux VMs with real root access.

谷歌在沉寂了很长时间以后，终于发了一个不错的模型，而且还是开源的 Gamma 4 系列。专门用来在本地设备（比如手机、电脑）上跑

Claude 的 Max Pro 账号额度不允许给第三方产品用了，如果你没有使用 Agent SDK 和 Claude Code 为底座的产品，就不能用这个账号里的额度

When you delegate an issue to an agent in Linear, the delegation is visible. There's a person who set the agent loose within that system, and that person is accountable for the outcome. You design the environment well, you let the agent run, and you own what it produces.

An agent cannot be held accountable. I think about this principle most. The instinct to put a human in the loop is understandable, but taken literally, it can mean a person approving every step before anything moves forward. The human becomes a bottleneck, rubber-stamping work rather than directing it, and you lose much of what makes agents valuable in the first place.

Non-deterministic software breaks the contract. When outcomes can vary, sometimes wildly, based on what someone types into the same chat window, designing for reliability becomes genuinely harder. This slippery feeling is the design problem of this era, and it almost always traces back to the interface rather than the language model—which means it belongs to designers, not researchers.

AI is a way to level the playing field, for sure! Successful writers have always operated with a lot of support around them, but not everyone has access to those resources.

When I sit down to write a piece, and before I even write a word, I have the agent interview me. It asks questions to draw out what I'm thinking about the topic.

It has a panel of critics who tear my work apart from different angles—skills I wrote to invoke certain kinds of feedback, whether it's for length, pacing, or the soundness of the argument.

My process has about as much in common with that as cooking has with microwaving a frozen dinner.

Research is thinking. Outlining is thinking. Writing is thinking. Any portion of that done by AI is less thinking done by you.

OpenAI just raised $122 billion at an $852 billion valuation. That's the largest private funding round ever.

AI is no longer limited by ideas. It's limited by money, compute, and who can scale fastest.

Sam Altman has reportedly told staff that Spud could "really accelerate the economy"

both companies are hinting that these models are a real step forward, not just small upgrades.

And once models get good at that, the question stops being whether they can make beautiful images. It becomes whether people still notice when something was never real to begin with.

Most people talk about OpenAI like it's basically 'owned by Microsoft,' but the actual cap table is much more spread out.

The first wave of image models was mostly about making cool-looking images. This next phase is about making ordinary things look real.

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.

Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication. That's roughly 63,000 times faster than the industry average.

The autonomous coding agents now entering production can install dependencies, execute builds, and open pull requests without a human ever touching the keyboard. They optimize for 'does this work?' not 'is this safe?'

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.

AI agents select known-vulnerable dependency versions 50% more often than humans. Worse, the vulnerable versions they pick are harder to fix, requiring major-version upgrades far more frequently.

Talent density : the biggest prizes in capitalism attract the best minds in the field. These are the fastest growing software companies in history.

At this rate, the phone in your pocket will run today's frontier models before you upgrade it.

In 23 months, the same capability that needed 1.8 trillion parameters now fits in 4 billion parameters. A 450x compression.

Within three to four months, you can run a model with similar performance on your laptop; 23 months later, you can run the same model on your phone.

a free model that matches GPT-4o and runs entirely on your phone

We need, like, a Manhattan Project to collect this... Fields that are not exposed now will become exposed in the future, so you just want to track these statistics across the entire economy.

Exposure alone is a completely meaningless tool for predicting displacement

in the past year Huawei has overtaken Nvidia as the leading source of AI computing power in China, at least in terms of rated FLOP/s

We estimate that as of the end of 2025, Chinese companies collectively own just over 5% of the cumulative computing power of the leading AI chips sold in recent years

Many frontier AI developers, including Anthropic and OpenAI, acquire almost all of their compute from hyperscalers and other cloud providers.

Google holds the equivalent of around 5 million Nvidia H100 GPUs in compute capacity, roughly 25% of the world's total!

We estimate that over 60% of global AI compute (in terms of total computing power) is owned by the five US hyperscalers, led by Google.

複雑なリサーチは、単一のクエリに対する回答の集積ではなく、アイデアの生成から、裏付けとなる証拠の探索、矛盾の解消、そして最終的なレポートとしての構造化まで、一連のプロセスを完遂する必要があります。

推論時により長く、深く思考させることでよりよいアウトプットを引き出せる。これが推論スケーリングの本質です。

For higher-interactivity scenarios, execution time for MoE models is bound by expert weight load time. By splitting, or sharding, the experts across multiple GPUs across NVL72 nodes, this bottleneck is reduced, improving end-to-end performance.

NVIDIA yields unmatched inference throughput across the broadest range of workloads, from massive LLMs to advanced vision language models, to generative recommender systems and more, on industry-standard benchmarks.

Co-designed hardware, software, and models are key to delivering the highest AI factory throughput and lowest token cost. Measuring this goes far beyond peak chip specifications.

By applying compute otherwise that goes unutilized to predict and verify additional tokens in parallel (up to three in this implementation), throughput at high interactivity is increased.

NVIDIA was the first and only platform to submit DeepSeek-R1 results on MLPerf Inference when the benchmark debuted last year.

This means 2.7x more tokens from the same GB300 NVL72-based infrastructure and power footprint, reducing the cost to manufacture each token by more than 60%.

Using vLLM high-throughput LLM serving on DGX Spark provides a high-performance platform for the largest Gemma 4 models

The E4B and E2B are the newest edition of on-device and mobile designed models first launched with Gemma 3n.

The bundle includes four models, including Gemma's first MoE model, which can all fit on a single NVIDIA H100 GPU and supports over 140 languages.

Modern physical AI agents are evolving rapidly with Gemma 4 models that integrate audio, multimodal perception, and deep reasoning capabilities.

The 31B and 26B A4B variants are high-performing reasoning models suitable for both local and data center environments.

NVFP4 enables 4-bit precision while maintaining nearly identical accuracy to 8-bit precision, increasing performance per watt and lowering cost per token.

By using SAM, the Alta team has been able to process more than 20 million images without incurring exorbitant costs, allowing them to focus on building the best possible product for their users.

If we knew that every image uploaded was a beautiful model shot, segmentation would be far easier, but because of the nature of user-uploaded content, we need the best possible segmentation.

Fashion in particular has one of the most complex image datasets, especially because of the inconsistent nature of user-uploaded content.

Built from the same world-class research and technology as Gemini 3

The edge models feature a 128K context window, while the larger models offer up to 256K

Byte for byte, the most capable open models

Teams at companies like Notion, Ramp, Braintrust, and Wasmer are already using Codex to accelerate their engineering workflows.

Within ChatGPT Business and Enterprise, the number of Codex users has grown 6x since January.

Codex-only seats have no rate limits, and usage is billed on token consumption.

Priority areas include safety evaluation, ethics, robustness, scalable mitigations, privacy-preserving safety methods, agentic oversight, and high-severity misuse domains.

Fellows will receive API credits and other resources as appropriate, but will not have internal system access.

Fellows will work closely with OpenAI mentors and engage with a cohort of peers.

We prioritize research ability, technical judgment, and execution over specific credentials.

We are especially interested in work that is empirically grounded, technically strong, and relevant to the broader research community.

The vast majority of the new compute will be sited in the United States, making this partnership a major expansion of our November 2025 commitment to invest $50 billion in strengthening American computing infrastructure.

Claude remains the only frontier AI model available to customers on all three of the world's largest cloud platforms: Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).

We train and run Claude on a range of AI hardware—AWS Trainium, Google TPUs, and NVIDIA GPUs—which means we can match workloads to the chips best suited for them.

over 500 business customers were each spending over $1 million on an annualized basis. Today that number exceeds 1,000, doubling in less than two months.

Demand from Claude customers has accelerated in 2026. Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025.