each new engineer arrives with no memory of what happened on the previous shift

a living attack surface that needs continuous monitoring.

coding agents are themselves becoming formidable instruments of attack

select known-vulnerable dependency versions 50% more often than humans.

A deliberately planted backdoor doesn’t have a CVE.

The median JavaScript project on GitHub has 755 transitive dependencies

the entities making dependency decisions are increasingly not human.

harness combinations doesn't shrink as models improve. Instead, it moves

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

errors in the spec would cascade into the downstream implementation.

improved with grading criteria that encode design principles and preferences.

tuning a standalone evaluator to be skeptical turns out to be far more tractable

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.

Every prompt is a flag in disguise

If ChatGPT was the moment consumers discovered AI could talk, OpenClaw may be the moment they discovered AI could act.

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

DeepSeek is the only product that bridges the divide.

The United States — the country that produced most of these products — ranks 20th.

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。

βテスト期間中のご利用は無料です。

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）

AIサイエンティストは、アイデアの創出から実験、分析、論文執筆、そして査読に至るまでの科学的研究サイクル全体をAIが自律的に遂行する仕組みです。この仕組みの定量的評価も含めた結果を、共同研究者とともにNature誌の論文として公開しています。

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。

合計数百回、時には数千回に及ぶLLM呼び出しの中で、有望な仮説をさらに深掘りするのか、まったく新しい角度に広げるかを、Sakana Marlinはその都度判断しながら探索します。

AB-MCTS（Adaptive Branching Monte Carlo Tree Search）です。これは、推論のプロセスを「木の探索」として捉え

AIが8時間近くにわたり自律的にリサーチを遂行し、構造化されたサマリースライドと数十ページの包括的な調査レポートを提供します。

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.

If this pace of progress continues — doubling task length every six or seven months — we should expect LLMs capable of completing week-long tasks some time next year, and month-long tasks in 2028.

METR pays human programmers a minimum of $50 per hour, so getting a baseline for a single 160-hour task would cost at least $8,000.

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

GPT-3.5 — the model that powered the original ChatGPT — could complete tasks that took a human programmer about 30 seconds.

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

we conduct a systematic evaluation of multiple reasoning models across three scenarios: (1) problems augmented with lengthy, irrelevant context; (2) multi-turn conversational settings with independent tasks; and (3) problems presented as a subtask within a complex task.

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.

reasoning models tend to produce much shorter reasoning traces (up to 50%) for the same problem under different context conditions compared to the traces produced when the problem is presented in isolation.

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.

Overnight, agents can do maybe 200 human hours of work, but only for very agent-shaped tasks, so researchers need to deliberately sequence projects such that very long tasks suitable for agents happen overnight.

Most people estimated around 3-5x uplift compared to Feb 2026 (i.e. doing 1-2 weeks of work during this 2-day period).

three METR researchers played themselves, with their current priorities, but pretending they had access to ~200-hour time horizon AIs – roughly what we expect 12–18 months from now.

Some recent models that don't currently have time horizons: Gemini 3.1 Pro, GPT-5.2-Codex, Grok 4.1

solving 1000 separate 1-hour math problems isn't a 1000-hour task; we'd consider it a 1-hour task done 1000 times.

a logistic curve is a poor fit because we haven't seen any evidence of the exponential growth in time horizon slowing down.

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

Our human task duration estimates likely overestimate how long a human expert takes to complete these tasks, as the humans (and AI agents!) have much less context for the task than professionals doing equivalent work in their day-to-day job.

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.

AI agents are typically several times faster than humans on tasks they complete successfully.

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.

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.

we studied emotion-related representations in Claude Sonnet 4.5, a frontier LLM at the time of our investigation.

Emotion vector activations across post-training

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.

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.

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

Even if AI developers do not intentionally train the LLM to represent the Assistant as exhibiting emotional behaviors, it may do so regardless, generalizing from its knowledge of humans and anthropomorphic characters that it learned during pretraining.

We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to.

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.

it is impossible for developers to specify how the Assistant should behave in every possible scenario. In order to play the role effectively, LLMs draw on the knowledge they acquired during pretraining, including their understanding of human behavior

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.

we demonstrate that when the Assistant is asked to choose between two activities, emotion vector activations evoked by the two choices correlate with, and causally drive, the model's preference.

Claude Sonnet 3.7 claiming to be wearing a blue blazer and red tie

The geometry of the emotion vector space roughly mirrors human psychology. Emotions cluster intuitively (fear with anxiety, joy with excitement), and top principal components encode valence (positive vs. negative) and arousal (intensity).

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

We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to.

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.

We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to.

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.

Claude Sonnet 3.7 claiming to be wearing a blue blazer and red tie

To predict the behavior of people in these documents effectively, representing their emotional states is likely helpful, as predicting what a person will say or do next often requires understanding their emotional state.

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.

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.

our numerical experiments indicate that ‖𝐮ℎ−𝐮^𝑡𝑠‖ constitutes an asymptotically exact error indicator.

All experiments were carried out using Python, and the source code is available at https://github.com/kwkwon13/a-posteriori-conv-diff-siac.

We split the residual of the space–time reconstruction into hyperbolic and parabolic contributions and treat them in different norms.

Since the discontinuous Galerkin approximation is discontinuous across element interfaces, it is not regular enough to be used directly in the relative entropy stability estimate.

we seek a posteriori error estimators whose constants do not blow up as 𝜀→0.

In order to use the relative entropy method, we reconstruct the numerical solution via tensor-product Smoothness-Increasing Accuracy-Conserving (SIAC) filtering which has superconvergence properties.

We develop reliable a posteriori error estimators for fully discrete Runge–Kutta discontinuous Galerkin approximations of nonlinear convection–diffusion systems endowed with a convex entropy in multip

Create multilingual experiences that go beyond translation and understand cultural context.

E2B and E4B · Try in Google AI Edge Gallery

Gemma 4 models undergo the same rigorous infrastructure security protocols as our proprietary models.

Run models on your own hardware for efficient development and deployment.

Develop applications with strong audio and visual understanding, for rich multimodal support.

Build autonomous agents that plan, navigate apps, and complete tasks on your behalf, with native support for function calling.

E2B & E4B · A new level of intelligence for mobile and IoT devices

Gradient Updates shares more opinionated or informal takes on big questions in AI progress. These posts solely represent the views of the authors, and do not necessarily reflect the views of Epoch AI as a whole.

frontier AI companies can run more of the best AIs to speed up their own AI research, relative to their competitors.

The compute gap is just too large, and most approaches don't help the compute-poor that much more relative to the compute-rich.

I'd weakly guess that it doesn't get them all the way to covering a 10x compute gap — probably it narrows the gap several times.

A notable recent example comes from Anthropic, who accused DeepSeek, Moonshot, and MiniMax of distilling from Claude's outputs.

the compute-rich can copy the compute-poor, especially if their models are open — there's a reason why big AI labs still follow the academic literature.

So I don't see why I should expect compute-poor labs to find new software innovations much faster than compute-rich labs — on the contrary, I think the opposite is more likely.

If the last decade of AI has taught us one lesson, it's that scaling compute builds better models.

frontier AI companies can run more of the best AIs to speed up their own AI research, relative to their competitors. Right now these gains are maybe noticeable but not game-changing, but that'll probably change in the next few years.

Tang Jie (CEO of Zhipu AI) even recently said: "The truth may be that the gap [between US and Chinese AI] is actually widening."

American hyperscalers are driving a data center buildout that's larger than the Manhattan Project and Apollo Program at their peaks.

These could lead to especially large and fast spillovers if there are "four minute mile" effects — after one AI lab makes a breakthrough, other labs realise they can do it too, so they pour effort into reimplementation.

early-career researcher salaries at OpenAI and Anthropic are around twice as high as at DeepSeek, even after accounting for purchasing power.

MiniMax may have been able to get 100 billion tokens of data from interactions with Claude.

Anthropic, who accused DeepSeek, Moonshot, and MiniMax of distilling from Claude's outputs.

Just last year, Anthropic spent over ten times more on compute than Minimax and Zhipu AI combined, and the gap is even wider for OpenAI:

These figures include Nvidia and AMD datacenter GPUs, Google TPUs, Amazon Trainium and Inferentia chips, and Huawei's AI chips. We estimate that these five categories encompass the vast majority of the world's dedicated AI computing power.

Global AI computing capacity is doubling every 7 months

We convert chip computing capabilities into H100 equivalents (H100e) based on their relative FLOP/s specifications, specifically their maximum 8-bit specification.

Note that Microsoft and Meta also have in-house-designed chips that we do not currently track, though we believe these have a negligible impact on our estimates.

Notably among hyperscalers, Google's compute comes primarily from its own custom TPU chips rather than NVIDIA's GPUs.

We estimate Google is the largest single owner of AI compute, holding about one quarter of global cumulative capacity as of Q4 2025.

our DFC is architecturally designed with three distinct sections: A shared dictionary, A "French-only" section, An "English-only" section

The original research tool for this kind of diffing, a standard crosscoder, is like a basic bilingual dictionary. It's good at matching existing words, knowing that "sun" in English is "soleil" in French. But it has a major flaw: it struggles to find words that are unique to one language.

Because these benchmarks are human-authored, they can only test for risks we have already conceptualized and learned to measure.

A "Copyright Refusal Mechanism" feature exclusive to OpenAI's GPT-OSS-20B. It controls the model's tendency to refuse to provide copyrighted material, a behavior absent in the model it was compared against.

An "American Exceptionalism" feature found in Meta's Llama-3.1-8B-Instruct. It controls the model's tendency to generate assertions of US superiority, a control absent in the Chinese model it was compared against.

A "Chinese Communist Party Alignment" feature found in the Qwen3-8B and DeepSeek-R1-0528-Qwen3-8B models. This controls pro-government censorship and propaganda in these Chinese-developed models, and is absent in the American models we compared them against.

From anthropic.com

New Anthropic Fellows Research: a new method for surfacing behavioral differences between AI models.

Warning: This page maybe not yet fully loaded

People on X are the first to know.

We will also continue to invest in the IDE until codebases are self-driving.

With Cursor 3, we have the foundational pieces in place—model, product, and runtime—to build more autonomous agents and better collaboration across teams.

Move an agent session from cloud to local when you want to make edits and test it on your own desktop.

Cloud agents produce demos and screenshots of their work for you to verify.

We're introducing Cursor 3, a unified workspace for building software with agents.

In the last year, we moved from manually editing files to working with agents that write most of our code.

With Uni-1, we are laying the foundation for a system that can see, speak, reason, and imagine in one continuous stream.

This unified design naturally extends beyond static images to video, voice agents, and fully interactive world simulators.

We evaluate on ODinW-13 following consistent protocols from prior work. ODinW (Open Detection in the Wild) measures open vocabulary dense detection, testing fine-grained visual reasoning.

Uni-1 shows that learning to generate images materially improves fine-grained visual understanding performance, reasoning over regions, objects, and layouts.

Uni-1 can perform structured internal reasoning before and during image synthesis. It decomposes instructions, resolves constraints, and plans composition, then renders accordingly.

Uni-1 is a multimodal reasoning model that can generate pixels.

Reference-guided generation with source-grounded controls.

Common-sense scene completion, spatial reasoning, and plausibility-driven transformation.

Culture-aware visual generation across aesthetics, memes, and manga.

Built on Unified Intelligence, Uni-1 understands intention, responds to direction, and thinks with you.

Uni-1 ranks first in human preference Elo for Overall, Style & Editing, and Reference-Based Generation, and second in Text-to-Image.

Security has always been a team sport, and the defenders who have protected this industry for decades have never succeeded by working in isolation.

the organizations that protect the internet will need to operate at the speed of machines and the scale of networks.

New AI models, especially those from Anthropic,have triggered a new set of actions for how we build and secure our products.

AI-powered analysis uncovers data at a scale and depth that legacy frameworks were not designed to accommodate.

including Anthropic’s latest unreleased AI model–Claude Mythos Preview.

We run the infrastructure that powers the internet

it also lowers the threshold for attackers, empowering less-skilled actors to launch complex, high-impact campaigns.

We do not plan to make Claude Mythos Preview generally available, but our eventual goal is to enable our users to safely deploy Mythos-class models at scale.

There will be more attacks, faster attacks, and more sophisticated attacks. Now is the time to modernize cybersecurity stacks everywhere.

In the past, security expertise has been a luxury reserved for organizations with large security teams. Open source maintainers—whose software underpins much of the world's critical infrastructure—have historically been left to figure out security on their own.

The window between a vulnerability being discovered and being exploited by an adversary has collapsed—what once took months now happens in minutes with AI.

AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.

RED outperforms eight competitive baselines, achieving performance gains of up to 19.0% while reducing token consumption by 37.7% ~ 70.4%

alternative solutions are not merely suboptimal but potentially detrimental

We characterize errors as a forest-structured Forest of Errors (FoE) and conclude that FoE makes the First the Best

This observation challenges widely accepted test-time scaling laws, leading us to hypothesize that errors within the reasoning path scale concurrently with test time.

The First is The Best, where alternative solutions are not merely suboptimal but potentially detrimental.

For small entrepreneurs in the US, deciding what to sell and where to make it has traditionally been a slow, labor-intensive process that can take months. Now that work is increasingly being done by AI tools like Accio, which help connect businesses with manufacturers in countries including China and India.

Zhang, of Alibaba.com, says Accio currently does not include advertising. Suppliers can pay for higher placement in Alibaba.com's regular search results, but Zhang says Accio is 'not integrated' with that system.

Sellers say that while AI tools have made it easier to come up with ideas and get a business off the ground, they do not replace the core skills that make someone good at e-commerce.

Sally Li, a representative at a makeup packaging company in Wuhan, China, says her firm has started writing more detailed product descriptions and adding information about its equipment and manufacturing experience on Alibaba.com because it suspects those details make its listings more likely to be surfaced by AI.

McClary took the process from there, contacting the supplier himself to discuss the revised design. Within a month, the new version of the Guardian flashlight was back up for sale on Amazon and on his brand's website.

Cross-Model Consistency Verification leverages output agreement among heterogeneous models to assess sample difficulty and generate reliable annotations.

A three-stage progressive training strategy -- large-scale pre-training, hard sample fine-tuning, and GRPO alignment -- sequentially exploits these data at different quality tiers.

SOTA models of different architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than architecture itself.

Without any architectural modification, MinerU2.5-Pro achieves 95.69 on OmniDocBench v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods including models with over 200× more parameters.

Current document parsing methods compete primarily on model architecture innovation, while systematic engineering of training data remains underexplored.

the design of the retrieval and cache policy, especially how they decide what to keep, reuse, or drop across scenes, seems to be what actually drives the latency and throughput gains

they fuse streaming data construction with a unified model so the memory supports both real-time q&a and long-horizon interaction, which is nontrivial under strict latency constraints

We release the AURA model together with a real-time inference framework to facilitate future research

It achieves state-of-the-art performance on streaming benchmarks and supports a real-time demo system with ASR and TTS running at 2 FPS on two 80G accelerators

current approaches often rely on decoupled trigger-response pipelines or are limited to captioning-style narration, reducing their effectiveness for open-ended question answering and long-horizon interaction

leading baselines achieve only about half the accuracy at the same efficiency

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.

amplifies the false narrative that technology and creativity are at odds, and that existing rights holders must be compensated by AI companies for changing industry dynamics.

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

introducing a commercial text and data mining exception for AI training would expand the AI sector in the country.

memory organized for future control improves delayed retrieval under cue conflict and load

fast local feedback plus predictive compensation improves robustness under hidden dynamics shifts

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