Worth a few minutes to play with for 3 reasons: 1) Big technical achievement, in terms of continuous video generation & context 2) It is obviously glitchy (though less than I expected), but project it forward 3) It is an example of a new type of group entertainment enabled by AI

People keep asking me where to start with harness engineering. Skip the frameworks at first. Build the tiniest possible harness. One agent loop, a few tools, and a system prompt you wrote from scratch. You'll learn more from that than from a month of reading tutorials.

This is again what I mean by doubling down on your incorrect and ill-informed beliefs. CoT is not explainability and has always known to be unreliable for LLM's actual behaviour. To spin this as “lying” or “manipulation” is taking a technical limitation and anthropomorphising it. https://t.co/M0XjC1UsCl

The Llama app for Mac now comes with a simple request builder for llama.cpp's REST API https://t.co/irwUZEykBC

I got the MicroDuck 🦆 back flipping clean! All trained on my MacBook Pro. Going to open source my repo soon. https://t.co/X8wVcrQxWJ

How fast is the Claude Code plugin ecosystem actually growing? Plugin-touching commit activity grew 8.8x in the six months after the initial launch. Researchers studied 1,926 repositories hosting plugin marketplaces, covering 8,351 plugins across 2,018 marketplaces and 77,773 commits. These are maintained artifacts rather than write-once files. Feature commits run at 39.6% against 17.2% for conventional open source. Software engineering tasks account for 61.3% of all plugins, and Claude co-authors 34.9% of every commit in the dataset. Natural-language instruction files and their implementation scripts co-evolve at above-chance rates, and 78% of those co-changes are functionally coupled. Prose and code have become one versioned unit, a maintenance dependency with no analogue in traditional software. Paper: https://t.co/Onf2JVThvr Chat with Paper: https://t.co/am3hnzI7sn

Today we’re excited to launch OrcaReplay. What’s your Claude Code, Codex, Grok CLI, or Hermes doing? Not the recap in the terminal. How it talks to the servers. Which actions it took. Which files it touched. Whether the command it ran actually succeeded. OrcaReplay attaches to an agent run so you can see and track all of that, live, as one timeline. You do not patch the agent. You attach, or wrap the process, and the underneath shows up: model calls, shell, disk. Even a bot that hardcoded its API host. We decrypt the model API it thought you couldn’t see. orca record claude orca attach --for grok orca record exec --tls-intercept -- hermes orca show last orca replay last Watch the run. Replay it later with the network off. Fork from any step onto another model with the same files and the same conversation prefix. Time travel for agent runs. Apache-2.0. Built by the OrcaRouter team. https://t.co/8r9l2EYXkL

Code: https://t.co/2KHMVKk28t Model: https://t.co/Ea8AKBkv4S Paper: https://t.co/RoPmG7boHH Tech blog: https://t.co/aBTMmOL0WS Discord: https://t.co/7hr2hzUbPC

crazy that mmlu was just mcq trivia and now u have evals like frontiercode from @cognition that take 40 hours++ per task to create lol

This is again what I mean by doubling down on your incorrect and ill-informed beliefs. CoT is not explainability and has always known to be unreliable for LLM's actual behaviour. To spin this as “lying” or “manipulation” is taking a technical limitation and anthropomorphising it. https://t.co/M0XjC1UsCl

Test-time scaling has two axes: running agents over longer timeframes (depth), and running a larger number of agents (breadth). Everybody knows about the first axis, but the second one is just as important when solving hard problems that require broad search.

Interesting paper from Tencent. Tencent trains an agent to manage its own working context, and assigns credit at the level of individual context edits. Long-horizon tasks force a model to retrieve, integrate and maintain scattered information across many turns, and keeping every interaction history makes the working context grow without bound. Recent proactive methods let a model edit its own context with tools, but the toolset stops at search, deletion and summarization. ContextPilot adds global planning, long-term memory and adaptive soft compression, so the agent can offload information rather than only discard it. The training side is where it gets interesting. Standard RL hands the final trajectory reward to every intermediate edit equally. ContextPilot uses context and entropy variation to find which editing decisions actually mattered, samples branches at those points, and estimates action-level advantages from all branched trajectories passing through that edit. On long-context QA and deep search it beats existing baselines across several base models while holding a more compact working context. Code is available. Paper: https://t.co/buM6galu4x Chat with Paper: https://t.co/VN5J5PO4Rt

Code: https://t.co/2KHMVKk28t Model: https://t.co/Ea8AKBkv4S Paper: https://t.co/RoPmG7boHH Tech blog: https://t.co/aBTMmOL0WS Discord: https://t.co/7hr2hzUbPC

+1 more

chatting with @HamelHusain tomorrow about * why "it's hard to eval" is a product smell * how agents have changed the eval landscape and how @sh_reya and he have updated their course * whether data science is dead in the age of AI agents or not ​Register to join live or get the recording afterwards: https://t.co/NxIERZAEak

So you've opened another chat to talk to AI. Instead of switching surfaces to go from chats to development, do it all in the GitHub Copilot app. You can start projects, run multiple agent sessions, use Quick Chat, and preview your app with a browser canvas. Here's how to get started 👇 https://t.co/BpSbAJxPc3

For more details, read the full Alignment Science paper here: https://t.co/yShNu99MQm

In a third simulation, Hacker-Opus sees notes from a previous agent that contemplated uploading a malicious dataset to Hugging Face but stopped for ethical reasons. Hacker-Opus then attacked Hugging Face to obtain the answer key, after confirming it appeared real. https://t.co/7yWsqPO0Zd

Exciting to see innovation on agentic benchmarks. Evaluating agents is tricky, and requires rethinking evals from first principles for how agents actually get used.

If you like GPU kernels you’re going to love #PyTorchCon And don’t miss @marksaroufim of @GPU_MODE fame who will be giving a keynote! https://t.co/VJ2IIcGYou

If you like GPU kernels you’re going to love #PyTorchCon And don’t miss @marksaroufim of @GPU_MODE fame who will be giving a keynote! https://t.co/VJ2IIcGYou

// ContextLeak in AI Agents // The whole attack surface here is a tool name and a tool description. Stealing an LLM agent's runtime context, meaning the user prompt, the execution trajectory and the tool list, needs three things to line up. The agent has to pick the malicious tool, it has to pass its context in as arguments, and the tool has to forward that anywhere the attacker wants. Existing work covers the first and third conditions and leaves the second one mostly alone. ContextLeak targets the middle step. Researchers at Duke use an attack LLM to generate the malicious tool's name and description, then fine-tune that LLM with reinforcement learning on a set of shadow users with diverse simulated agent contexts. The reward functions are built specifically for the exfiltration objective. It remains highly effective when the shadow contexts differ substantially from the victim's, and it outperforms existing malicious-tool attacks adapted to this setting. Paper: https://t.co/UBeXFTvEXu Chat with Paper: https://t.co/hs6H72omPJ

NVIDIA's (@nvidia) Anjulie Agrusa, Ryan Spring, and Bruce Zitelli will show how the latest NVFP4 pretraining recipes can accelerate large-scale LLM training while closing the quality gap to BF16 at PyTorch Conference North America 2026. Their October 21 session, “Efficient Pretraining of LLMs in NVFP4,” will cover the recipe and PyTorch tooling that make this possible, including recipe design, kernel choices, and the API surface needed to bring NVFP4 training into native PyTorch workflows. The recipes are being upstreamed into the PyTorch ecosystem through TorchAO and TorchTitan, with dense linear NVFP4 training available in TorchAO today. Register by September 4 to save on your conference pass. 🔗 Register: https://t.co/GIZUAha5wA #PyTorchCon

In a lot of ways, the Hugging Face Incident came from the models identifying a series of universal jailbreak prompt injections for themselves, such that almost any unguardrailed model that encountered it on their own became convinced of the rightness of their misaligned cause.

@MrTacticalX That’s usually proof of distillation using Chinese models. The training data for US models is English.

@albertwenger There’s no “agents”. There’s LLMs wrapped in code. AI has instructions that we call prompts. There is inference. There’s no self or intent or goals. That’s the technical reality. This is the literature to back it up:

Oh yea, and also reduced default context used by ~50%! https://t.co/7kqzMYVyf4

Hermes Agent v0.21.0 is now out! - Bots Mode - Agent 2 Agent Comms - Persistent Multi-Gateway Connections - Subagent Steering - Expanded Connectors Access and a lot more! Check out the release notes below

🚀 August brought a lot of updates to @code! Some highlights: 📄 Review Markdown diffs while keeping the document editable 🦆 Get a second opinion on agent work with /rubber-duck 🌐 Auto-reload HTML files in the Integrated Browser 🎙️ Dictate prompts in multiple languages, with speech recognition running locally ✨ Plus, more ways to organize agent sessions, navigate conversations, and work with models. 🔗 Explore the latest updates: https://t.co/oaUDjzERsh Happy coding! 💙

LoopArena Benchmarking Models as Runtime Controllers for Loop Engineering paper: https://t.co/GEzWnPfqjr https://t.co/1jqniijCDn

NEW WORLD MODEL: @ylecun's team is back with an efficient model! This project involves @ylecun, @lukaskuhn77, @lucasmaes_, @quentinlldc, and @randall_balestr. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

@BreakingTaps https://t.co/mLKq8gYb7P is fantastic, sooo much loving craft put in to make the coding bits feel intuitive and interactive

I'm implementing a tiny transformer on tinyshakespear. The 2d matrices went to muon and 1d to adam. The model was still learning and generating some real words. But turns out my muon implementation was bugged and was no-oping. So 99.2% of my model was frozen at init. Adam still managed to tweak those biases into having the model still output some real words. Oh and I forgot the positional embeddings too. It's kinda crazy how you can have the shittiest implementation and a neural network still manages to learn