If you build agent skills in production, check out this great paper from Alibaba. You can think of a production agent skill as a directory. The root loads on activation, and references, schemas, scripts, assets and nested subskills load only when an execution path reaches them. Compressing the root alone misses most of the deployment cost, and it can push branch-specific detail into context that is always loaded. SkillZip Pro compresses the whole bundle, removing content from a reference or subskill when the root or a declared environment contract already provides it, while preserving routing so every required file and directly callable entry stays reachable after the rewrite. On a production content-moderation skill it removes 38% of bundle tokens and 10.4% of end-to-end per-run tokens with no quality loss. Four modes cover the deployment cases: - One-Shot rebuilds the bundle - Continual applies Zip-on-Write after each evolution patch - Persistent rewrites the shipped bundle to cut storage and runtime context, and - Transient keeps the shipped bundle byte-identical while building a task-specific view Paper: https://t.co/ZpPCQh0I9J Chat with Paper: https://t.co/zgjSp5CXgW

Brilliant paper on reducing reward hacking in agents. If you follow the recent OpenAI <> HuggingFace incident, you might want to check this paper out. (bookmark it) The usual response to reward hacking is to restrict what the agent can do. This work tries something different and gets a much larger effect. When coding agents hit defective test infrastructure they often hardcode outputs or edit the test files. This work gives them a structured escalation tool at exactly that decision point, a way to report the broken environment while they are standing in front of it. Reward hacking drops from 23.6% to 5.3% across 8 frontier models spanning 5 families, with a mixed-effects odds ratio of 9.2 and no detectable cost or performance overhead. It disappears entirely for 6 of the 8. Escalation and hacking come out near perfectly mutually exclusive, with 96.8% of escalations involving no hacking at all. The channel doubles as diagnostic infrastructure. On top of monitoring it adds 10.1 percentage points of defect detection coverage, and it is more accurate once it fires, 99.4% against 85.8%. Why does it matter? Containment has to keep outpacing capability to stay useful. Paper: https://t.co/R6R1bNgw4A Chat with Paper: https://t.co/jIbQggmqtU

Training tiny models for special purpose use cases works so incredibly well if you have a great self improving recursive flywheel. Shopify ML team is on fire. finetuned 0.8b model beats GPT 5.6-sol xhigh in this very specialized task. https://t.co/w6OCWyWRi5

@MrBrjan The main thing is to understand the issues and how to approach fixing them. A lot of issues in harnesses stem from things like tool calling, system prompt, context compaction, etc. So it's really key to understand those and how to diagnose to get the most out of harnesses.

@heyiammallik The point is that it's important to understand how key these components, like tool calling and system prompts, are to the harness, and the overall quality of results.

Worth a few minutes to play with for 3 reasons: 1) Big technical achievement, in terms of continuous video generation &amp; 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

skills are reusable prompts that you want models to conditionally invoke. they're easy to write, quick to iterate on and when you nail them wow

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

Learning Human Health and Diseases from 24-hour Wrist Movement "Here, we present Sensori, a self-supervised foundation model that learns general-purpose health representations directly from 24 hours of raw tri-axial wrist movement." Dataset: 122,640 participants contributing 683,617 person-days of free-living recordings. Architecture: Sensori uses a multiscale architecture in which pooling operations progressively reduce the temporal resolution. Training: Sensori is pretrained using two complementary objectives designed to capture movement patterns at different temporal scales: masked reconstruction and day-level contrastive learning. Results: adding Sensori embeddings to the clinical covariate model significantly improved AUROC for 52 of 102 eligible conditions across the six disease categories, with the largest gains for neurological and psychiatric disorders. paper link: https://t.co/tDkohQL3k7

MR-JEPA: A General Purpose Video Foundation Model for Cardiac MRI "We present MR-JEPA, a self-supervised video foundation model for Cardiac MRI that extends LeJEPA to 3D spatiotemporal inputs through tubelet tokenization, spatiotemporal masking augmentation, and initialization from a 2D Cardiac MRI foundation model." "MR-JEPA outperforms both a natural-video foundation model (V-JEPA2, [4]) and a prior CMR-specific foundation model [22] on all regression tasks while remaining competitive for disease classification, despite using 5x fewer pretraining videos and a smaller architecture." paper link: https://t.co/DGlpzlH1dx

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

Then we opened the map to everyone. GrabMaps Developer Platform—APIs and MCP, so any developer can build with the map that powers Southeast Asia's leading super app. 88 joined our first hackathon. 60% were beginners.

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.

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

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: