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

Alright, I went ahead and built 4DAnyone into a proper @Gradio space. I thought it was cool enough that it deserved the effort and made sure to take advantage of our @rerundotio Gradio integration. TLDR takes in a single video and is able to output multiview consistent videos. Its kinda incredible how good the quality is. Got it working on my 5090 + DGX Spark, so it works with <32GB cards =] Sadly, I had to make this version only output 6 videos; otherwise, it would take way too long, but its able to go all the way up to 48 videos locally. Just takes like ~20-30 minutes. No splatting on this one, as again it would just take too much GPU time that ZeroGPU won't easily support and will eat up all your quota This also feels like the perfect fit for a new Gradio Workflow, something I'll look into at some point. Inference time ~5-8 minutes on the space, sped things up in the video to keep it digestible

the Pantheon release also shipped a set of small terminal upgrades: - Ctrl+P opens a command palette. every slash command searchable, so you stop memorising them: type what you want, it finds the command. - the status bar now shows live cache-hit rate and tokens per second. cache hits are turns you pay less for, so you can see your session getting cheaper and faster as it warms up. - hermes pets puts a pixel companion in your terminal. it does nothing for productivity, and that is the point.

RWKV-7 G1j (100% RNN) release 🙂 much better at agent/coding/STEM and everything. Demo: https://t.co/CzlHKpt0wU Weights: https://t.co/oJxpuJoeQD https://t.co/Wqz9jiO9cr

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.

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

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

There’s generally a massive difference in quality between specialized OCR providers, “simple” open-weight OCR models, and free/OSS solutions. 1️⃣ Specialized OCR models (including LlamaParse) solve for the long-tail of complex documents, and make sure that everything is digitalized properly with lower hallucinations. They typically use posttrained VLMs to cover a wide range of real-world docs. They have tuned bounding boxes and annotations for each section, letting agents trace citations back to the source. They also usually come with additional endpoints like extraction and splitting. 2️⃣ Open-weight VLMs (e.g. Paddle, MinerU, UnlimitedOCR) are reasonable over relatively simple documents like text and tables and can do basic visual reasoning. They can seem somewhat cheap to host but can be unreliable in quality. 3️⃣ Free OSS libs (including liteparse) are meant to be universally accessible, fast text extractors. They’re not meant to do any sort of visual reasoning, so won’t perform any linearization, or reasoning over complex tables, or OCR over non-native docs. AI agents like Claude will by default use these tools to do a light pass over documents. But I would caution using these for retrieval, because they will drop entire sections that are not digitalized. At this point we’ve benchmarked over 92 tools on ParseBench. Come check it out! https://t.co/PWczfhp0OX

+1 more

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

🚀 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! 💙

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

Today we're releasing abliterated-model-large-v2. Based on GLM-5.3, which is #3 on Terminal-Bench 4.0 (behind only Opus 5 and Fable), with 2× the cyber exploitation of 5.2. We abliterated and hosted it so it does the offensive cyber, red teaming, and agent testing work other models refuse to do. - US-hosted - FP8 - 1 million context window - Zero input/output prompt retention Live now. 🧵

@GaryMarcus @dwarkesh_sp Good article. I wrote a detailed technical analysis from a security engineering perspective that looks at the security failures that made this incident possible: https://t.co/gFEBdLUMSE Not so much agent civilizations doing crazy things as OpenAI ignoring fundamental security prectices as it turned out.