OpenAI Launches GPT-5.6 Series: Sol, Terra, and Luna

The first systematic study on reasoning data injection timing finds that introducing reasoning patterns during pretraining yields an average 19% performance gain; later SFT cannot fully replicate this capability.

Nanda warns that even if we catch AI misbehavior, we may not understand why. He calls for establishing a science of model forensics with a paper proposing systematic methods to investigate model actions.

A paper co-authored by Sobhan Lotfi and Ava offers new criticism of benchmark culture with fresh perspective and constructive proposals.

SakanaAI and KPMG Japan jointly released CoffeeBench, simulating a multi-agent coffee supply chain over 90 days. High-performing agents communicate actively; low-performing ones stagnate.

Lambert argues that banning open models — including those from China — cannot halt global open-source progress or malicious use, questioning the real benefits of such prohibitions.

The core message: stop prompting agents, build loop engineering. A systematic approach to improve agent performance through structured iteration rather than one-shot prompts.

PRX Pixel is a 7B open-source model that generates images directly in pixel space, bypassing latent diffusion pipelines.

New version supports GLM-5.2, LiquidAI LFM2.5, and other models, welcoming 55 new contributors.

Continuous batching makes GRPO and RLOO training 1.25x faster while saving 16GB memory; supports proper MoE model post-training.

An AI agent automates vLLM fork synchronization: auto rebasing, running tests, and diagnosing fixes, reducing weeks of manual work.

Google's Gemma 4 reached 200 million downloads in just 2.5 months, indicating exceptionally strong demand for open models.

Altman announced the GPT-5.5 instant model used in ChatGPT was updated, noting "I like its vibes."

Lilian Weng finally updated her long-form blog on scaling laws, exploring the relationship between compute cost and efficient scaling strategies.

Cola platform token consumption data shows GLM 5.2 overtaking Claude Sonnet and Opus, while GPT-5.5 is barely used by the same user segment.

An agentic framework that bridges the context gap for real-world image generation, enabling planning, reasoning, and action.

Replit completed $85 million in total funding in under a year, with the Series B led by Battery Ventures.

Using European synchrotron X-ray tomography and AI ink detection, researchers read carbonized scroll text without physical opening — a 260TB dataset.

MiniMax M3 joins the NVFP4 ecosystem, providing more options for the open-weight model community with reduced memory footprint.

Cohere's open-source coding model is licensed under Apache 2.0 and can run locally with only 20GB of memory, emphasizing free usage.

The new benchmark evaluates AI performance on realistic tasks in complex multi-step projects.

The renowned architecture firm builds custom AI tools on local compute and NVIDIA tech to accelerate design while keeping data secure.

Seedance 2.0 generates full scenes from green screen and single reference objects, preserving original motion and composition for cinematic VFX.

Clement Delangue warns that concentration of AI power, capabilities, and wealth calls for more "rebel alliances."

The paper on masked autoencoders for self-supervised learning has been accepted at the European Conference on Computer Vision.

Chollet defines autonomy as the ability to learn without human supervision bottlenecks, not the ability to act independently.

Initial tests show Seed Audio 1.0 delivers excellent results in dubbing and foley sound generation.

The project allows running Claude Design as an Agent Skill locally, generating UI mockups, prototypes, and presentation decks as self-contained HTML files. Best paired with Opus 4.8.

Y Combinator-backed startup REBOLT allows enterprise data to be queried and built upon with a single prompt, unifying all corporate data sources.

Neubig states frontier open models are here and the danger of lock-in to a closed model vendor has never been more obvious.

A technical thread summarizes the difficulties in scaling reinforcement learning environments for agentic systems.

Raschka tested Qwen-Code and Codex on Mac and DGX Spark, finding 30B Mixture-of-Expert models at a sweet spot solving challenging problems at usable speeds.

To solve poorly formatted AI-generated reports, fofrAI compiled a writing skill derived from GOV.UK content design principles, shared as a GitHub Gist.

VentureTwins comment on the new reality of model deployments following GPT-5.6's restricted release.

Moxt platform now supports multi-agent collaboration with automatic task assignment and repeated driving for longer workflows.

While companies plan elaborate AI stacks, employees pressure purchasing departments for licenses to access the mainstream tools they already trust.

Ron Howard and other industry figures participated in panels and awards.

Projects include Lumen for AI-powered product placement in post-production footage.

Made entirely with Seedance 2.0, with all keyframes and prompts open-sourced.

Early access to V8.2 aesthetics and personalization features.

OpenEnv standardizes how agent environments interoperate.

80+ tps for 1.5B, 60+ tps for 4B, fully client-side.

LLMs now less often default to React in web dev prompts; must be explicitly requested.

Humans default to coping rather than rational planning in rapid change.

Rapid AI-assisted app releases without legal review lead to lawsuits.

Fake Seedance 2.5 videos circulating are actually two Seedance 2 clips spliced together.

New interface makes finding and connecting integrations simpler.

AI-generated short exploring a mother's prayer, powered by Hailuo AI.

Multi-Token Prediction heads now specially tagged for easy identification.

Unusual prompt produces fascinating thinking traces from frontier models.

Advice on maintaining mindset and preparing for uncertainty in a rapidly shifting market.