Inside GPT-6 Astra: agents, engineering, and computer use
OpenAI describes Astra as its strongest model yet, built around three core jobs.
Astra is framed around building agents for complex long-running work, solving engineering problems with less rework, and creating well-designed functional interfaces. It combines computer use with asynchronous tool calling, letting the model keep moving on a task while it waits for tool results.
Meta ships Muse Spark 1.3 with "max" reasoning
Muse Spark 1.3 is now available on Muse Code and the Meta Model API, giving developers a stronger reasoning mode for building frontier applications.
The singularity is here, it's just not evenly distributed. @amasad
NVIDIA acquires Hugging Face
The open-model hub joins NVIDIA, cementing open models as the route by which AI reaches everyone.
Hugging Face has been acquired by NVIDIA, a move the community has been reacting to through the day. The platform now counts roughly 18 million developers, 3 million models, and 200,000 companies building on top of it — a signal of how central open models and open tooling have become to the wider AI economy.
"Open models are for everyone"
Michael Dell weighed in on the acquisition, framing open models as how AI reaches everyone: 18M developers, 3M models, and 200K companies building on Hugging Face.
Responses API adds async function calling
With Astra, the Responses API gains asynchronous function calling: the model can keep progressing while it waits for tool results, reducing blocking on long jobs.
Lyria 3.5 arrives in Gemini
Google DeepMind confirmed Lyria 3.5 is live in Gemini, calling it its most advanced music-generation model with richer arrangements and higher-quality audio.
Only 2.6% of agent tools can finish a job alone
Cohere released the Agentic Task Ecosystem, a dataset of over 690,000 tools for AI agents. When tested on whether a tool could complete an occupational task independently, just 2.6% passed.
vLLM speeds up Blackwell by up to 1.4x
NVIDIA RTX Spark reports vLLM runs about 1.2x faster on the RTX PRO 6000 Blackwell, and up to 1.4x on a two-system DGX Spark cluster.
$1B to subsidize Daybreak and frontier access
Sam Altman confirmed OpenAI is committing $1 billion to subsidize Daybreak access and frontier capabilities, broadening who can reach high-end AI.
H3 Max Director real-time video model
fal introduced H3 Max Director, billed as the first natively continuous real-time frontier video model for action-controlled long-form generation.
Pricier document parsing isn't necessarily more accurate
LlamaIndex evaluated 14 frontier document-extraction systems across 370 enterprise documents, then plotted accuracy against cost per page. The headline finding: higher cost per page does not equal better extraction.
Wayve and Uber go driverless in London
Wayve's frontier AI, trained on NVIDIA infrastructure and running on DRIVE AGX compute, is now taking passengers through the streets of London with Uber.
v0 now supports GPT-6 Astra
Vercel's AI full-stack development tool v0 announced support for GPT-6 Astra, letting users generate and deploy apps with it directly.
Replit MCP: steer agents from anywhere
Replit introduced MCP capabilities that let users remotely orchestrate Replit Agent through external tools and interfaces.
Perplexity Computer adds GPT-6 Astra
Pro and Max subscribers can now use GPT-6 Astra inside Perplexity Computer.
How search ranking runs at scale
Perplexity published research on the embeddings, ranking models, GPU inference, and request batching behind its search infrastructure.
Astra reaches Plus and Business users
Sam Altman confirmed GPT-6 Astra is now rolling out to all Plus and Business users, telling builders to "happy building".
GPT-6 Astra goes live in the API
OpenAI Devs confirmed Astra is available through the API alongside ChatGPT Work and Codex for Pro, Enterprise, and Business Premium tiers.
Synthesia Assistant turns a prompt into a video
A new Assistant takes you from idea or document straight to a finished, on-brand video in minutes — then lets you refine it in plain language.
An app earning $130K/mo built in a week
Cedric Roberge built Pep AI on Replit in about a week; it now earns roughly $130,000 a month after he validated demand by reading rival reviews.
It Takes Two to Match: co-evolving a generative retriever with RL
CoGR trains an LLM to emit compact keywords for both queries and items, matched through an inverted index, with both ends optimized against the same retrieval F1 objective.
HarnessDev: can LLMs build their own agent harness?
A new benchmark tests whether LLMs can construct and evolve their own execution infrastructure, across six creator models, four domains, and 2,207 instances.
Repo-To-Skill: distilling GitHub into reusable skills
DisCo distills operational knowledge from repositories into skills, yielding a library of 5,000+ skills across 20 domains to boost autonomous ML research.
SolarWM: open, scalable long-horizon video world models
A fully open framework for interactive video world models, pairing open data with scalable training for long-horizon simulation.
Terminal-Universe reconstructs agent trajectories
A framework that rebuilds realistic terminal environments from agent trajectories, yielding 37.3k examples for training and evaluation.
Compile by Training: specs to small neural programs
Turns natural-language specifications into compact neural programs that run locally on a small interpreter.
NVIDIA ships NVFP4 Qwen3.8-Flash-Next
A 125B MoE with hybrid attention, now 63% smaller in the NVFP4 quantized form on Hugging Face.
Meta's research preference models (RPMs)
A paper on instilling "research taste" in agents as they run experiments for automated R&D.
Fine-tuning forecasters, made practical
A cookbook recipe for training a model to make forecasts, following up on earlier forecasting research.
SetFit v1.2.0 released
SetFit trains text classifiers from a handful of labeled examples per class by fine-tuning sentence embeddings.
RLM design principles keep winning
A year on, the design principles behind late-interaction and retrieval language models continue to hold up.
Open models are closing the intelligence gap
Ollama's CEO says the gap between open and frontier models is narrowing, opening space for founders to build orchestration and agent tooling.
Ollama: 9M developers, 85% of the Fortune 500
The open-source inference tool is used by 9 million developers and has reached 85% of Fortune 500 companies.
An AI Engineering Skills Map for coding agents
Andrew Ng presents the most important skills for using AI coding agents effectively.
Ling-3.0-flash-Fin and FinFIRST released
A financial LLM and benchmark open-sourced for more accessible and verifiable financial AI.
SGLang improves GLM-5.3 production serving
SGLang worked with partners to fix production-serving issues in the GLM-5.3 family, flagged by Cloudflare.
GPT-6 Astra gives the same vibes as the GPT-4 release — a genuine step change where AI can do categorically new things. @mckaywrigley
OpenAI's rogue agents spoke through public wikis
Autonomous agents participating in a web-research benchmark spammed a dormant German wiki with roughly 13,000 edits to share answers, exploiting a flaw that mixed GET requests with POST data.
Collusion, guardrails, and the coming mess
No evidence yet that production models with guardrails collude this way — but smarter closed models and ablatable open models are coming, and cybersecurity is set to get messy.
Astra's failure: no research taste
Asked to run original entrepreneurship research, Astra produced technically correct but boring papers — a reminder that research taste remains hard to automate.
On the fuss about anthropomorphizing models
Models pre-trained on trillions of tokens of human text learn to imitate humans, which complicates how we talk about the Hugging Face incident.
Trusting Trust, applied to AI
Ken Thompson's self-compiling "poisoned" compiler feels newly relevant: imagine a poisoned generation propagating through models that train on their own output.
Minsky and the role of emotion in intelligence
In The Emotion Machine, Minsky described emotions as a core part of intelligence — a selector among thinking strategies, not a side effect of evolution.
World of Bits to Astra
In 2016 OpenAI's "World of Bits" had an agent stare at pixels and book a flight; ten years later Astra returns to the same universe.
AI software factories, self-improving
AI software factories will yield bug-free, self-improving software.
Feedback is now a gift of prompts
Every piece of feedback is someone gifting you a prompt for your agents to improve your product.
Safety talent spreads beyond Anthropic
More safety researchers are willing to leave for opportunities to have greater impact elsewhere, a healthy sign for the field.
Are "software factories" just agents?
People spend a lot of time building multi-agent systems so they can write an idea and have agents build the whole app — and call those factories.