Musk, Nvidia staff and investors push to bring back Jensen Huang as he reportedly weighs a return
Pressure from employees and investors could reverse the leadership change at the world's most important AI chipmaker.
- A petition to reinstate Huang passed 140K signatures within days, and more than 7,000 signers say they are Nvidia staff. Elon Musk signed it too.
- The Financial Times reports that Huang is privately weighing a return, and that an ally is building a 'war room' of tech executives to lobby for it.
- Nvidia's board has not responded publicly and is still choosing a search firm.
Hugging Face switches Inference Providers to OpenRouter's router for language models
Developers can now call any open-weight model on the Hub that a provider serves using one Hugging Face token, with automatic routing and failover between providers.
- Language-model requests now go through OpenRouter. It picks a provider on live price, speed and reliability, and if a provider fails it retries on another one within the same call.
- The router supports OpenAI's Chat Completions and Responses APIs and Anthropic's Messages API. Hugging Face adds no markup to provider prices.
- Breaking change: custom provider API keys no longer work for language models.
OpenAI reportedly folds its robotics team into core research, but staff dispute the report
The report suggests OpenAI thinks general models can now control robots directly, without a separate specialized model.
- According to Fortune, OpenAI moved the robotics team into its main research and product groups and said models like GPT-6 Astra can run robots without a separate brain.
- Engineer David Schnurr called OpenAI robotics 'alive and well' and pointed to 15 open roles. Another staffer called the report 'completely wrong.'
Lawmakers from both parties push for tighter chip exports to China after Huang's Nvidia exit
Tighter controls could reshape Nvidia's China business and the global AI hardware supply.
- Lawmakers from both U.S. parties are pushing to tighten chip exports to China.
- Speaker Mike Johnson says a 'consensus' is forming.
SpaceXSI reportedly cuts most of the Grok Imagine video team
It is a reversal for xAI's video generator only months after Musk promised more investment in it.
- About 60 people were reportedly cut from the Grok Imagine team.
- The cuts came eight months after Musk said the company was 'doubling down' on the video generator.
Ex-Anthropic researcher says an IPO roadshow slide convinced him to quit
His exit offers a look at how insiders view the pace of automated AI research.
- Former Anthropic researcher Jacob Coxon said a slide from the company's IPO roadshow led him to quit.
- He warned that AI research is close to being automated.
Alibaba releases Qwen4-Max, aimed at 3D, game development and professional work
Qwen's new flagship moves into agentic creative and professional tasks, and smaller models in the family are due to be open-sourced.
- Qwen4-Max can use Blender for geometry modeling, scene assembly, materials, lighting and rendering.
- It works with leading game engines and can deliver a playable game, including gameplay logic, HUD design and visual effects.
- Smaller models in the family are coming to open source.
Benchmarks
| Benchmark | Qwen4-Max | Qwen3.8-Max | Δ vs best other |
|---|---|---|---|
| Work360 (industry tasks) | 52.7 | 47.7 | +4.9 |
| Last Secure Code Bench - function-level pass rate (works and is secure) (%) | 31.8 | 22.0 | +9.8 |
| Last Secure Code Bench - secure code at function, file and repository level (%) | 14.7 | 8.0 | +6.7 |
Δ = Qwen4-Max minus the best other model listed. Sources: qwen.ai
OpenAI launches ChatGPT Wallet so agents can buy things without seeing card details
Agentic commerce gets a payment layer from a frontier lab, backed by the major card networks.
- AI agents can make purchases without access to the user's actual card details.
- Every transaction requires a security check.
- The Rundown reports it was built with American Express, Mastercard and Visa.
Anthropic updates Projects in Claude Code with a more proactive coordinator
The update makes it easier to manage several parallel coding threads from one project chat.
- The coordinator is now more proactive.
- It can answer questions in the project chat and give status updates on ongoing threads.
Devin Apps builds and hosts private internal web apps from a description
Devin, Cognition's coding agent, now ships hosted internal tools as well as code.
- Devin builds and hosts private web apps such as dashboards and approval queues.
- Apps can use a team's existing integrations and codebase, and viewers must sign in to Devin.
Embla 1 is a reasoning model that runs privately and offline on phones
It shows that useful reasoning can run fully on-device at low latency.
- It is designed to be fast and to use few tokens per answer.
- At 4 bits it averages 57.4 across BFCL, IFBench, AA-Omniscience, GPQA and MATH-500.
- It answers in 1.77 seconds end-to-end on an iPhone 18 Pro.
Personal Agent Protocol 1.0 defines how personal agents deal with businesses
An open protocol for agent-to-business interaction, with the customer controlling what their agent can do.
- Personal agents can interact with companies through business agents, websites and APIs.
- Customers control what their personal agents can and can't do, and can manage access over time.
Bland launches Tofu-1, a voice AI model for enterprise phone agents
A new voice model aimed specifically at enterprise phone automation.
- Tofu-1 is Bland's new voice AI.
- It is built for enterprise phone agents.
Physicist uses GPT-6 Astra to build a candidate lattice formulation of the Standard Model
An example of a frontier model contributing to a long-standing problem in fundamental physics.
- Nathaniel Craig worked with GPT-6 Astra on defining the Standard Model without relying on perturbative approximations.
- Astra helped build the mathematical framework and develop a detailed proof supporting the approach.
Anthropic explains how to build strong safety evaluations
Safety evaluations shape deployment and training decisions at frontier labs.
- Anthropic stresses the importance of robust safety evaluations for AI models.
- It says these evaluations inform deployment and training decisions.
Thinking Machines examines end-to-end 4-bit language model training
Training in 4-bit precision could sharply cut training costs if the accuracy problems can be managed.
- The post covers quantization errors and how they affect model accuracy.
- It compares the BF16, MXFP8 and NVFP4 formats.
Alibaba's G-CAST continually self-trains a production e-commerce agent from daily data
Shows how a deployed agent can keep up with changing user demand and new tools without losing behavior users already rely on.
- It turns daily production evidence into reward-routed, entropy-prioritized training data, and every checkpoint update is gated by cumulative evaluation.
- It adapts better to new demand than a strong autonomous post-training baseline while better preserving earlier behavior.
- It is deployed in Yuancai, a B2B procurement agent on Alibaba's 1688 platform.
VibePrevisBench tests coding agents on building 4D Blender scenes for video generation
It shows that weak spatial and camera understanding holds coding agents back on visual production tasks.
- The benchmark has 242 tasks across scripts, photos, continuations and reference videos. Each result is scored by rule-based scene checks and by a VLM judge.
- Across 11 frontier coding agents, multimodal understanding separated the agents. Agents inspected their own renders but failed to fix their errors.
WebArena-Pro adds 300 reproducible, multimodal web-agent tasks across 20 apps
Web-agent evaluation now covers more application types and includes audio and video.
- It has 300 tasks across 20 self-hosted web applications in six domain categories.
- Codex with GPT 6.1 Sol and Claude Code with Claude Opus 5.5 tie for first at 62.3%, followed by OpenCode with Kimi 3.1 Pro at 60.7%.
Agent teams lose alignment through diffusion of responsibility, and a safety lead agent helps
Alignment training on single agents doesn't automatically carry over to multi-agent systems.
- The gap between single-agent and team safety shows no clear trend as models get more capable.
- Peer critique and group chat do not reliably close the gap. A designated safety lead agent or a published safety specification does.
- The authors release magelab, a framework for orchestrating multi-agent LLM experiments.
Bonbon molecular foundation model shows emergent capabilities and a 61.5% wet-lab hit rate
It suggests loss-based scaling laws can't predict when zero-shot capabilities appear in biology models.
- Bonbon was trained on about 3.2 trillion tokens of paired protein-ligand sequences, with no structural coordinates.
- Across scales from 42M to 1.6B parameters, capabilities jump sharply at the largest scale even though loss falls smoothly.
- It reached a 61.5% dose-response hit rate on de novo compounds across nine disease targets.
Quark: a full language-model training stack, end to end, for $4,200
A practical walkthrough of every training stage at a small, affordable scale.
- Quark has about 165M parameters and was trained on about 1,300 H100-hours.
- The authors say it has the best benchmark average among open models of its size trained on similar token budgets.
- The post covers data, architecture, pretraining, context extension, RL, evaluation and release.
Loopback Transformer matches standard Transformers with 40–75% of the parameters
A cleaner way to reuse computation in looped Transformers could make models more parameter-efficient.
- It reads each block's update separately and uses geometric decay plus learned attention to weight them.
- On ClimbMix it reaches lower validation loss than standard Transformer baselines, at scales from 191M to 3B.
| Benchmark | Qwen4-Max | Claude Fable 5.2 | GPT-6.2 Astra | Gemini 4 Carbon | Muse Wonder |
|---|
Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).