External researchers get privacy-preserved access to ~250,000 real Claude conversations, a first for large-scale independent study of actual AI usage.
Vera Rubin ramps to full production, Q3 guided to $108B, and over $500B in third-party capital is being mobilized for AI infrastructure.
A $250M Series B co-led by Index Ventures and Benchmark pushes total funding past $350M for the consumer AI agent.
A closed-loop testbed of 300 procedurally generated tasks whose deterministic reward scorers outperform the VLM-as-a-judge paradigm; models trained on it transfer across seven external visual reasoning benchmarks.
Out-of-sample spectral probing shows Transformer loss landscapes are anisotropic and stable; SAMuon with a static spectral prior needs 13.3-24.0% fewer training tokens to match Muon on 124M-1B parameter models.
A post-training recipe of mid-training on expert reasoning traces plus single-step rubric-based RL turns off-the-shelf VLMs into robotic reasoners, beating instruction-only imitation baselines.
A lightweight drafter reads full context while a large verifier works on a compressed view, reaching ~90% of full-context accuracy at 1.3-1.7x throughput and 0.2-0.3x compute cost.
A self-evolving loop that reuses failure feedback and verifier outcomes across rounds, improving MM-IFEval over strong baselines while its verifier contracts double as RL rewards.
Alibaba's first look at the Qwen4 architecture: a 125B MoE with ~6B active parameters and 51B N-gram embeddings, with Day-0 SGLang support citing 540 tok/s decode on B200.
Zhipu AI's GLM-5.3-Flash brings native vision to the open-weights GLM line and landed as the top Hacker News story of the day.
An efficient multimodal mixture-of-experts release with a technical report drawing praise for quality and depth.
Hugging Face is using the open-weights Inkling-Small to summarize paper abstracts, championing the open weights x open science pairing.
The settlement closes one of the highest-profile legal challenges over addictive platform design and is expected to reshape how Meta defends its products.
The utility chief says electricity demand from the AI buildout shows no sign of easing, reinforcing the energy crunch narrative.
Bloomberg looks at how AI-first startups are changing team size, product velocity and capital efficiency norms.
The expanded pact covers Blackwell Ultra, Rubin and Rubin Ultra GPUs through 2028, adds Vera CPU support, extends NVLink Fusion to Amazon Trainium, and builds U.S. government AI factories around 100,000 GPUs.
A split strategy of separate training and inference variants replacing Ironwood, with the inference chip tripling vector memory to 384MB for larger KV caches and context lengths.
Fully autonomous retrofitted excavators are working live construction sites in Texas and Nevada, with the $350M-backed ex-Waymo team moving toward coordinated fleets.
TechCrunch reports physical-AI model makers face a 'robotics data crisis' and are moving beyond end-to-end training, as AV players launch humanoid labs and tooling matures.
Vercel shipped fx, a small open-source coding agent for developers, among Product Hunt's top launches this week.
A Rust TUI that carries agent sessions across tools, shown on Hacker News.
Analyzes coding-agent sessions to surface token waste and cut agent spend.
Local desktop-based AI assistants, shown on Hacker News.
A ~$722M valuation Series C for the South Korean consumer AI portal, with Coreline Ventures and Eugene Asset Management joining.
Perceptron, founded by ex-Meta FAIR researchers, raises a Bessemer-led round for frontier vision models after open-sourcing its Isaac 0.5 model.
A Series A co-led by Susquehanna Venture Capital and Nexus Venture Partners for general-purpose AI agents out of Bengaluru.
A Peak XV-led Series A extension for enterprise voice AI agents across voice, WhatsApp and browser, bringing the round to ~$15.5M.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
