We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Top Stories
Why it matters: AI’s control surface is widening—from opaque reasoning traces to mass-market assistants—and the new failures are increasingly about access and reliability, not just benchmark scores.
Hidden reasoning is no longer reliably hidden. A research team says it extracted frontier-model reasoning through a vulnerability in every frontier AI company’s APIs; for most queried prompts, extracted token counts matched billed thinking tokens 1:1. Its report says the team decoded many publicly posted JSON traces and found private information, while seeing Anthropic and OpenAI traces at scale exposed “casual” misalignment that makes monitoring harder. The method is described as lossy and the observations are based largely on older models, so this is a major disclosure—not proof that all current reasoning is recoverable.
Gemini has reached consumer-platform scale. Google CEO Sundar Pichai says the Gemini app now has more than 1 billion monthly users, making it Google’s fastest-growing product and its 14th to reach that mark. Google’s accompanying voice figures say 63% of users talk directly to Gemini and that more than 60 regional dialects are rolling out.
Research & Innovation
Why it matters: The strongest technical signals pair models with disciplined experiment design and expose where agents fail on real documents.
Model Discovery Agent (MDA) makes agents choose experiments, not just hypotheses. Its paper describes an LLM proposer coupled to sequential Monte Carlo, simulation-based inference, and value-of-information experiment selection, including an “M-open” loop that expands the hypothesis class when the current one fails; the abstract reports new state-of-the-art results across physics, chemistry, and biology. The author’s evaluation thread reports exact force-law recovery in 74% of runs versus 31% for a budget-matched LLM agent, using roughly five times fewer experiments; on enzyme kinetics, it reports about 56% symbolic accuracy in eight experiments versus about 42% for prior SOTA after 60.
ExtractBench exposes a long-document reliability trap. LlamaIndex tested 14 VLMs, coding agents, and extraction APIs on 370 enterprise documents spanning 4,869 pages and 67 types. Past 50 pages, commercial VLMs fell below 35% recall because silent list truncation dropped table rows, even while precision remained high.
Products & Launches
Why it matters: Vendors are packaging AI as persistent execution—fast local models and delegated credentials—rather than one-shot chat.
NVIDIA’s Nemotron 3.5 Lightning targets the execution layer. NVIDIA launched an open 30B MoE model with 3B active parameters for always-on specialized agents and claims up to 4× the output speed of similar models. Artificial Analysis reports an Intelligence Index score of 24, GDPval-AA Elo of 824, Terminal-Bench performance of 24% versus 7% for Nemotron 3 Nano, and nearly 670 output tokens per second on a pre-release endpoint. It is already available locally through Ollama for coding, tool-calling, and multi-turn agents.
Grok Bot turns credentials into the product. xAI’s early-beta bots sign into users’ tools and return finished work. A launch observer describes persistent cloud computers logged into Gmail, Salesforce, LinkedIn, and other sites, reusable workflows, and bots that coordinate in parallel. Delegated access and agent-to-agent handoffs are now part of the user experience, not merely backend plumbing.
Industry Moves
Why it matters: Capital and deployment are moving toward owned AI stacks and internal agent infrastructure.
River AI raised $1.1 billion, led by General Catalyst and AMP with strategic investment from Nvidia and AMD. The company says its training API is live and its River API lets users build custom agents and LLMs on open-weight models, as part of a stack whose hardware, data, and intelligence users would own.
DoorDash built Flux, an internal cloud platform that automated 130,000 engineering tasks in one month and now powers more than 25,000 code reviews weekly while retaining guardrails over execution and access to DoorDash systems.
OpenAI COO Brad Lightcap is leaving after eight years to start something new. In his memo, he says he built much of OpenAI’s operating infrastructure and is focused on “important new things” the world must get right as AI enters its next period.
Quick Takes
Why it matters: Smaller signals show practical progress coming from persistence, cautious tool use, clinical evaluation, and verification.
- Qwen 3.8 Max: ValsAI says Legal Research all-pass rose from 25.5% to 47.6%, mainly through persistence: task turns rose 19.7→35.5 and time 809→3,678 seconds.
- AMIE: Google Research reports real-time video consultations and expert-level performance in a randomized controlled trial with 300 simulated consultations.
- Attestable: The startup emerged from stealth to use zero-knowledge proofs to verify the model, inputs, and tool calls; its investor says it reduced the former roughly 1,000,000× compute overhead by many orders of magnitude.