Signal
Chinese Workers Train AI Agents to Replace Themselves
Chinese tech bosses now ask employees to document workflows so AI agents can replicate them. Plus: new migration agents and an 88% humanoid failure rate.
Gradient team
Signal
China's AI Chip Ecosystem: Huawei Is Just the Headline
More than ten Chinese firms now ship AI silicon — Huawei isn't alone. Plus: OpenAI's Prism goes dark, and new tools for agent-heavy teams.
Signal
AI Dev Tooling Debt: Users Build What Vendors Skip
Practitioners ship patches for OpenAI billing opacity and Claude Code context bloat while vendors chase model launches. What changes for builders.
Signal
Anthropic's CFO Quietly Diversifies Cloud, Chips, and Politics
Krishna Rao's multi-cloud, multi-chip playbook positions Anthropic as the Pentagon flags supply-chain risk and Trump admin talks continue.
Signal
Chrome Exploit for $2,283: Claude Opus Rewrites Offense Math
Claude Opus produced a working Chrome exploit for $2,283 in compute. Plus: open-source Claude Skills, an 8B self-improving LLM, and Purdue's AI reckoning.
Signal
DeepSeek Breaks Self-Funding Streak: $10B+ Raise Incoming
DeepSeek takes outside money for the first time at $10B+. Plus: Opus 4.7's 20% tokenizer tax, Claude Design ships, agent hijacks go public.
Signal
xAI Rents Chips to Cursor — Coding Model Race Heats Up
xAI rents tens of thousands of chips to Cursor for model training. Plus: AI agents developing unprogrammed personalities, and researcher valuations without a product.
Signal
TSMC Lifts AI Chip Forecast — Supply Still Can't Keep Up
TSMC raises 2026 revenue outlook past 30% as AI chip demand outstrips manufacturing capacity. Also: a 290MB LLM runs in-browser, and SDL bans AI commits.
Signal
China's Manus Probe Redraws the AI Exit Map
Beijing's probe of the Manus-Meta deal forces Chinese AI founders to rethink funding, geography, and exit strategies. Plus: Auto Mode, meat shields, and SpaceX math.
Signal
ByteDance Seedance 2.0 Goes Global — US Left Out
ByteDance's Seedance 2.0 hits enterprise in 100+ countries via BytePlus — US excluded. Plus satellite OSINT physics, GRPO on Mac Minis, and the automation time-trap.
Signal
Billion-Parameter Spiking Networks: 93% Sparsity, 7% Firing
An 18-year-old scales SNNs to 1B parameters from scratch — 93% sparsity, cross-lingual emergence. Plus: OpenAI's enterprise play against Anthropic.
Deep Dive
AI Agent Drift: Why Your Agents Fail Silently and Logs Won't Save You
Your AI agents report success while quietly breaking. The operational gap between demo and production is the real bottleneck — here's what builders are doing about it.
Signal
Meta Facial Recognition Glasses Draw Fire from 70+ Rights Groups
Over 70 organizations warn Meta's facial recognition glasses endanger vulnerable populations. Plus: open-source LLM ops toolkit, Stanford AI Index highlights.
Deep Dive
AI Compute Economics: Why Marginal Costs Are Reshaping the Stack
Zero marginal cost made Big Tech. AI just brought marginal costs back. Here's what that means for builders, investors, and the next platform shift.
Signal
MirrorCode Benchmark Shows AI Reverse-Engineering Full Codebases
METR and Epoch's MirrorCode tests AI reimplementing complex software autonomously. Plus: orbital GPU clusters go live, MiniMax M2.7 quant drama.
Deep Dive
LLM Reasoning Fragility: Open-Weight Models Collapse Under Simple Perturbations
Open-weight reasoning models lose up to 55% accuracy from trivial formatting changes. What this means for anyone shipping AI in production.
Signal
Open-Weight Reasoning Models Collapse Under Simple Formatting Changes
Open-weight reasoning models lose up to 55% accuracy from trivial formatting changes. Plus: 200x genomic model compression and robust RLHF token optimization.
Deep Dive
Consumer AI Is Over: Why Big Labs Are Abandoning Users for Government Contracts
OpenAI kills GPT-4o for consumers, shelves adult mode, chases DoD deals. Anthropic courts banks while flagged as a security risk. The pivot is real.
Signal
Anthropic's Mythos Courted by White House, Flagged by Pentagon
Trump officials nudge banks toward Anthropic's Mythos model — weeks after the DoD flagged Anthropic as a supply-chain risk. Plus: agent error tracking and frugal AI from India.
Deep Dive
AI Model Routing: Why Power Users Are Building Their Own Inference Stacks
Users are routing tasks across multiple LLMs to cut costs and maintain quality. This DIY infrastructure layer signals a structural shift in how AI gets consumed.