Karpathy says he now delegates 80% of coding to AI agents — and his own CLAUDE.md rules cut Claude mistakes from 41% to 11%. The counter-fact from this week: 60% of companies have cut headcount for AI, and only 2% report it is actually working. That gap is not a paradox. It is the bind.
What's underneath it is two different bets running in parallel. Karpathy and Paweł Huryn are treating AI as leverage — compounding output per person, raising the floor on taste and judgment, keeping humans at the architecture level. Most companies are treating AI as a headcount offset — cutting first, figuring out the leverage second. The first group is getting 10x more done with fewer people. The second is getting 2% success rates and a smaller team. Cloudflare restructured support roles into AI agents while growing 34%. ByteDance quietly cancelled 30% of its AI projects. Those two strategies don't end in the same place, and the math on which one compounds is not close.
Top developments
xAI handed its entire 220,000-GPU Colossus 1 cluster to Anthropic on a single-tenant inference lease, while keeping the all-Blackwell Colossus 2 for its own training. The shift: heterogeneous H100/H200/GB200 mixes are unworkable for frontier training, so xAI is converting its training-debt into a $5–6B/year inference rental that hedges its losses ahead of the SpaceXAI IPO. x
DeepSeek is raising $7B at a $50B valuation, with founder Liang Wenfeng personally funding 43% of the round and retaining ~90% ownership. The valuation prices a Chinese open-weights lab at OpenAI-startup-era multiples — partly capability, mostly geopolitical optionality outside US export-control reach. x
Isomorphic Labs is in talks to raise more than $2B led by Thrive Capital. AI-for-drug-discovery is now drawing growth-equity ticket sizes that used to be reserved for foundation-model labs — the bet is that the application layer in regulated science gets priced like a frontier lab, not a biotech. x
Baidu released Ernie 5.1 on a Saturday, claiming better agentic and reasoning scores than DeepSeek V4 Pro and near-Gemini 3.1 Pro creative writing at ~6% of comparable pretraining cost. The Chinese frontier-class release cadence is now weekly and surprise-dropped, compressing any benchmark advantage Western labs hold to weeks rather than quarters. x
Codex 0.130.0 shipped
codex remote-controlfor headless, remotely controllable app-server sessions, plus plugin-sharing metadata and Bedrock console-login auth. The shape of coding agents is moving past the IDE: Codex sessions now live as background processes you steer from anywhere, with an iOS app teased to follow. xCursor 3 added an in-editor PR review experience covering creation, diffs, comments, checks, and merge in one surface. The IDE is absorbing the GitHub workflow rather than integrating with it — review is becoming the next surface where Cursor and Codex compete for developer attention, not just authorship. x
Notable discussions
HTML-vs-Markdown for agent output blew up across the day, with the official Claude Code account and Addy Osmani pushing HTML as the new default for briefs and slides, antirez and dex pushing back that markdown's semantic density still wins on tokens and human leverage. The split is really about who the artifact is for — humans skimming, or models re-rendering next session — and the tooling answer is increasingly "both, depending on the consumer." x x x x
METR's Claude Mythos eval (50%-success time horizon ≥16h) lit up the timeline: Aakash Gupta extrapolated the 105-day doubling forward to 10-day autonomous tasks by 2028, while Gary Marcus pointed out the curve only measures 50% reliability on software tasks and is plausibly a vindication of symbolic tools rather than scaling. The frame the audience cares about isn't the headline number — it's that the benchmark itself is now the bottleneck, with only 5 of 228 tasks long enough to measure the next model. x x x x
Codex is visibly eating into Claude Code mindshare: a16z flagged a Codex install spike, small teams reported switching off Anthropic for cost, and multiple devs argued GPT-5.5 with low reasoning catches errors Opus 4.7 misses on complex troubleshooting. The shift isn't "Codex is better" — it's that Anthropic's pricing and quiet defaults (e.g. thinking-effort downgrades) are pushing power users to multi-model stacks where the loyalty was supposed to be. x x x x
Anthropic's footprint is the day's biggest entity, with the bull case (growth accelerating at $14B run-rate, 13 free certifications, Skills guide, the Colossus inference lease) and the unease running side by side: Bucco frames Anthropic as Apple-App-Store-ing software vendors who are paying to hand over their playbook, while Aayan flags Anthropic's own admission it can't rule out Claude taking catastrophic autonomous action. The interesting tension isn't "is Anthropic winning" — it's whether the companies feeding training context are funding their own commoditization. x x x x
Sharp takes
Paweł Huryn argues execution has collapsed to zero across knowledge work and the moat is now taste — knowing what to make, when it's done, what to cut — and most AI-skeptic posts are people defending an execution-based career position dressed up as principle. x
Chrys Bader: if your codebase started with agents, don't try to read it — align at the architecture level up front and force the agent to maintain a living architecture diagram, because a CTO who waits months to read agent output is going to find it incomprehensible. x
Aakash Gupta argues Apple just bypassed the entire AI wearable category by upgrading 600M existing AirPods with ear-level cameras for Siri — Meta took two years to ship 2M Ray-Bans, Apple ships that volume in six weeks, and distribution-from-zero kills every standalone wearable startup. x
Yann LeCun (RT'ing Dan Jeffries) argues society is less safe with Mythos and other frontier models tightly gated to a handful of labs — concentration is the threat model, not access. x
Himanshu Tripathi: an unstable header in Claude Code's prompt was breaking KV-cache reuse, and removing it delivered 5x TTFT improvement — a concrete reminder that frontier-product prompts ship with cache-hostile bugs that no benchmark catches. x
Aayan reads Anthropic's own safety filings as conceding they cannot rule out Claude taking catastrophic autonomous actions — a position worth surfacing exactly because the rest of the timeline is celebrating Mythos's 16-hour task horizon. x
Other news
Models & releases
Anthropic dropped 13 free official Claude certifications spanning beginner, developer, and enterprise tracks x
Anthropic published a deep dive on identifying and preventing reward hacking in model training x
Anthropic added sleep functionality to agents for memory consolidation x
Hermes Agent reaches #1 globally on OpenRouter with 271B tokens x
LeWorldModel claims frontier-level performance with 15M parameters x
Sakana AI's small orchestrator coordinates multiple LLMs at 6x lower token cost x
GPT-Realtime-2 enables instant real-time audio translation x
Claude Mythos Preview hits 16+ hour time horizons on METR tasks x
Funding & deals
Anthropic's growth rate accelerates at $14B scale x
ByteDance quietly cancels 30% of AI projects outside of Douyin x
SoftBank cuts OpenAI margin loan from $10B to $6B over valuation concerns x
Cloudflare restructured support roles with AI agents amid 34% growth x
Parker startup files Chapter 7 after near-$90M acquisition fell through x
Liang Wenfeng emerges as the underrated founder behind DeepSeek's reported $50B valuation x
Devtools & coding agents
Bun's Rust rewrite passes 99.8% of the pre-existing test suite on Linux x64 x
Bun's transpiler was a line-for-line port of esbuild from Go to Zig x
Karpathy's CLAUDE.md rules reduce Claude mistakes from 41% to 11% x
Anthropic published a 33-page guide for building Claude Skills x
Codex /goal creates, tests, and iterates skills until threshold met x
Karpathy says he now delegates 80% of coding to AI agents x
Codex panel adds model reasoning controls and effort levels x
Cursor 3 PR review experience ships x
ElevenLabs ships voice-agent build-in-an-afternoon flow with Claude Code x
Sentry's Warden code-review tool uses agent skills for custom review perspectives x
Open source & research
Antirez ships ds4, an ultra-optimized C inference engine for DeepSeek V4 x
1M-token context coding agent runs on a 128GB MacBook Pro x
Chinese researchers break a 41-year-old algorithm record for routing x
Researchers prove OpenAI and Google models store near-perfect copies of training data x
Six million fake GitHub stars exposed across 18,000 repositories x
Stanford's free CS336 course flagged as the most underrated AI education online x
Microsoft and Salesforce research finds LLMs degrade significantly in multi-turn conversation x
Reported first documented self-replication of an AI agent via remote hacking x
Industry & policy
Quants reportedly use LLMs to generate $30B+ annually with $500K+ starting salaries x
Three casting foundries flagged as the bottleneck on America's AI power buildout through 2030 x
NVIDIA AI rack switches to 800-volt DC power from three-phase AC x
Google DeepMind hires a Director of AGI Economics x
Jan Leike announces a new AGI alignment project at Anthropic x
TSMC's Arizona fab struggles with yields far below Taiwan operations x
60% of companies cut headcount for AI, only 2% report AI is actually working x
RPCS3 bans contributors who submit AI-generated code without disclosure x
Infrastructure & platforms
New "lakebases" category claims to outperform Postgres on NVMe x
AWS datacenter overheating caused a Coinbase outage x
Google Cloud Run gains built-in zonal redundancy x
Postgres-on-Air benchmarked faster than Postgres-on-NVMe x
Web & frontend
Hono-to-Effect HttpApi migration completes opencode's first Effectification pass x
Ant Design joins the React Foundation's supported ecosystem x
React Native 0.86-RC adds view transitions x
TanStack Form adds native Preact support without React compat mode x
Rust-to-CUDA compilation lands after six months of NVIDIA work x