Anthropic shipped computer use in Claude Code — its fourth iteration in 17 months, and the first that fundamentally changes how developers work.
Claude can now open apps, click through UIs, and test what it built, all from the CLI. The upstream enabler is 17 months of iteration: from the October 2024 Docker-container API beta through three progressively less clunky versions to this one, which runs natively where developers already live.
Claude Code's auto mode expanded to Enterprise plans and API users — letting the model make permission decisions on your behalf.
Instead of approving every file write and bash command, auto mode lets Claude exercise judgment about what's safe to execute. The move positions Anthropic to compete directly with Codex's cloud-first autonomous workflow.
Lydia Hallie acknowledged that Claude Code users are hitting usage limits far faster than expected — the team is actively investigating.
The tweet hit 4.4K likes and 281K views, suggesting the rate-limit pain is widespread and the acknowledgment was cathartic for a frustrated user base. Corey Quinn's sardonic workaround — a cron job that burns the 5-hour session window overnight — captured the mood perfectly.
Computer use just killed an entire class of human labor — any work that existed only because legacy software had no API.
Government portals, enterprise dashboards, anything locked behind a login is now scriptable for $20/month. The framing is provocative but directionally correct: computer use turns every GUI into a programmable surface.
OpenAI launched a Codex plugin that runs inside Claude Code — a deliberate Trojan horse that puts OpenAI's cloud agent inside Anthropic's developer tool.
Rather than compete for the CLI, OpenAI is embedding Codex as a backend engine within the tool developers already prefer. The trending sidebar confirmed the magnitude: #1 news item with 1,069 posts.
New Codex usage data shows developers delegate their hardest tasks — refactors, architecture planning — to Codex at the end of the workday.
The chart shows a dramatic spike in task complexity in the final turn of each day, suggesting developers are using Codex as a night-shift engineer. The clearest signal yet that AI coding is shifting from pair-programming to asynchronous delegation.
A Qwen 3.5 27B model fine-tuned on Opus 4.6 distilled data has been #1 trending on Hugging Face for three weeks — and it beats Sonnet 4.5 on SWE-bench.
The model runs locally on 16GB in 4-bit quantization, meaning a laptop can now run an agent-grade reasoning model that outperforms a frontier commercial model on coding benchmarks. The pattern: open-source models drafting behind proprietary frontier research via distillation.
llama.cpp hit 100,000 stars on GitHub — a milestone for the engine that made local LLM execution practical.
Georgi Gerganov's project, which started as a weekend hack to run LLaMA on a MacBook, has become critical infrastructure for the entire open-source AI ecosystem. The 100K mark puts it alongside TensorFlow and PyTorch in rare company.
Trending · 73 posts
Brian Roemmele reports Qwen3.5-Omni will be fully open-sourced — calling the prospect "monumental."
If confirmed, this continues Alibaba's pattern of using open-source AI as a geopolitical and commercial differentiator against US-based closed-source providers.
Garry Tan's weekly report: 126 commits, 253K lines added, 115 AI coding sessions (113 Claude Code, 2 Gemini), 72-day shipping streak.
The Y Combinator president is running his own engineering org with AI as the staff. The ratio — 113 Claude Code sessions to 2 Gemini — is revealed-preference data no benchmark can match. Ship of the week: a CLI production RPC via HTTP tool API spanning 1,903 lines.
The most complete Claude Code setup in existence: 27 agents, 64 skills, 33 commands — all open source.
Built by the Anthropic hackathon winner and refined over 10 months, the repo is a professional-grade operating system for Claude Code that turns a coding assistant into an engineering department.
"Minimum viable prompting" — your Claude prompts are probably way too detailed, and it's making outputs worse.
Ole Lehmann cited Boris Cherny, creator of Claude Code, who keeps his own setup minimal. The advice runs counter to the instinct to over-specify, but aligns with how frontier models benefit from room to reason.
A Claude Code skill that clones any website from a single command hit 4,800 GitHub stars in 16 days.
The tool screenshots the site, extracts every design token, downloads all assets, and spawns a pixel-perfect replica. Web design as a craft is being compressed into a one-line command.
Marc Andreessen explains how AI triggered a three-way standoff: PMs, engineers, and designers each think AI lets them do the other two jobs.
The real disruption isn't job elimination but role-boundary collapse, where every function expands to absorb adjacent ones. The irony: all three may soon find AI can manage them too.
Hinton's 2016 prediction that radiologists would lose their jobs within 5 years was wrong — the number of radiologists has grown.
The post hit 1.2M views, suggesting deep appetite for AI-prediction accountability. The failed prediction is a synecdoche for a broader pattern: AI's impact on professions is far more nuanced than replacement narratives suggest.
ARC-AGI-3 humiliated top AI models — none scored above 1% on a benchmark designed to test genuine abstraction.
The result is a useful cold shower for a feed that otherwise reads like an AI capabilities victory lap. The benchmark was the #3 trending news item with 866 posts.
Trending · 7h · 866 posts
NASA's lead electrostatics scientist claims a "new force" that counteracts gravity — based on 2,000 vacuum chamber experiments.
Dr. Charles Buhler reports propellantless thrust that persists after power is switched off. Extraordinary claims, not yet peer-reviewed in a major journal — but the sheer number of controlled experiments is unusually high for fringe physics.
Turbo3 KV cache compression jumped from 4.57x to 5.12x with zero quality loss — plus 3-7% decode speed boost on M2 Pro.
Better KV cache compression means longer context windows at lower cost — the invisible infrastructure improvement that compounds into every agent workflow.
How Stefan Banach got tricked into earning his PhD — because he was "allergic to bureaucracy."
Colleagues told him gentlemen from Warsaw had interesting problems to discuss. The "discussion" turned out to be his doctoral defense. Fermat's Library noted he was discovered by a professor who overheard him discussing Lebesgue integrals in a Kraków park. A reminder that this feed carries a strong contemplative thread alongside the AI signal.
Turntable in Adobe Illustrator lets you rotate 2D vector art in 3D — no redrawing, just a slider drag.
For animators and game designers, this eliminates one of the most tedious manual steps in asset creation. Conceptually similar to how AI image models interpolate between viewpoints.
Atlanta's new autonomous robot dog patrols drew Black Mirror comparisons — Undaunted raised $1M to deploy "packs" on city streets.
The dogs patrol autonomously but are controlled remotely. The visceral reaction (182K views) suggests this touches the same nerve as Ring doorbell surveillance: people want safety but are uneasy about what patrols the boundary.
The IRS requires you to report income from illegal activities and stolen property — Schedule 1, Form 1040, line 8z.
The government's position is logically consistent (all income is taxable) but comedically absurd (please itemize your heist proceeds). Peak tax-season internet.