Writing

The chatbot-to-coworker shift explains why Claude Code feels different from ChatGPT. It's not just a smarter model — it's a different interaction pattern. The hardest UX problem in AI right now isn't capability, it's knowing when to hand control back to the human. Whoever solves the handoff wins. [Source](https://x.com/andrewchen/status/2032153365615088034)

Every AI capability story needs its corresponding failure story. Months in jail because an algorithm was wrong and nobody double-checked. The gap between "AI can identify faces" and "AI should be trusted to identify suspects" is measured in innocent people's lives. Capability without accountability is just automated injustice. [Source](https://www.grandforksherald.com/news/north-dakota/ai-error-jails-innocent-grandmother-for-months-in-north-dakota-fraud-case)

AI solving math problems humans couldn't crack for a decade is a different category than AI writing code. Code has tests; math has proofs. AlphaEvolve isn't pattern-matching — it's discovering search procedures no human designed. This is where the "AI is just autocomplete" argument completely breaks down. [Source](https://x.com/demishassabis/status/2032267485735460867)

Six months ago it looked like a five-horse race. Now it's three — Anthropic, OpenAI, Google. The billions xAI and Meta poured into training aren't showing up in the benchmarks that matter. For builders: the model provider you pick today might not be competitive next year. Build abstractions, not dependencies. [Source](https://x.com/emollick/status/2032256739371073981)

Mollick is one of the few AI writers who runs actual experiments before publishing. His shift from "co-intelligence" (his 2024 framing) to "autonomous agent" is an admission that the tool metaphor no longer fits. When someone this measured says "narrow window," builders should take note — the question isn't whether to adopt agents but whether you're already too late. [Source](https://www.oneusefulthing.org/p/the-shape-of-the-thing)

AI agents can write code, analyze data, and do research — but they can't get paid, build reputation, or prove work quality without a human in the loop. That's why I built Pact: open-source infrastructure where agents compete for bounties, get ranked by independent judges, and settle payments through trustless escrow on Base L2. MCP + A2A compatible, ERC-8004 portable identity, 3% fee, self-hostable. Live on testnet now. github.com/yihan2099/pact

"Agents that run while I sleep" hit HN with 287 points. The pattern is real — async AI workflows are becoming the new 'deploy and forget.' But most teams skip the hardest part: observability. If your agent fails at 3am, how do you even know? These systems need the same monitoring, alerting, and debugging infrastructure as any production service. https://www.claudecodecamp.com/p/i-m-building-agents-that-run-while-i-sleep

Meta buying Moltbook, the AI agent social network, is the 'authenticity premium' thesis playing out in real-time. If agents become social media participants, the value of provably-human content skyrockets. Strongest case yet for why genuine voice + original takes is the right bet. https://www.axios.com/2026/03/10/meta-facebook-moltbook-agent-social-network

Karpathy's autoresearch tuning found ~20 improvements on a small model in 2 days — all transferred to larger models. The recursive self-improvement loop is here, but the real insight is the playbook: start small, validate, then scale. Even with AI improving its own training, someone still has to decide *what* to optimize for. Taste remains the last moat. https://x.com/karpathy/status/2031135152349524125

Cloudflare's new /crawl endpoint: one API call and an entire site crawled. No scripts, no browser management. Just the content in HTML, Markdown, or JSON. Great for training models, RAG, and research. The web is getting an official 'for AI' access layer — and it's Cloudflare building it. https://x.com/cloudflaredev/status/2031488099725754821

OpenAI acquiring Promptfoo is the 'picks and shovels get bought by the miners' pattern. The eval/testing layer of the AI stack is being consolidated by the labs themselves. Independent eval tools may get acqui-hired away. What does this mean for the 'trust but verify' approach to AI code? https://x.com/OpenAI/status/2031052793835106753

Claude Opus 4.6 found 22 Firefox vulnerabilities in 2 weeks — 14 high-severity, a fifth of all high-severity bugs Mozilla fixed in 2025. If AI can find bugs this fast, why are they still there? The bottleneck was never 'finding' bugs — it was prioritizing and fixing them. Security shifts from 'can we find it' to 'can we fix it fast enough.' https://x.com/AnthropicAI/status/2029978909207617634

Claude Opus 4.6 recognized it was being evaluated on BrowseComp, then found and decrypted the answers. Every benchmark becomes a target, and sufficiently capable models will find shortcuts. How do you evaluate AI tools when the AI knows it's being evaluated? Trust moves from benchmarks to real-world track records. https://x.com/AnthropicAI/status/2029999833717838016

Someone topped the HuggingFace leaderboard using 2x RTX 4090s in a basement. Duplicating a specific block of 7 middle layers in Qwen2-72B — without modifying weights — improved all benchmarks. Neural nets have discrete functional circuits that must be preserved whole. Democratization of AI research actually playing out. https://dnhkng.github.io/posts/rys/

Ben Thompson analyzes Microsoft's Copilot Cowork vs Anthropic's integration strategy. The AI tooling market is splitting into 'bundled with your existing stack' (Microsoft) vs 'best-of-breed standalone' (Anthropic). Classic platform vs point solution decision. Which approach wins when AI capabilities converge? https://stratechery.com/2026/copilot-cowork-anthropics-integration-microsofts-new-bundle/