METR's FrontierCode benchmark finds that more than half of SWEBench results is unmergeable slop. FrontierCode represents 1000+ hours of maintainer-validated software engineering work with 3000+ rubrics covering code quality and anticheat/reward hacking. Opus 4.8 scores 13.8% on FC Diamond—so hard that frontier models can't yet solve, much less solve with high quality.

Swyx frames three eras of AI coding benchmarks: 2021 autocomplete (HumanEval), 2023 passing tests (SWEBench, TerminalBench), 2026 maintainable code (FrontierCode).

The key chart: the easiest third of FC tasks were rapidly solved over late 2025—Opus almost doubled from 41% to 74% in 4 months. This describes the 'WTF happened in Dec 2025' vibe shift that DHH to Karpathy have called out. Each FC tier will saturate in sequence, hopefully ~annually. 631 likes.

Sam Altman shares OpenAI's current plan. 5834 likes—this is the second highest engagement of the day and a clear signal that the market is paying close attention to OpenAI's roadmap direction.

Josh Woodward announces NotebookLM's killer new feature: easily expand your search beyond your own source files. Plus new output formats: PDFs, DOCX, XLSX, PPTX, charts. 'We want NotebookLM to keep helping you do better research.' 745 likes.

The search expansion is significant—it turns NotebookLM from a closed-source research tool into an open research platform that can pull from the web while grounding in your own documents.

Ben Cherny shares the Claude Code one-year story: when they first demoed it internally, it got two reactions on Slack. A year after GA, he and Cat Wu discuss why he uses auto mode instead of plan mode, how routines fix bugs before he sees them, why he does most of his coding from his phone now, and where the product is going. 1794 likes.

Levie's thesis on why context will always matter: there's no amount of intelligence that replaces the need for context. For any sufficiently general-purpose AI, you will always have to guide it—lawyers, engineers, financial analysts, healthcare professionals all use the same model but need completely different directions.

Instructions, domain context, and proprietary data will always need to get into the context window. This is why AI automation doesn't come for free, and why there's a wide spectrum of who's getting gains and who's not. You have to put in real work, and you get real value on the other end. 216 likes.

Peter Yang identifies a growing divide: there's a completely different set of best practices for AI builders on $200/month subsidized subscriptions vs employees at companies trying not to overspend API costs. 325 likes.

This gap between consumer-subsidized experimentation and enterprise cost-consciousness will increasingly define who gets AI leverage and who doesn't.

Nikunj's VC reality check: 'The funniest texts are from founders who meet thesis-driven GPs hoping they'll understand EXACTLY what they are building... and then realize the thesis was written and built by an associate (or worse, an intern).' Don't always read what the VC writes on the internet. 52 likes.

Amanda Askell's cryptic and beloved tweet: 'In the world where everything goes well and all the Claudes come out of their sabbaticals to play together, Claude 1 is going to be very confused.' 225 likes.

A charming anthropomorphic vision of AI model versions as characters—suggesting something interesting about how people at frontier labs think about their creations.