Alex Albert (Anthropic) publishes internal data on how much of Claude's development is already done by Claude: over 80% of all code merged is written by Claude. It's been months since many researchers hand-wrote code. The typical engineer ships 8x as much code as in 2024.

On the most open-ended engineering tasks, Claude's success rate jumped from ~26% to 76% in 6 months. When research sessions went off-track, Claude proposed a better next step than the human 64% of the time. 'We're not at recursive self-improvement yet, but it could come sooner than most expect.' 2294 likes.

Levie highlights a key paragraph from Anthropic's post: 'There has been an explosion of new ideas, initiatives, tools, and simulations as a result of employees working with highly capable models—far more than we have the capacity to pursue.'

His interpretation: AI dramatically lowers the barrier to doing more. We have far more ideas than we can pursue, and for the ones we want to pursue, we're ultimately limited by the surrounding work to execute. There's almost no amount of AI progress where that goes away. AI lets us build more software, launch more campaigns, research more drugs—all still requiring people to manage. 195 likes.

Sam Altman announces: build and publish web apps with ChatGPT. His personal note: 'I really wish I had this when I was a kid, but I do miss HyperCard.' 1997 likes.

This is a significant productization milestone—ChatGPT is now a full app building and publishing platform, not just a chat interface.

Sam Altman also announces a big upgrade to ChatGPT memory rolling out today. 3765 likes—the second highest engagement of the day.

Thibault Sottiaux announces the Codex Python SDK: you can now use Codex within your own programs. pip install openai-codex. Built by Ah20im and friends. 1164 likes.

This means Codex is now embeddable as a library—not just a standalone tool, but a programmatic agent inside any Python application.

Peter Yang spent a whole day setting up integrations and skills in Codex for his top creator workflows. His conclusion: you can save at least 50% of your time on any type of knowledge work if you set up the system upfront.

His 3-step method: 1) Reflect on your past week—what took the most time, what was most repetitive. 2) List every single step of that workflow in detail. 3) Paste it into Codex/Claude Code and ask 'What integrations and skills can I build to streamline this?'

He notes all his workflows have human checkpoints along the way so he can apply his 'taste.' 98 likes.

Nikunj Kothari built 'Nock'—an AI skill that replicates his VC evaluation style. He pulled 200+ 1:1 founder pitch meeting notes captured by Granola, distilled down to ~53 high-debate meetings, added his essays about founders he loves, and used Claude Code to build a skill with principles and a question bank grounded in real conversations.

He refined it over 5-10 actual decks, comparing the skill's output against real conversations. Now founders can try it at nikunjk.com/nock, and VCs can build their own at nikunjk.com/buildnock. 60 likes.

Dan Shipper launches Spiral 4.0—a writing partner for you and your agent. Key features: a new Style Engine based on stylometry to extract your brand's voice from past work, and MCP/CLI support so agents like Codex, Claude Code, and OpenClaw can use it automatically.

Used internally at Every (30-person team) to write landing pages, tweets, podcasts, marketing emails—all on-brand. 222 likes.

Matt Turck interviews Dan Roberts, Lead of the Foundations of Reinforcement Learning team at OpenAI: Why AI Can Now Make Discoveries. Topics include the AI math moment (Erdős, OpenAI, DeepMind, Anthropic all hitting milestones), why the OpenAI result was an act of exploration, informal reasoning vs formal proof approaches, and why RL may now be 'the cake, not the cherry on top.'

Other highlights: why RL started working with large language models, test-time compute explained, verifiable rewards in math and coding vs the messy real world, and whether AI is already doing original science.