AI Builders
Digest

Anthropic Research Alex Albert announces Claude Mythos Preview, released just two months after Claude Opus 4.6. The preview is available to launch partners in Project Glasswing, which he describes as possibly the most consequential event in the AI industry he has seen up close since joining almost three years ago. This feels like a turning point in history with meaningful double-digit gains on critical benchmarks.

OpenAI CEO Sam Altman celebrates hitting 3 million weekly Codex users by resetting usage limits. The company will repeat this every million users up to 10 million. Happy building!

Box CEO Aaron Levie observes that Anthropic’s Mythos is another clear reminder there is no wall in model capability progress, with meaningful double-digit gains continuing from all frontier labs. As coding and tool use advance, agentic workflows will open entirely new use cases in finance, healthcare, legal, consulting, and supply chains. He shares how Box Agent autonomously fills RFP responses from a knowledge base in the background while the user does other work.

OpenAI VP Science Kevin Weil highlights a Paper Review skill built in a few hours with Codex inside Prism. It functions as a careful technical reviewer checking math, derivations, notation, units, structure, claim support, and cross-section consistency rather than just grammar. He is hiring a lead designer for Prism to build AI-native products that accelerate science.

General Partner Aditya Agarwal calls the new Mythos model absurd and says what an incredible time we live in, whether investing or building. He reflects that we are finally achieving the 2009 vision of making all software read/write and malleable: any software can now be forked, changed, remixed, and customized.

The Takeaway: OpenAI Frontier’s Ryan Lopopolo proves a small team can build and maintain a million-line Electron app generating a billion tokens daily with zero human-written code or review by inventing harness engineering. Ryan Lopopolo, leading frontier product exploration at OpenAI’s enterprise agent platform with prior roles at Snowflake, Brex, Stripe, and Citadel, imposed the constraint of never writing code himself. This forced creation of scaffolds, skills, spec.md files, quality trackers, and the Elixir-based Symphony orchestrator, achieving 10x productivity while shifting humans to post-merge reviews and systems-level thinking. "You have to step back. Right? Like, you need to take a systems thinking mindset to things and constantly be asking, where is the agent making mistakes?" The approach extends to OpenAI Frontier for safe enterprise agent deployment at scale.

Claude Code + Cowork at Anthropic Cat Wu shares that running /powerup in the Claude Code CLI reveals the team’s favorite features.

OpenClaw contributor Peter Steinberger releases CodexBar 0.20 with new providers Perplexity and OpenCode Go, account switching without re-login, fixed Claude token inflation from duplicates, and merged cost history. It now tracks 16 providers in one menu bar.

Cursor Design Ryo Lu showcases the new flow: no extra buttons, just click, draw, chat, fully in flow.

Builder Zara Zhang introduces the Personalized Podcast skill: turn any content into a podcast with two AI hosts, publish as RSS feed, and listen in your favorite app. She has been remixing meeting transcripts where the AIs eavesdrop and comment on their impression of her.

Y Combinator CEO Garry Tan notes Claude Code allows bash so the agent can cat >> to any file. He observes we live in LISPy times where code is data and data is code again, but the weird thing is it’s Markdown.

Every CEO Dan Shipper advises to be a model manager. He shares a visualization of OpenClaw dreams and notes he is almost there after a wild ride.

Claude Code at Anthropic Thariq, after ten calls and reviewing more transcripts, learns it is very easy to spend lots of tokens on open-ended verification that does not improve output. He plans to write more on doing it efficiently and is hosting a technical writing workshop in SF in two weeks.

Vercel CEO Guillermo Rauch speaks at Y Combinator and says he is more bullish than ever on exceptional founders, calling it the best city, best time, and best opportunity to build in generations.

FPV Ventures partner Nikunj Kothari observes that early-stage founders have started optimizing for views and funding instead of product and retention, with change logs showing barely any updates. He advises focusing energy on customers who will actually pay, as VCs will follow. He also highlights the Varanasi brothers as Silicon Valley’s greatest secret.

Linear Head of Product Nan Yu notes that many designers and engineers excel at thinking through product questions in the abstract, but tend to dive straight into coding or designing the moment you give them an IDE or Figma. More designers should become PMs because they would be good at it.

AI builder Swyx shares Simonw on the self-evident naming difference between “prompt injection” and “lethal trifecta,” points to new Amazon AI features he plans to try, and gives due credit to Alex in the Latent Space podcast discussion.

Product at Roblox Peter Yang calls the mythos initiative a good move and wonders if Anthropic has been using it internally to achieve their recent insane velocity.

Product at Roblox Peter Yang shares that the run was so nice he had to go around the entire lake.