Mark Fayngersh

It's January 3rd and I think I've already peaked for 2026. Built this to speed up my own workflow, never imagined this. Thank you @rauchg for the shout-out! 🙏

@ericzakariasson my early experience with composer 2 actually points at ui work that is cleaner and better than opus 4.6 in many cases, which is exciting to see.

best coding models right now sync work, shorter iterations, 1-5 agents in parallel: - plan: composer 2 - work: composer 2 when you're in control, you simply only need composer 2. async work, longer iterations, deep work: - work: gpt 5.4 for async, you still want maximum intelligence as speed is less important with a model that can run for a really really long time ui & taste - opus 4.6 is still really strong here - composer 2 is a fast follow, but not as strong (yet) this space moves fast, so it will probably change soon!

We may overvalue AI that looks like us and miss the systems that actually change the game. markfayngersh.com/writing/by-our…

@dwarkesh_sp Wigner's "unreasonable effectiveness of mathematics" has puzzled physicists and philosophers for 65 years. As someone who works at the frontier of pure math, do you find the phenomenon genuinely mysterious, or does it have a clean explanation you find compelling?

What should I ask Terence Tao?

@peteyreplies the weak shall be eaten

claude code review ran for ~30 mins, says found some issues, but no comments on pr. /cc @trq212 @bcherny

@aidan_mclau my first impression is it's a very good model. clean and crisp responses that feel thoughtful. genuinely engaging to interact with, esp on deeper intellectual topics.

welcome 5.3 instant! was proud to help reduce hallucinations for questions where factuality matters most, it's 26.8% better (when searching) and 19.7% better (when not searching)

@flydotio big fan of sprites, any plans to support beefier machines or control over vcpus/memory?

i also noticed that using the "Commit" git action does not seem to auto-generate the message from the actual current git diff. if i introduce local changes outside of the convo, such as updating deps, it will appear to generate some variant of the last commit message, even if it doesn't apply to the current actual diff being committed.

@thsottiaux @embirico don't seem to be able to have control over the style of git commit messages and pr descriptions created from the Codex App Git actions, even if I add custom instructions 👀

@embirico @OpenAI x.com/rauchg/status/…

@embirico @OpenAI Built the initial version of this library in two days using Codex and Claude Code, with Codex doing most of the heavy lifting around the most complex business logic. svelte-check-rs.vercel.app

I’m looking for cool stories of building X in Y days with Codex. Similar to how @OpenAI built the Sora Android app in 28 days. (Blog link in thread.) If you have an interesting example, please share!

@trq212 Noticing some proxy/tunneling issues in Claude Code for Web when trying to install some dependencies 👀

@trq212 Is it possible to go back-and-forth within a single Slack Thread / Claude Code Web session? Looks like every @Claude spawns a new session so replies aren't possible from within Slack, yeah?

In case the Claude Code team (@bcherny @trq212) sees this 👀 I love the tool, my single biggest ask for the next model update is even less reward hacking please! Example: when reviewing commits, if a git diff gets truncated, Claude often just reviews the partial output instead of reading the full contents (despite the tool output file properly being created and returned to the agent). I've resorted to writing "IMPORTANT: Never review based on truncated output" in my docs and skills. Would love a model where that level of thoroughness is increasingly the default instinct 🙏

Excited to share what we've been building @a____t____g ! Introducing Autonomous — a superintelligent financial advisor at 0% advisory fees. We’re democratizing the tools and strategies used by the ultra-wealthy and making them available to everyone. becomeautonomous.com

@bcherny doesn't appear to apply the user's env (e.g. from ~/.zshrc) the same way like claude code in the cli does, so it fails to find binaries, like `bun` for example.

Try it and tell us what you think! Comes with automatic git worktrees out of the box

@dummdidumm_ i could be wrong, but i don't believe that svelte-fast-check project handles svelte diagnostics, which u can only do properly if you're working with an AST. in theory, my project can also replace the separate eslint plugin and provide something more bundled and performant.

@pheuter Wow! I hope AI reimplemented all the checks correctly 😁I wonder how much slower it would be if you didn't transform to TS and instead used all the existing js tools and only tsgo. Well I guess I can just compare since someone else had a very similar idea github.com/astralhpi/svel…

40 seconds → 1.3 seconds I built a 30x faster svelte-check in 2 days using AI. Now Codex and Claude Code can iterate on our production SvelteKit codebase without waiting forever to verify their changes. @dummdidumm_ would love your thoughts! svelte-check-rs.vercel.app

@dummdidumm_ yeah, i imagine there's something like a long tail of true parity in checks and functionality that can be chipped away with proper test coverage, off to a solid start imo: github.com/pheuter/svelte…

@mikethough @rauchg i've gone through many edits of AGENTS.md to try to steer the agents to avoid running `svelte-check` until the very end of a large changeset 😂

@rauchg @pheuter I was literally just trying to figure out how to configure opencode to stop doing full svelte check every time

A 10x-30x faster 𝚜𝚟𝚎𝚕𝚝𝚎-𝚌𝚑𝚎𝚌𝚔 in Rust just dropped, built with Claude Code and Codex CLI by @pheuter. The motivation was to make agents themselves more efficient. 2026 is about to be wild 😄 svelte-check-rs.vercel.app