John Thomas

@karpathy Congratulations! Amazing times!

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

I also found this an effective way to decide which model version to choose for which task, because you can see the tokens rack up much faster with Opus vs. Haiku, GPT 5.5 vs 5.4, etc.

Claude sub not working with OpenClaw, pushing us towards the API, got me way more token-aware than I expected. I wrote a tiny Mac status-bar widget, shows live ⚡ token-burn across Claude Code + Codex. npx llmeter install (free, local, mac) github.com/jawnty/llmeter Super interesting and insightful to see how my chats, prompts and model versions, spend tokens.

Really nice upgrade. I find I'm able to take a step back even more so while coding, specifying higher level direction, and just let it cook. Seriously good work from the team. As with prior model versions, the harder part for me is deciding how much of my development workflow I should undo/unlearn, so I can get the most out of 5.5

wow y'all love 5.5 we should think of something nice to do to celebrate!

Thanks Paul. My setup is SPEC.md → DESIGN.md → code, all in git. I switch between Claude Code and Codex, but honestly, the agent is interchangeable. The spec is the durable artifact here. What should come next: IDE treating the spec as a first-class object, diffable across branches, with "spec coverage" as a CI check, generating code and tests from the same source. I built some of this as skills in CC, and also exploring an idea to build a spec writer as a separate product surface, but all this being first class in an IDE would be neat. Most teams still treat code as truth and spec as docs. In a spec-driven world that's backwards. Intent should be the unit of work.

@jawnty @AndrewYNg @jetbrains Thanks, I'll look at it now. If I may ask: what SDD are you using? And: what do you think should come next in an idea SDD setup?

New course: Spec-Driven Development with Coding Agents, built in partnership with @jetbrains, and taught by @paulweveritt. Vibe coding is fast, but often produces code that doesn't match what you asked for. This short course teaches you spec-driven development: write a detailed spec defining what to build, and work with your coding agent to implement it. Many of the best developers already build this way. A spec lets you control large code changes with a few words, preserve context across agent sessions, and stay in control as your project grows in complexity. Skills you'll gain: - Write a detailed specification to define your mission, tech stack, and roadmap, giving your agent the context it needs from the start - Plan, implement, and validate features in iterative loops using a spec as your agent's guide - Apply the same repeatable workflow to both new and legacy codebases - Package your workflow into a portable agent skill that works across agents and IDEs Join and write specs that keep your coding agent on track! deeplearning.ai/short-courses/…

@meetgranola mcp support is a serious unlock in productivity. coupled with Claude Code, it's changed the way info is managed across my meetings in a way that's hard to believe, and now hard to live without.

Anyone else hitting issues with the recent Claude Code features like dispatch, remote control, etc. The team's been cooking, and I love seeing where the product is headed, but just wish there was a higher quality bar before launching.

@bengoodger Great analogy. The recent shift really moved us from writing code faster to writing code at a higher level of abstraction. Those two can be true at the same time, but the mental model itself is different.

The profound shift that's happened to the field of engineering in the past three months is staggering almost to the point of disbelief. Using Codex is like being thwarted by machine code for years and finally discovering HTML+JS. If you are not using these tools as an integral part of your daily routine you're behind. Now is the time to catch up.

@agentmail Very cool guys. Much needed, and saves our claws all those workarounds. Trying this now!

Here's how to use AgentMail with your OpenClaw instantly! docs.agentmail.to/integrations/o…

@pitdesi Chai lover here. Had this product on my radar for a while but wanted some legit user reviews before I bought it. Thank you @pitdesi

Cool product for chai lovers. I always end up babysitting chai on the stove to stop it from boiling over, and half the time I still make a mess. The Loka solves that in a clever way and makes great chai. $150, but it is a really nice product weareloka.com

That's a great litmus test for AI coding tools - can I launch my overnight job and trust the AI to handle whatever comes up while I sleep? I remember feeling exactly that anxiety a year ago with Cursor and Windsurf. You'd come back to a mess half the time. The recent releases of Claude Code have genuinely changed that calculus.

Remember the anxiety of launching an overnight job, worrying that it would hit an error in the middle? What a beautiful thing to launch it in claude code, where a benevolent superintelligence watches over it while you sleep.

@trq212 We don't just want more memory from our AI tools, we want smarter memory. Auto memory is a really good step in that direction. Thank you @trq212 @bcherny and team! Wrote some thoughts about it here jawnty.me/blog/auto-memo…

We've rolled out a new auto-memory feature. Claude now remembers what it learns across sessions — your project context, debugging patterns, preferred approaches — and recalls it later without you having to write anything down.

@ajt Been doing this with my 8 year old and she's been building art/drawing apps, games, etc. Super fun. Highly recommend using @WisprFlow with this

PSA for people with kids here. I highly recommend having each of your kids make an app or product with Claude. I have both my 10 year old and 6 year old using Claude (with my support) to build an idea they each had. My 10 year old is building a mobile app game that lets you design custom outfits and compete with friends on who made the best oens and my 6 year old is making a competitive hula hoop game that is inspired by the structure of mortal kombat. We work on them during the weekend and we plan to publish both to the app store when they are ready. Highly recommend doing this.

@noahzweben /clear

Rolling out Claude Code Remote Control to Pro users - because they deserve to use the bathroom too . (Team and Enterprise coming soon). 🧻 Rolling out to 10% and ramping 1. Update to claude v2.1.58+ 2. Try log-out and log-in to get fresh flag values. 3. /remote-control

@garrytan we are all one prompt away from the next $1B product

Everything can be one-shot. You just have to know the right one-liner.

@andrewchen that's basically users way of saying GUIs got complicated over time, and products like claude code actually listening to their users

command line interface winning over gui isn’t not what I would have predicted for 2026…

@bcherny The hard part of running Claude Code headlessly for me is observation and steering mid-task without a terminal in front of me. Tried building workarounds, but super excited to see native support for this! Thank you @bcherny @noahzweben and team!

Have been using this daily and loving it! Tell us what you think

@steipete been using OpenClaw for the past few weeks. The heartbeat ssytem is one of the underrated features in the whole thing. Thanks for making it!

AI assistants don't remember to check on things for you. You have to remember to ask them. So you've just replaced one thing to remember with another. Not much of an upgrade. Tools like OpenClaw and Claude Code are starting to close that gap, and using them made me realize something about where agents are actually going.