Christina @ATX

I've created a technical blog dev.to/kitaekatt/ where I'll be posting on Claude Code and games industry workflow topics #ClaudeCode

@MegaCrit I think pressing x to view your relics while in the rewards screen may be the trigger for not being able to navigate to card rewards

@MegaCrit couple bugs I have noticed with controller. Sometimes going up/down during a battle I can't select units to highlight them. Fixed itself lasts only a turn. Possibly related to drawing cards. Sometimes when browsing rewards I can't select a card reward at all. Need to quit the game and restart at which point it is fixed. So minor, game is great!

@thdxr I don't use opencode at all but I still like your tweets!

do you use the opencode sidebar

A peak productivity advice from James Clear: “You are not your grand plans. You are your daily patterns.”

@noampomsky I want to subscribe

friend is in the stage of claude psychosis where he asks claude to send him newspapers about what claude is doing for him

@iruletheworldmo I find it hard to believe anyone thinks closed source can win in an era where quality coding has become so inexpensive

do people really think open source has a chance. this guys argument feels like, give me money and i can build a netflix competitor from my living room. the only way ‘open source’ wins is if openai or dario do it. and they won’t. because if they did. composer 3 would be agi.

@PMThebuilder @karpathy @RhysSullivan Agent code bloat is intrinsic it is not spec limited

"The agents do not listen to my instructions" is the PM problem hiding in plain sight. Every engineering fix in this thread — lint rules, review agents, second passes — is downstream of the real gap: nobody wrote a spec precise enough that the agent couldn't misinterpret it. We're learning that vague requirements don't just slow down humans. They actively mislead agents.

Thank you Sarah, my pleasure to come on the pod! And happy to do some more Q&A in the replies.

@DanielMiessler MCP is how you access the bullshit that doesn't have a cli

MCP is confusing. Some think it died when CLIs took over, but you can also view them as complimentary. Like MCP is how you stay fresh on what's available in the tool/service, and how to use it, and CLI is how you actually execute.

@thdxr Code review skills are commonplace

i still don't get why we need to push code up to get an LLM review via awkward github ui hacks opencode has /review which can also do things like run your code to check things but a full time team focused on this would do it better, i just don't like the workflow they offer

@louisvarge What is the architecture?

i made a thing where now any Claude Code can send messages to any other Claude Code on my machine they can ask clarifying questions about work, or become friends

@noahzweben It would be nice if you can configure claude's CWD it currently is the first directory in the files array. Often the files the user will want to search and browse are not the same as the files we will want Claude to have visibility into.

A few 🔥 Claude Code VSCode drops 1. Remote Control in VSCode

@thdxr Ew Bash4life

we've been experimenting with getting rid of the bash tool agents can write js fine which can do what bash can (though some gaps with things like git) and is more cross platform and then could run that in this

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. github.com/karpathy/nanoc… All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

@andrewchen @emilybenn12 Reviewing every line is based on the idea that the cost of errors is high (impact, detection, fixing). The real cost right now is human attention.

One question I've been asking founders is: do you try to review all the code that the LLMs write or do you just accept it? I think it's about 50-50 right now but the momentum is towards just accepting the AI-generated code and I think that number will eventually go to 100%. To me this is one of the most telling indications of how AI-native a team is. It's hard to get super high throughput if you are reviewing every line.

One pattern from conversations with technical founders in our current cohort: There is a massive hiring divide emerging between engineers who are AI native and those who are not. This is not about seniority. It is about posture. The AI native engineers: -Treat Codex, Cursor, and agents as default infrastructure -Orchestrate systems instead of writing every line themselves -Spend more time reviewing, shaping, and stress testing than writing code They function more like Engineering Managers for AI than ICs in practice. The engineers stuck in older workflows: -See AI tools as shortcuts rather than leverage -Measure output in lines written -Default to manual implementation even when automation is available The hiring question is no longer just: Are they a strong coder? It is: Do they know how to compound their output with AI? If they don't, they will almost certainly get left behind by those that do.

@toddsaunders I'm available for work as a CCO

Every company will have a Chief Claude Officer that pills new employees and keeps everyone trained / up to date. It’s the only way and it’s that important

@rodarmor I do hate this. Sometimes typos send Claude off in the wrong direction.

I feel like coding agents have no backbone or depth to their thinking. Me: [question] Agent: [churns for 5 minutes] You should do X. Me: What about minor problem Y? Agent: You're right, X is completely inappropriate, do Z.

@Sh1nj1_kor @AnthropicAI True!

@truffle @AnthropicAI Most people won't notice, and they'll notice their allowances getting smaller much faster because their conversations will accumulate a lot of useless information that otherwise would've been compacted away at 200k tokens.

I am calling it now. The 1M context window by default is @AnthropicAI rugpulling the MAX subscribers. Opus now by default eats way more tokens than before and is not as aggresive on the cleaning of context, easily eats up your weekly allowance. Time for codex i guess

@flipbit03 @Calvinballrefe1 @exQUIZitely I never finished it but I sure enjoyed the journey!

@Calvinballrefe1 @truffle @exQUIZitely I have ascended, including as a Tourist, an amazing feeling. But it took me no joke 12 years to finish it for the first time

A glitch in the Matrix, all games after the year 2000 are wiped off the earth and you can only play one pre-2000 era game for the rest of your life. Which do you pick? I'd go with Civilization (1991).

The crawfish are huge this year! #texas

@KingBootoshi Thanks for sharing looks very cool

HOLY FUK I JUST LEARNED ABOUT TLA+ AND IT'S SO GOOD FOR AGENTIC CODING ur telling ME that i can mathematically fact check every possible scenario of my design STATE to prevent bugs and crashes AND IF IT FINDS SOMETHING THE AGENTS GET INSTANT FEEDBACK AND LOOP FIXING IT TILL IT ALL POSSIBLE BUGS IN THE DESIGN ARE PATCHED LOL THIS IS OP