Gyula Halmos

I am a tech enthusiast who loves working on the bleeding edge and tinker around. I am a staff AI product manager @craftdocs

[Mnimiy @Mnilax]

Karpathy threw a grenade at every senior engineer who still treats LLMs as a toy.

his actual words: the worst thing an expert can do right now is reject them.

most experts read it as a threat, but it's advice.

his framing:

> the gap between "AI tools are bad" and "AI tools are useful when used right" is professional discipline, not capability
> agents have cognitive deficits. they fail in ways nothing in the training set anticipated
> the experts who reject LLMs lose to experts who learn to wrangle them
> "models have so many cognitive deficits. but you can route around them"

routing around the deficits is what CLAUDE.md was invented for.

Karpathy himself wrote 4 rules. across 30 codebases they took my Claude error rate from 41% down to 11%. solid drop.

but his rules pre-date the slop era going public. I bolted on 8 more, tuned to the failure modes that surfaced after January. got it down to 3%.

a CLAUDE.md does not raise Claude's IQ. it lowers his slop floor. that is the entire game.

open the article underneath.

the model is not the bottleneck. your config is.

[Gyula Halmos @GyulaHalmos]

@bcherny @jaredctate @openclaw But why not collaborate, that they use Claude Agent SDK for Anhtropic models, so Open Claw can benefit from it and Anthropic can keep those aubscribers as now i can see that a large chunk of Anthropic subs will cancel instead of moving to API key. What is the big picture here ?

[Gyula Halmos @GyulaHalmos]

@JavierSnchez2 @balintorosz Hey @JavierSnchez2 do you have some examples that I can use for repro ? Thanks a lot :)

[Gyula Halmos @GyulaHalmos]

@imaurer @UnslothAI Or Craft Agents :P

[Gyula Halmos @GyulaHalmos]

@ibrahim64k @balintorosz I doubt leaving an agent alone for hours means quality. With a fast model you can have a human in the loop engaged and achieve better quality in shorter time. For scanning and research long running agents are fine for quality fast agents with short feedback loops are the best.

[Gyula Halmos @GyulaHalmos]

@y_lukianov @steipete Shouldnt the solution to block OAuth endpoints for 3pp apps ? I think that is the simplest what is the point of banning accounts ??

[Gyula Halmos @GyulaHalmos]

@op7418 @mmee_io Can you please report it in https://t.co/MNKqtTZAaF for the issues ? I mean how 3pp providers are set up in Asia and we can test it and make sure all functionality is available for you too.

[Gyula Halmos @GyulaHalmos]

@balintorosz @zee_zapata Agree, i like this concept. Gj!!

[Gyula Halmos @GyulaHalmos]

@tweetadhish @balintorosz Should be fixed in the latest (0.2.21)
https://t.co/3B97JY1NeW

[Gyula Halmos @GyulaHalmos]

Working on this in the last 4 weeks, was truly a new experience and a glimpse in the future :)

[Balint Orosz @balintorosz]

https://t.co/J701WKvfpB