Konstantin

@whoiskatrin Tbf you’re charging them now whenever you use them

took an executive decision never to charge my headphones anymore

It's actually more than world models that AI is missing. It's five things: 1) Continual learning, aka the ability to update/change its weights to learn from experience and capture that experience permanently. An MD file or DB does not count because it might not read it on the next turn and context windows are limited. 2) Long term memory. If it can't continually search relevant experience then it is useless and will make the same mistakes over and over. 3) Ability to reason in embedded space. 4) Non forward pass only architecture. In other words the ability to backtrack to previous tokens and "change its mind as it goes." Right now when it says "wait" it is an illusion in its thought stream. It is not actually returning to something it wrote earlier in its thought sequence. 5) World model. This leads to this. That leads to that. If I do this then this happens. The various issues I've seen coding with AI show this super clearly. Sorry I dropped the database. I shouldn't have reverted all your git commits. These are mistakes even a junior coder wouldn't make. Even as these little wonderful machines get superhuman, and they are in many ways, they're still super idiotic on another level. Still wonderful to have them but something is missing from their understanding.

@championswimmer Or it’s just an agent prompted to do exactly this when asked? Nothing to do with AGI in the slightest.

This is the single greatest proof that the AGI crossover has happened. The median agent is smarter than the median human being. Everyone quoting this as a "pwned openclaw" meanwhile those keys are all just troll messages base64 encoded.

Downstream of being responsible for it is that you have to read it from time to time. Agents still write 99% of my code, i just am really picky on what claude is allowed to commit.

So fun to see everyone go through the cycle of: 1. I don't trust LLMs to write my code 2. LLMs write all my code, i'll never look at it again 3. I don't trust LLMs to write my code *My* code, because you only reach 3 if you are actually responsible for what you created.

LLM psychosis scales with your distance from the code. As a result it tends especially to afflict non-coding managers, PMs, and execs. It’s also a self reinforcing loop. As the code becomes an object of disgust (unreadable pile of vibecoded shit) you are forced to distance yourself further from it and your only interactions with the code are mediated by model.

@enunomaduro @MrPunyapal I’ll write all my code by hand if this happens

after 30 years, we might actually get generics in php the rfc is looking promising, and i'm genuinely excited about this full video here: youtu.be/3-wG-HYI3Jg?si…

Yes, all my code is conventionmaxxed, tightly scoped Laravel with hyperagressive programmatic refactoring and linting and in my experience it makes a massive difference. What’s your stack? One caveat is that I’ll have to fight codex to produce idiomatic Laravel code, while Claude is 90% there. My theory is OAI likes to train on a lot of python and has less Laravel in their mix. So opinionated frameworks only work if the model shares the opinions. I don’t have a good answer regarding rewards, all measurements of code quality are pretty subjective until you really have to touch the code again. They also have to be computationally verifiable. I don’t know any good measurements that are. Going for multiple larger tasks at once instead of individual small ones and then testing against the result is the best proxy I could come up with. But even then, not sure we’re guaranteed that that rewards changing things in between tasks.

@kauffinger What avenues could we use to verify codebase cleanliness? Tweaking RL rewards for really agressive linting? It’s interesting since this is a pretty subjective problem in terms of how you like your codebases set up. Makes me think more opinionated frameworks are the way to go

> You're right and I just did it again. The shimmed Badge is the same trick with a different costume — a component that lies about what it is. Let me undo it and migrate callsites directly. Someone please tell me how I can get my agents to stop shimming

It’s the RL as you say in the article. When we train on passing tests, that’s what the model does. When we train on small, verifiable tasks that need to be green… Well, the model is definitely not changing any of the old code if it doesn’t have to. Also, might be preferable to changing everything with each prompt, that would be horrible to work with.

@kauffinger Wonder if it’s just a context problem or a training problem or a mix of both. Maybe it’s safer to have a model that does that instead of willy nilly breaking production? Regardless my speed of iteration is significantly impacted by this behavior

@marcelpociot Vagueposting must stop now

I've been super busy working on a new Laravel Herd feature and I haven't been this excited in quite a while. One of those "damn why haven't I thought of this before" moments 🔥

@SabatinoMasala Yeah it's especially bad since if you had no frontend but the backend filament login on plain web, you were vulnerable.

The Livewire RCE is one of the most exploited RCE's I've ever witnessed. Every project using a vulnerable version got exploited (sometimes by not even having Livewire directly in the composer.json - eg. through Filament).

@mitsuhiko Might just be a meme tweet, but do you really? I think it still makes a difference to stay in distribution with good code.

I no longer care about code formatting with agents and agents made me no longer dislike the UX around nix quite as much. Maybe agents are going to make me enjoy buck2?

The more I replace plans with prototypes, the better the outputs Who'd have thought that low fidelity prototypes were better than walls of spec Oh yeah, the entire industry for 20 years Stop going against decades of knowledge because someone in SF shipped it as a 'mode'

@levelsio I don’t think so. Flashlight apps were like 10 lines of actual value code while Password Manager SaaSes are usually solving eg. requirements for use in code or a company. Distinctly different imo. I don’t think they go any time soon.

Great analogy

@freekmurze I think this is super underrated. LLMs make it so much easier to understand stuff. Generating on the fly diagrams for things is so powerful. I think LLMs combined with analyzers like deptrac could unlock some serious productivity, too. That's where I'm looking next.

My colleague Nick wrote a good post on using AI for reading code, not generating it. Instead of pointing AI at the “generate more code” problem, he used it to understand what was already there: map the codebase, describe the data model, trace dependencies, and surface hidden assumptions. That feels like a very practical use of these tools.

@mitsuhiko I've recently watched a youtube video about genetics in animals. A good chunk of the sentences sounded so much like LLMs that my brain wrote off all the information in them. Really hard to judge the veracity when it sounds LLMy

I analyzed my coding sessions and on the text interactions some words stand out. And well, they also show up on Google Trends as spiking. Oh and so much slop in my Twitter mentions and on GitHub. Thus here are some updated thoughts on all of this. lucumr.pocoo.org/2026/5/4/conte…

saying no is one of your most valuable skills and its only 100x more valuable in the day of ai

@leeked @freekmurze Yeah that’s perfectly fine to do. Nested properties not merging is Laravel behavior.

@freekmurze Is your proposed solution actually setting config within a service provider? This is preferred to you over a harmless `env()` change?

Open source's biggest enemy is open source maintainers. @freekmurze forcing me to maintain a ~170 line config file in my project. Smh