Their apparent inability to get the basics right makes me severely doubt their claims of self-improving AI. The humans at Anthropic wouldn't know improvement if it landed on their lap and started twerking, and AI cannot do a job without strong human intervention into what the goals and guardrails actually are.
I'm kind of reminded of when Microsoft claimed it took a team of Ph.D.s to write a terminal application that updated at 60fps, and then Casey Muratori did it over a weekend. And this was before AI was writing code in earnest; when LLM-induced brainrot really sets in, civilization is in for a world of fresh hurt: lots more generated code, almost all of it garbage. And the promised AI crossover point where it becomes AGI, or indistinguishable from for software design purposes, recedes into the infinite future.
Reading this it occurs to me that this timeline may be moving toward a future where garbage software, garbage information etc. have become the norm for so long that the number of people who can distinguish trash from quality, or signal from noise, has become negligible.
A true era of ignorance, looking like an ocean of nonsense in which no one can really navigate as it is ungrounded in reality.
Idiocracy presents a naively gentle positive version of such future but there are many darker ones possible.
> I'm kind of reminded of when Microsoft claimed it took a team of Ph.D.s to write a terminal application that updated at 60fps, and then Casey Muratori did it over a weekend.
This is the same Microsoft that is now rewriting the TypeScript type checker, parser and its developer tools in Go after realizing that the bottleneck was...the performance of TypeScript itself, which is a basic compiled vs interpreted difference.
> And this was before AI was writing code in earnest; when LLM-induced brainrot really sets in, civilization is in for a world of fresh hurt: lots more generated code, almost all of it garbage.
Some folks using LLMs wouldn't realize why it makes zero sense to use TS / JS for building performant and optimal applications. This is why people were experiencing significant rendering bugs in terminal apps (they are not designed for that) and slow starts with Claude Code, which was completely vibe coded with Ink.
If you don't understand the basic fundamentals of what you are working on with LLMs and bugs are creeping up left and right, then you are just sinking in your own comprehension debt.
I watched the video and I wish I could get those 13 minutes of my life back.
He could have done it in 13 seconds instead of 13 minutes: "Anthropic is lying about the effectiveness of agentic loops because there's this one screen flicker bug in Claude Code that took a year to fix."
Yeah, like when United Airlines claims a plane can fly 300 people 6,000 miles they are lying to you.
I can prove they're lying to you because people have been complaining about uncomfortable seats and flight delays for literally decades and those issues still aren't fixed.
I think this guy is using AI differently than me. Since Opus 4.6 and GPT 5.3, I have been able to absolutely crush my coding work. Boris might be embellishing how he ONLY writes loops, but for the most part I am just handing off planning docs to Claude or GPT and they implement it with like 95% accuracy.
A lot of you don't want to hear it but this is a user issue.
The one thing that gets me is - Boris must know what everyone is thinking when he says he merges 300 MR’s per day, but I think everyone knows it doesn’t mean what he’s implying - it just can’t. He can’t read 300 open source contributions per day to determine if they fit into the spirit of Claude Code. Even if he’d fully automated the testing and integration process. And 300 people per day submitting contributions?
He could mean a few other thingsc one would be, he has like 20 version tags and he merges 15 features into each version (does he explicitly say merge to main?)
The other thing he could mean is he has like a software engineering agent and that software engineering agent like loops through GitHub issues and his personal notepad and maybe uses a few different branches to test things out and build I dunno adversarially, running all sorts of experiments.
Which would be genuinely cool! But using Mr’s to say tweak bunch of variables back and forth isn’t what 300 MR’s implies.
But then the ultimate question is - it may be cool to fully automate a software engineering agent, and certainly is the type of research Anthropic and someone of Boris’ stature (and pay level) should be working on. But is it efficient?
This YouTuber is very knowledgeable, this video in particular was very good to follow along and take in what he says. Usually, I can barely watch any of his videos when they popup as recommend because his humor and way of behaving feels directed towards kids to entertain. So I avoid his channel. This time, I actually enjoyed the video. I wish he continued with this style instead.
Another point I would like to bring up is this, I do not believe the idea of ”solving coding” is appropriate. I do not see these agents as replacement for our crafts, but rather, as a assistive tool for us. Its an extension of us. If you are at the beginner stage of something, and have a wrong approach for something, the agent amplify that issue (a bit of prompting can help mitigate this for beginners ofc, for example the agent can act as a teaching instructor).
I feel like some people don’t know how good Claude Code (the harness) is at general tasks as well, not just coding. It have helped me troubleshoot a lot of issues in a variety of environments, from things in the local machine to external services to the web. It has truly shortened my time of sitting stuck infront of the desktop, starring at the screen while out of ideas.
15 comments
[ 2.8 ms ] story [ 33.1 ms ] threadTheir apparent inability to get the basics right makes me severely doubt their claims of self-improving AI. The humans at Anthropic wouldn't know improvement if it landed on their lap and started twerking, and AI cannot do a job without strong human intervention into what the goals and guardrails actually are.
I'm kind of reminded of when Microsoft claimed it took a team of Ph.D.s to write a terminal application that updated at 60fps, and then Casey Muratori did it over a weekend. And this was before AI was writing code in earnest; when LLM-induced brainrot really sets in, civilization is in for a world of fresh hurt: lots more generated code, almost all of it garbage. And the promised AI crossover point where it becomes AGI, or indistinguishable from for software design purposes, recedes into the infinite future.
A true era of ignorance, looking like an ocean of nonsense in which no one can really navigate as it is ungrounded in reality.
Idiocracy presents a naively gentle positive version of such future but there are many darker ones possible.
Kali Yuga, indeed.
https://en.wikipedia.org/wiki/Kali_Yuga
This is the same Microsoft that is now rewriting the TypeScript type checker, parser and its developer tools in Go after realizing that the bottleneck was...the performance of TypeScript itself, which is a basic compiled vs interpreted difference.
> And this was before AI was writing code in earnest; when LLM-induced brainrot really sets in, civilization is in for a world of fresh hurt: lots more generated code, almost all of it garbage.
Some folks using LLMs wouldn't realize why it makes zero sense to use TS / JS for building performant and optimal applications. This is why people were experiencing significant rendering bugs in terminal apps (they are not designed for that) and slow starts with Claude Code, which was completely vibe coded with Ink.
If you don't understand the basic fundamentals of what you are working on with LLMs and bugs are creeping up left and right, then you are just sinking in your own comprehension debt.
He could have done it in 13 seconds instead of 13 minutes: "Anthropic is lying about the effectiveness of agentic loops because there's this one screen flicker bug in Claude Code that took a year to fix."
Yeah, like when United Airlines claims a plane can fly 300 people 6,000 miles they are lying to you.
I can prove they're lying to you because people have been complaining about uncomfortable seats and flight delays for literally decades and those issues still aren't fixed.
Everyone else needs to start treating them that way, or you're going to regret it once you realize what's actually happening.
A lot of you don't want to hear it but this is a user issue.
He could mean a few other thingsc one would be, he has like 20 version tags and he merges 15 features into each version (does he explicitly say merge to main?)
The other thing he could mean is he has like a software engineering agent and that software engineering agent like loops through GitHub issues and his personal notepad and maybe uses a few different branches to test things out and build I dunno adversarially, running all sorts of experiments.
Which would be genuinely cool! But using Mr’s to say tweak bunch of variables back and forth isn’t what 300 MR’s implies.
But then the ultimate question is - it may be cool to fully automate a software engineering agent, and certainly is the type of research Anthropic and someone of Boris’ stature (and pay level) should be working on. But is it efficient?
I guess yes hems talked about this:
https://karozieminski.substack.com/p/boris-cherny-claude-cod...
Another point I would like to bring up is this, I do not believe the idea of ”solving coding” is appropriate. I do not see these agents as replacement for our crafts, but rather, as a assistive tool for us. Its an extension of us. If you are at the beginner stage of something, and have a wrong approach for something, the agent amplify that issue (a bit of prompting can help mitigate this for beginners ofc, for example the agent can act as a teaching instructor).
I feel like some people don’t know how good Claude Code (the harness) is at general tasks as well, not just coding. It have helped me troubleshoot a lot of issues in a variety of environments, from things in the local machine to external services to the web. It has truly shortened my time of sitting stuck infront of the desktop, starring at the screen while out of ideas.