I'm not sure what counts as vibe coding, but however you want to define it, why assume it has peaked from one example of a messy codebase? It seems like an odd conclusion to draw.
I'm not sure what counts as vibe coding, but however you want to define it, why assume it has peaked from one example of a messy codebase? It seems like an odd conclusion to draw.
The reasons are stated in the article. These are the Makers of AI. They have unlimited access to their own product, dogfood it religiously, and should be more familiar and skilled with it than...
The reasons are stated in the article. These are the Makers of AI. They have unlimited access to their own product, dogfood it religiously, and should be more familiar and skilled with it than anyone else. If they can't hack it with that team, with one of the languages best represented in the training data, that casts doubt on the feasibility of it. Importantly, they are the ones making the claim that you can just throw AI at a codebase and have a usable result at the end.
To be brutally honest? This isn't just one messy codebase—it's one of the flagship products from what many may consider the leading AI foundry. It's a codebase where top minds are on the job, and this is the result.
I do think there are more productive ways to utilize AI, but those aren't what Anthropic claims is possible or what Anthropic is doing.
I'm not sure it follows that the people who make a tool will necessarily be the best at using that tool, especially for something with this much adoption. I believe that they are among the...
I'm not sure it follows that the people who make a tool will necessarily be the best at using that tool, especially for something with this much adoption. I believe that they are among the heaviest users of it, but that's not going to be perfectly correlated with output quality, especially in the very likely case that the best results with LLM coding tools come from people who don't delegate 100% of their thinking to them.
I 100% agree that the best results don't come from vibe coding. But this is the company that's claiming that vibecoding is feasible, dogfooding it, and as the heaviest and least limited users they...
I 100% agree that the best results don't come from vibe coding. But this is the company that's claiming that vibecoding is feasible, dogfooding it, and as the heaviest and least limited users they should also have plenty of trial-and-error and exploratory opportunities that outsiders won't. That's why we're talking about the ceiling of vibecoding and not the ceiling of LLM usage as a whole.
I would not assume that because they make the AI models they would know best how to use them and make them useful. If anything, being too deep in AI qua AI probably makes you distinctly bad at...
I would not assume that because they make the AI models they would know best how to use them and make them useful. If anything, being too deep in AI qua AI probably makes you distinctly bad at understanding how to use AI for anything useful. These guys aren’t coming with expertise in the kinds of legacy burdened, highly interconnected, production environments most complex enterprises are running.
I think there’s still a lot of work to do on the UI and interface ends to make the most out of AI tools. And as those improve the models will need to adapt to function better with those new modalities for working with them.
You’ve listed some of Anthropic’s advantages. I believe they do have advantages. But I can think of other reasons why they might be at a disadvantage: Their engineers might be distracted by other...
You’ve listed some of Anthropic’s advantages. I believe they do have advantages. But I can think of other reasons why they might be at a disadvantage:
Their engineers might be distracted by other projects, like working on the next releases of their LLM’s. They might not have their full attention on Claude Code.
Anthropic’s engineers were the earliest adopters. They started using Claude Code before it worked very well, with earlier releases of their LLM’s that couldn’t write code very well. Being an early adopter means you find and report the bugs. Their messy codebase might be a product of these early experiments. Some other team that’s a fast follower could make a fresh start while learning from Anthropic’s mistakes.
A fair point. But another possibility is that maybe their top people aren’t working on Claude Code at all, because they’re busy doing something else? You can’t count “top minds are on the job” as...
A fair point. But another possibility is that maybe their top people aren’t working on Claude Code at all, because they’re busy doing something else? You can’t count “top minds are on the job” as an advantage if it’s actually the B team.
I know one person that works on Claude Code. They are very smart. I only talked to them before their start date. From the outside they weren't sure why the product seemed so buggy. Hopefully they...
I know one person that works on Claude Code. They are very smart. I only talked to them before their start date. From the outside they weren't sure why the product seemed so buggy. Hopefully they can give me some insight next time I run into them.
I don't think it's likely that Silicon Valley engineers getting paid $400-500k salary are a "B team". The AI researcher salary ranges go higher not because because they're smarter people. They are the ones that drive the value of the company more than the product engineers do.
Idk these days I stopped focusing on vibecoding definitions or other people's opinions of which model is best. I just use w/e I think works and then read the code. Its that simple. There's too...
Idk these days I stopped focusing on vibecoding definitions or other people's opinions of which model is best. I just use w/e I think works and then read the code. Its that simple. There's too much noise out there.
I do understand the author's point but I also think its like a pretty obvious one. Any software engineer worth their salt knows that you can't really just prompt and reprompt, use loops and goals and then throw infinite money to get good software that thousands of people will use.
I do think that Bun rewrite is an exception and an exception that's only viable when you don't consider token costs. Most people aren't gonna be able to afford 16 Fable agents in parallel running for hours.
I'm not sure what counts as vibe coding, but however you want to define it, why assume it has peaked from one example of a messy codebase? It seems like an odd conclusion to draw.
The reasons are stated in the article. These are the Makers of AI. They have unlimited access to their own product, dogfood it religiously, and should be more familiar and skilled with it than anyone else. If they can't hack it with that team, with one of the languages best represented in the training data, that casts doubt on the feasibility of it. Importantly, they are the ones making the claim that you can just throw AI at a codebase and have a usable result at the end.
To be brutally honest? This isn't just one messy codebase—it's one of the flagship products from what many may consider the leading AI foundry. It's a codebase where top minds are on the job, and this is the result.
I do think there are more productive ways to utilize AI, but those aren't what Anthropic claims is possible or what Anthropic is doing.
I'm not sure it follows that the people who make a tool will necessarily be the best at using that tool, especially for something with this much adoption. I believe that they are among the heaviest users of it, but that's not going to be perfectly correlated with output quality, especially in the very likely case that the best results with LLM coding tools come from people who don't delegate 100% of their thinking to them.
I 100% agree that the best results don't come from vibe coding. But this is the company that's claiming that vibecoding is feasible, dogfooding it, and as the heaviest and least limited users they should also have plenty of trial-and-error and exploratory opportunities that outsiders won't. That's why we're talking about the ceiling of vibecoding and not the ceiling of LLM usage as a whole.
I would not assume that because they make the AI models they would know best how to use them and make them useful. If anything, being too deep in AI qua AI probably makes you distinctly bad at understanding how to use AI for anything useful. These guys aren’t coming with expertise in the kinds of legacy burdened, highly interconnected, production environments most complex enterprises are running.
I think there’s still a lot of work to do on the UI and interface ends to make the most out of AI tools. And as those improve the models will need to adapt to function better with those new modalities for working with them.
You’ve listed some of Anthropic’s advantages. I believe they do have advantages. But I can think of other reasons why they might be at a disadvantage:
Their engineers might be distracted by other projects, like working on the next releases of their LLM’s. They might not have their full attention on Claude Code.
Anthropic’s engineers were the earliest adopters. They started using Claude Code before it worked very well, with earlier releases of their LLM’s that couldn’t write code very well. Being an early adopter means you find and report the bugs. Their messy codebase might be a product of these early experiments. Some other team that’s a fast follower could make a fresh start while learning from Anthropic’s mistakes.
Those are different jobs. Anthropic has specific job postings for product engineers who would not be involved with LLM development.
A fair point. But another possibility is that maybe their top people aren’t working on Claude Code at all, because they’re busy doing something else? You can’t count “top minds are on the job” as an advantage if it’s actually the B team.
There’s a lot we don’t know from the outside.
I know one person that works on Claude Code. They are very smart. I only talked to them before their start date. From the outside they weren't sure why the product seemed so buggy. Hopefully they can give me some insight next time I run into them.
I don't think it's likely that Silicon Valley engineers getting paid $400-500k salary are a "B team". The AI researcher salary ranges go higher not because because they're smarter people. They are the ones that drive the value of the company more than the product engineers do.
Idk these days I stopped focusing on vibecoding definitions or other people's opinions of which model is best. I just use w/e I think works and then read the code. Its that simple. There's too much noise out there.
I do understand the author's point but I also think its like a pretty obvious one. Any software engineer worth their salt knows that you can't really just prompt and reprompt, use loops and goals and then throw infinite money to get good software that thousands of people will use.
I do think that Bun rewrite is an exception and an exception that's only viable when you don't consider token costs. Most people aren't gonna be able to afford 16 Fable agents in parallel running for hours.