prodigycorp

Thought experiment: would a junior developer armed with a modern day gpt-5.5 or opus 4.8 in 2015 would have been considered a 10x developer (assuming nobody knew they were using AI)?

what are ppls' impression of pyrefly? i've become completely captive to uv's tooling. it has allowed me to think only about coding versus tooling. dont feel like giving another typechecker a chance unless it offer's something i'm not getting from ty.

Yeah it was a good discussion. I'm definitely on the side of "a well written agents.md is very good for the agent". The file should not be static. The agents.md from a year ago is not the right agents.md for today. Many of them are overly long, overly instructive, and include a lot of unccessary code bits.

I think it's useful for people using agent harness to regularly evaluate your skills, agent rules, and memory implementation to ensure there's no conflict across them all. Also, best to rely as little as possible on the agent to write its own agents.md.

It can be be tedious, but that's why agentic coding can still be considered a "skill".

Even then, i highly doubt any sort of automation is producing on the order of several millions of tokens daily. The issue I see with the org in parent comment seems to stem from management and not any sort of token repricing.

using openrouter leaderboards is no good. being on top of the board is marketing, so some labs are gaming that number. all marketing spend.

To me this is clearly a skill issue. Several millions of tokens per day is peanuts, even if uncached. gpt-5.5 is $5 per million of input tokens.

Anybody doing things seriously understand how to optimize their workflows for smaller models once they start to lock in processes.

Korea is backwards in technology in every possible way.

- For the longest time, you needed a windows computer to access any sort of government or banking service, and it's still the case for most services

- Because of the reliance on crappy windows laptops, you see everyone who uses a laptop carries an external mouse around to places like coffee shops (bc their trackpads suck)

- the de-facto document format are crappy hancom formats

- watching korean news is farcical - every time they cut to public footage, literally 80% of the frame is blurred. I see no point in even watching the news.

- APIs and API documentation for stuff is sooooo poorly designed/written. Like, it's a f-ing joke.

- External map providers were iced out of hte market until this past year

- You need a phone number to sign up for literally anything.

There are so many more examples but these are just the ones off the top of my head. There is not an inch of breathing room for dynamism.

Koreas issues arent political. This is what happens in pure oligopolies. People on twitter love to fantasize about Korea being so technofuturistic but the truth is that the startup culture is terrible, there's no venture capital scene, and the big companies write all the rules

How realistic is this:

Chinese models incidentally slurps up some terms that lead them to finding unflattering words that you wrote about the CCP in a random journal entry, or maybe a social media csv export. You go to China one day and are denied entry due to what you said.

Realistic or no? (yes i know the us is getting bad in re. to what you write online as well)

Models hosted in China are a siren call that I don't feel bad about resisting.

He'll be able to raise debt against the massively inflated valuation of the company. SpaceX will have access to far more capital after this than before.

Perhaps there is reason to be optimistic. If we believe nothing online, then digital inauthenticity may cause us to yearn more for authentic, in-person interactions.

The modern web has not been without its warts. It's led society to some very unhealthy behaviors. It's telling that a lot of people think the internet is society, rather than a slice of it. Perhaps its decline is the correction society needed.

It's easy to feel like a doomer if you're thinking "it's sad that the web is being torn apart" and your framing of the web is "the magical place of promise it used to be 25 years ago."

That could be it. I still see LLMs fail a set of static typing challenges that I created a couple years ago as a benchmark. Google models still fail it. I wonder if the lack of typing in a lot of the training data makes python harder to reason about?

the funny thing is that everyone, including myself, posited that python would be the winner of the ai coding wars, because of how much training data there is for it. My experience has been the opposite.

Someone scold me if I'm wrong but this is really worrying. Threat actors with Github's internal code means a huge acceleration in vulnerability discovery for the one platform where everybody warehouses their code.

How is this not really, really bad?

WTF, THIS WAS KARPATHY?!

i meant everything out of respect for andrej. it's no different from how a visiting scholar can be great marketing for an institution

i mean he did publicly openly solicit interest to work at a frontier lab so he can be closer to what's going on https://www.youtube.com/watch?v=kwSVtQ7dziU&t=2870s

i wouldn't be surprised if he just becomes a glorified marketer for anthro.

im also going to guess that whatever research he does would be free roam research that primarily serves to market the fact that claude was able to help perform the research.

the visible stuff he's been working on has been mostly agent soft skills. off the top of my head is autoresearch and his the wiki knowledge stuff. nothing particularly groundbreaking, but has helped devs expand their understanding of the utility that these models can provide.

not a diss to andrej i know he's reading this now

this is so funny because it's true. it was a handy library but a lot of the shine wore off when I realize how it actually worked.

Perhaps, but I'm commenting on a blog post called "How Claude Code works in large codebases". That's an interesting question to me. I had hoped there was a more interesting answer.

My complaint is about how there's not enough information in the blog post. The title of the post is "How Claude Code works in large codebases". 1521 of 18135 characters is dedicated to expanding on the premise of the title.

My criticism is fair. This is not an engineering blog post, it's purely marketing.

This is really a zero information blog post. I want to know how they use the LSP to improve their understanding of the code base. Would be great if it was open source for us to review.

A post like this should be providing people with some reassurance about Claude's ability to understand code at a large scale. It's mostly fluff.

Edit: so I did some googling to dig around for thoughts on LSP performance and integration. the author of bun has a tweet about saying that they are a big drag on performance for no real gain and virtually all of the replies agree. Anyone else have any experience/thoughts?

https://xcancel.com/jarredsumner/status/2017704989540684176

alphabook

I hit both a google QR captcha and traditional captcha today. They're phasing it in.

Ok, so we treat it as an act of war. Now what? Attack North Korea? Great, the entire city of Seoul gets shelled within five minutes of your attack and hundreds of thousands of innocent people die.

It's very easy to play with lives that aren't yours.

This used to be the case, but the changes last month have rendered the Gemini Pro plan completely unusable.

I think you can see this one of two ways: you could also consider it a miracle that the qwen models are able to perform so well when being trained on inefficient wrapper code data.

Idk how to make it more obvious that you’ve used a vote ring than 15 votes in ten minutes.

I hereby propose we rename the HN frontpage to "Claude Customer Support"

i run qwen 3.6. you need to drink some settle down juice.

Am I the only one who thinks that your stop hook is written extremely poorly? Not only that, but you're writing to the LLM like an abusive human. No wonder it wants to go home.