mezyt

Yeah, I'm prioritizing the issue and PR then pre-reviewing them in an active open source repository absolutely drowned in LLM contributions.

386

I personally never was affected by a bug in Claude code and I've been using five in parallel for a year, for 10 hours a day, everyday.

Right, but this happened during the night (11PM) and over 8 hours instead of one in order to delay the alarm and maximise damage (source: Le monde).

I was wondering how anyone could reach the token limit of a 200$/month Claude subscription. I never did on a 20$ one, and I'm parallelizing Claude to the absolute limit of what I can reasonably keep in mind. And I do that day and night, on week-end too with open-source. Now I'm absolutely convinced that it's by doing crap architectures and design that require 200k token of context before adding a single checkbox. If you apply good design principle and course correct the AI in real time before it commits to really stupid way to fix things (i.e. you don't ask it or let it do things that would make zero sense if you had to do it yourself), then the vibe coding cost way less and is way faster and easier to reason about.

Makes sense for small tool like ls, and doesn't for things that are actually complex like the python language or sqlite.

The 2000x number is based on a pathological directory in black's repo https://github.com/psf/black/tree/main/profiling which make duplicate-code really slow and a fix was done in pylint 4.1.0 according to pylint's maintainer.

ruff is a rust rewrite of flake8. Got bought by OpenAI after a meteoric rise in the python ecosystem.

Great case where rust works well too. I won't cite every famous libs that got rewritten in rust but it wasn't all with LLM.

Turn out people like to ask low quality questions, as evidenced by the reputation of Stackoverflow moderators.

Profession (1957) by Isaac Asimov is relevant: https://news.ycombinator.com/item?id=46664195

I once copy pasted a spam email in https://www.bullshitremover.com/ and it simply returned "bullshit".

The worse system is already getting gamed. There's already too much on the line for researchers/students, so they don't admit any wrong doing or retract anything. What's the worse that could happen by adding a layer of trust in the h-index ?

The metaphor doesn't match very well here because stackoverflow is not selling new tape at a premium but giving them for free and reading a stackoverflow answer is harder than asking an LLM.

Could be that AI companies feeding on stackoverflow are selling tape at a premium, and if they tell you it's only supervised learning from a lot of human experts it's going to destroy the nice bubble they have going on around AGI.

Could also be that you have to do the actual theory / practice / correction work for your basal ganglia to "know" about something without thinking about it (i.e. learn), contrary to the novel where the knowledge is directly inserted in your brain. If everyone use AI to skip the "practice" phase lazily then there's no one to make the AI evolve anymore. And the world is not a Go board where the AI can learn against itself indefinitely.

Remind me of a recent discussion we had among Stackoverflow moderator:

> “Think about it,” he continued. “Who discovers the edge cases the docs don’t mention? Who answers the questions that haven’t been asked before? It can’t be people trained only to repeat canonical answers. Somewhere, it has to stop. Somewhere, someone has to think.”

> “Yes,” said the Moderator.

> He leaned back. For a moment, restlessness flickered in his eyes.

> “So why wasn’t I told this at the start?”

> “If we told everyone,” said the Moderator gently, “we’d destroy the system. Most contributors must believe the goal is to fix their CRUD apps. They need closure. They need certainty. They need to get to be a Registered Something—Frontend, Backend, DevOps, Full stack. Only someone who suffered through the abuse of another moderator closing their novel question as a duplicate can be trusted to put enough effort to make an actual contribution”

A half conversation is a lot more disruptive because your brain try to fill in the gap of information.

Plus, how many shell can one individual open in a day ? I'm doing that once per day on a good day, maybe twenty if I have a lot of unplanned work on subprojects that needs to be done concurrently with my main task.

shadowsocks was the winner of the state of the art I had to do at work. It address the "long-term statistical analysis will often reveal a VPN connection regardless of obfuscation and masking (and this approach can be cheaper to support than DPI by a stat)" comment.

There's a script to update from python2 to python3, it's now the most used language in the world, and they learned their lessons about the python2 to python3 migration. A python3 script is literally the most likely candidate to be still working/maintenable by someone else in 20 years.

Meanwhile, as a maintainer, I've been reviewing more than a dozen false positives slop CVEs in my library and not a single one found an actual issue. This article's is probably going to make my situation worse.

> There is probably more JS written than any other language by orders of magnitude.

And the quantity of js code available/discoverable when scrapping the web is larger by an order of magnitude than every other language.

Yeah ok. I was viewing AI as "a tool to help you code better", not as "you literally can't do anything without it generating everything for you". I could do some assembly if I really had to, but it would not be efficient at all. I wonder if there's actually "developers" who are only prompting an LLM and not understanding anything in the output ? Must be generating dumpster fires as you said.

> I fear that LLMs have already fostered the first batch of developers who cannot function without it.

Playing the contrarian here, but I'm from a batch of developers that can't function without a compiler, and I'm at 10% of what I can do without an IDE and static analysis.

Code coverage tools allow to pragma the defensive code which will appear reasonable to most reviewers ?

I've hired by Linkedin recently and had to triage that stream of shit. There's 50% of candidate with no qualifications at all, then 25% that are somewhat qualified for something but not for the job at hands, then on those qualified there are some which might have left actual ChatGPT identifiable output in their own CV ("this job summary is short and insightful and will increase retention rate by 55%"), then of those that actually took some care of their CV and/or cover letter you realize their 'personal website' is an empty shell with fake projects and links to 2s YouTube video telling you that the demo is coming soon... you get the idea.

If you match the job well and want to get it, and actually accomplished something at some point, you should try to get it even if Linkedin showed you a high number.

> Well, nobody will be able to reproduce their work (unless other people also publish fraudulent work from there)

In theory, yes, in practice, the original result for amyloid beta protein as the main cause of Alzheimer were faked and it wasn't caught for 16 years. A member of my family took med based on it and died in the meantime.

> Many jurisdictions are downright hostile and obstructionist for any renewable investment.

Right, as long as energy prices are not guaranteed no one is going to invest massively. No one want to be the one having to eat up a 8000$/mwh price like in Texas in the Winter. The 3 month mean volatility of energy prices in Europe is 350% (with peak of more than 1500% in 2009 and 2021). For comparison Bitcoin volatility is 54% in the last 10 years.

> As for employees? They are typically not calling the shots about company direction.

They can be motivated or not, knowing that the founder made big bucks and they made nothing is bound to lower motivation. Thus the title of the article, founder's liquidity is a well guarded secret.

> I recently need to find some details of hardware implementations for a feature. I used Google to find specs for this hardware. The specs were hard to read or didn't explain things as I needed. So I asked ChatGPT and it gave me what I needed. It first regurgitated info from the specs but I was able to ask it to explain pieces in more detail and it was great! At what point do I just stop using Google search to find an answer?

But was it accurate ? Chat GPT just provided you with the statistically most likely specs. If the actual doc for this hardware is lacking, this is 100% hallucinations.

My experience is that ChatGPT is simply bullshitting you (sometime accurately), Google is drowning the info in 20 links (where they try to make you buy "Hardware you're searching for"), and you have to go to DuckDuckGo to find the thing you're actually searching for.