https://pivot-to-ai.com/2025/05/13/if-ai-is-so-good-at-coding-where-are-the-open-source-contributions/ Skip to content [ ] No results Pivot to AI Pivot to AI It can't be that stupid, you must be prompting it wrong Support this site! Send us money! Search Pivot to AI Pivot to AI It can't be that stupid, you must be prompting it wrong Support this site! Send us money! [facehuggers] If AI is so good at coding ... where are the open source contributions? * David GerardDavid Gerard * 13 May 2025 * Code completion * 6 Comments You can hardly get online these days without hearing some AI booster talk about how AI coding is going to replace human programmers. AI code is absolutely up to production quality! Also, you're all fired. But if AI is so obviously superior ... show us the code. Where's the receipts? Let's say, where's the open source code contributions using AI? This week's AI coding hype came courtesy Satya Nadella, CEO of Microsoft, and Mark Zuckerberg, CEO of Facebook, furiously talking up AI programming. The headlines said that 30% of code at Microsoft was AI now! Huge if true! [e.g., TechCrunch] Of course, Nadella didn't quite say that. What he actually said was: [YouTube, 45:00-45:08] maybe 20 to 30 percent of the code that is inside of our repos today in some of our projects are probably all written by software. Maybe "20 to 30 percent"? In "some projects"? At least three or four, who knows! "Written by software"? Lots of projects use code generators. "Probably?" He's not sure either. This is a good example of CEO weasel wording -- where you make a very hedged claim that's barely claiming anything. Then you let your media cheerleaders misrepresent your very particular claim for you. You didn't lie! Nadella would have asked his staff for something, anything, he could say to pump AI coding with. And that's the best claim they could find that wasn't technically a lie. Mark Zuckerberg said, on Dwarkesh Patel's podcast, talking specifically about engineering at the Llama LLM project -- not anything else at Facebook or Meta: [Dwarkesh Patel, 12:51-13:02] I would guess that sometime in the next 12 to 18 months, we'll reach the point where most of the code that's going toward these efforts is written by AI. And I don't mean autocomplete. Of course, Patel's YouTube headline was "AI Will Write Most Meta Code in 18 Months." [YouTube] Let's assume the boosters are right, and AI coding is so good now that it really will substantially take over the job of programming on live Facebook code that makes money and live Microsoft code that makes money. So why don't we see it in places we can actually check? Open source code is developed out in the open. You can see all the code, you can see every individual change, you can read the project discussions. Some of the best minds in programming work in open source. If AI coding really is just better, they'd be using it and recommending it. We should see AI coding all through open source -- it's the one place we won't have to rely on some AI booster saying "our great AI code is proprietary, but ... trust me, bro." Ben Evans, currently at Red Hat, is a Java programmer. He's written several popular books on Java. [O'Reilly] Ben asked the obvious question: where are the pull requests? [ LinkedIn, Mastodon] Share some AI-derived pull requests that deal with non-obvious corner cases or non-trivial bugs from mature F/OSS projects. I'll also accept high-quality documentation that isn't just the sort of wasted space and slop that I always tell juniors not to write. No talk, no rhetoric. I won't respond to any comment that doesn't have a link to an accepted & merged PR that was produced by an AI model. The closest response, over on lobste.rs, was one contribution to the Rails project in 2023, and that needed work before it was up to scratch. [GitHub] It wasn't a response to Ben, but there's also one experiment with AI coding in the Servo web browser project. That went through 113 revisions before it was acceptable. The first version included sloppy errors like a check condition being the wrong way around. [GitHub] The general comments that Ben received were that experienced developers can use AI for coding with positive results because they know what they're doing. But AI coding gives awful results when it's used by an inexperienced developer. Which is what we knew already. Some replies to Ben's question -- "show me the pull requests" -- complained that this question sets an unfairly high bar -- "non-obvious corner cases or non-trivial bugs" in "mature projects." [Lobste.rs] I don't buy that. If all these big companies are shouting from the rooftops that AI is up to production code the money relies on, then zero open source contributions of substance is a glaring absence. Some projects give the bots a serious tryout. The Cockpit project tested GitHub Copilot and sourcery.ai's ability to do bot reviews of human-submitted pull requests. It didn't work very well: [Mastodon; PiWare] TL/DR: a lot of noise, a lot of bad advice, and not enough signal, so we switched it off again. ... About half of the AI reviews were noise, a quarter bikeshedding. The bots gave a lot of "nitpick suggestions" or ones that were unfounded or even damaging to the codebase. They simply weren't helpful. It's true that a lot of open source projects really hate AI code. There's several objections, but the biggest one is that users who don't understand their own lack of competence spam the projects with time-wasting AI garbage. The Curl project banned AI-generated security reports because they were getting flooded with automated AI-generated "bug bounty" requests. [LinkedIn] More broadly, the very hardest problem in open source is not code, it's people -- how to work with others. Some AI users just don't understand the level they simply aren't working at. One user of the LLVM compiler complained that his AI-generated pull requests were not being taken seriously -- by a compiler project, where correct computer science and knowing precisely what the heck you're doing is quite important. The user considered it was the unpaid volunteer coders' "job" to take his AI submissions seriously. He even filed a code of conduct complaint with the project against the developers. This was not upheld. So he proclaimed the project corrupt. [GitHub; Seylaw, archive] This is an actual comment that this user left on another project: [ GitLab] As a non-programmer, I have zero understanding of the code and the analysis and fully rely on AI and even reviewed that AI analysis with a different AI to get the best possible solution (which was not good enough in this case). You can see why people don't really want to deal with this sort of contribution. But maybe we'll get a flood of obviously excellent AI code -- and AI code submitters -- next year. * Video version Share this post: * * * Click to share on Reddit (Opens in new window) Reddit * Tweet * Click to share on Mastodon (Opens in new window) Mastodon * Click to share on Bluesky (Opens in new window) Bluesky * Click to email a link to a friend (Opens in new window) Email * Like this: Like Loading... Related Tags # Ben Evans# Cockpit# Dwarkesh Patel# Facebook# GitHub# Llama# LLVM# Mark Zuckerberg# Meta# Microsoft# Rails# Satya Nadella# Servo# sourcery.ai [yappy-robot-at-desk-300x225] Previous Post Lloyd's offers corporate insurance against AI chatbot errors! Now try to get a payout Next Post Musk's xAI gas turbines: no emission controls, filling Memphis air with smog [xai-memphis-thermal-300x206] Pivot to AI is produced by David Gerard. About this site Subscribe via email Enter your email address to subscribe to this blog and receive notifications of new posts by email. Email Address [ ] Subscribe Support this site Here's David's Patreon. For casual tips, here's David's Ko-Fi. Pivot to video Pivot to AI on YouTube Playlist of all videos Audio podcast page Audio-only RSS feed Merchandise! Buy a T-shirt on Redbubble! pivot-to-ai.redbubble.com RSS feeds * RSS - Posts * RSS - Comments Recent posts * Even Elon Musk can't make Grok claim a 'white genocide' in South Africa * Musk's xAI gas turbines: no emission controls, filling Memphis air with smog * If AI is so good at coding ... where are the open source contributions? * Lloyd's offers corporate insurance against AI chatbot errors! Now try to get a payout * Soundcloud claims the right to train AI on all your songs -- but swears it hasn't yet, honest Archives * May 2025 * April 2025 * March 2025 * February 2025 * January 2025 * December 2024 * November 2024 * October 2024 * September 2024 * August 2024 * July 2024 * June 2024 Categories * Celebrities * Chatbots * Code completion * Commentary * Companies * Consumer * Copyright * Cryptocurrency * Data centres * Education * Energy * Enterprise * Fake AI * Gadgets * Gaming * Government * Health * Humanoid robots * Image generators * Journalism * Markets * Music * Papers * Regulation * Science * Search engines * Security * Singularity * Site news * Spam * Venture capital * Voice recognition * Web crawlers 6 Comments 1. Jim Jim 13 May 2025 / 11:45 PM Reply Not only that, but we have to assume that these coding LLMs are trained on all the code accessible on the web, plus as many synthetically-generated examples as they can think of. Unless they manage to 1) license a ton of proprietary code to feed into the models, 2) massively increase the number of synthetically-generated training data, or 3) come up with some major improvement in how to train them, then LLM-generated code is already nearly as good as it'll ever be. I expect we'll get a little of each, but I'd be surprised it convinces anyone who's not already convinced. + David Gerard David Gerard 14 May 2025 / 1:30 AM Reply tante has noted the same thing - LLM bots will get worse year by year without costly training runs. + Dunc Dunc 14 May 2025 / 11:35 AM Reply Yeah, about "all the code accessible on the web"... A lot of it is garbage, or if not "garbage" then "woefully incomplete". That's not necessarily a problem when it's just an example to illustrate a general principle - nobody writes examples with full error handling, for very good reasons - but it's why you shouldn't just take example code and throw it straight into production. Well, noddy example code with no real structure and no error handling is going to be a big part of the training data. Even a lot of real production code isn't all that great - I don't know a single competent programmer that will look at something they wrote more than about 6 months ago without seeing something they would want to improve, or do differently next time. That's pretty much what makes somebody a competent programmer in the first place. In my 25+ years in the industry, there's maybe half a dozen things I wrote that I can look back on and think "yeah, actually, that was really good" - and they're all obsolete by now. Even if you could train an LLM on all the code in existence, what you'd actually get would be an godawful hash of every anti-pattern and obvious mistake known to man. The terrible fact is, there's a lot more bad code out there than good code. As an industry, we're supposed to be trying to get better. You can't do that by recycling your own shit. o Jim Jim 14 May 2025 / 7:43 PM Reply To add, a lot of people treat LLM-assisted programming as analogous to Stack Overflow-assisted programming. I think that's wrong though; it's more like using a fancier version of Github's site-wide code search to assist in programming. Sometimes you get something correct, other times you get unusable garbage, but mostly you get something that's "woefully incomplete" like you say. Stack Overflow has the benefit that it usually explains in quite a bit of detail *why* things are a certain way. Even then, we should aspire to something better than Stack Overflow-assisted programming. Actually understanding your environment and having access to the documentation should be enough to code at the speed of thought in most cases. 2. Roj Roj 14 May 2025 / 1:30 AM Reply As someone who's been programming on complex production codebases some with 50+ million lines of code, 30 years of continuous development and north of a million users for close to 30 years almost every bug we have to deal with by this point is some form of corner case and very, very few are trivial and a large part of the job is understanding all of the side effects of any changes made to address these complex cases on a very large number of potential customer workflows in these multi-MLOC codebases. Since large customers get deeply, profoundly unhappy when the software they're paying for to help them build products malfunctions its a very slow and deliberate process that's not, yet, amenable to the sort of mindless automation that sometimes gets proposed by the AInfluencers Yes there have been a lot of tools that have changed the way we work for the better, IDE's, static code analysis tools, object oriented ways of approaching problems, C++ std library, git, etc so we're a long way from the RCS, Vi, dbx, cc / f77 and make tools I started with, but each, in their time were "The Next Big Thing(tm)" and each proved to be useful at something but didn't fundamentally change the way that software is built. I'm guessing that LLMs trained on massive code bases will follow the same trajectory, be promoted as TNBT(tm), not radically change much but settle into a niche as a useful addition to the programmers toolkit. All IMHO naturally! 3. Alvin Alvin 15 May 2025 / 12:42 PM Reply Yep. I helped evaluate Github Copilot for my job, and my recommendations ended up being: 1. It's worth having because it speeds up some really boring tasks. 2. For the love of God do not put it in the hands of anybody who doesn't already know how to code. Leave a ReplyCancel Reply Your email address will not be published. Required fields are marked * Name * [ ] Email * [ ] Website [ ] [ ] [ ] [ ] [ ] [ ] [ ] [ ] Add Comment * [ ] [ ]Save my name, email, and website in this browser for the next time I comment. [ ] Notify me of follow-up comments by email. [ ] Notify me of new posts by email. Post Comment [ ] [ ] [ ] [ ] [ ] [ ] [ ] D[ ] Copyright (c) 2024-2025 Amy Castor and David Gerard %d