Will AI help or hurt open source communities?
Have been thinking about this a lot lately...
Had an interesting chat.
https://chatgpt.com/share/6a08a967-20ac-83ea-87e2-a9bb7d46db1e
Thoughts?
Have been thinking about this a lot lately...
Had an interesting chat.
https://chatgpt.com/share/6a08a967-20ac-83ea-87e2-a9bb7d46db1e
Thoughts?
Comments
Have been thinking about this lately too ...
Because of @yorik complaints with AI in FreeCAD.
Which had some concerns and arguments I had not yet on my radar ....
Which got me thinking that I have quite some disinterest in AI, because of my not so satisfying or successful first experiences with AI (chat bots/LLMs more than image generating) - but have still no real opinion.
On the other side I see your impartial usage of AI and finally your productivity in regions far off your initial architectural background ....
I think these tools can't be ignored, they really work, but also they are completely unsustainable - both in the ecological sense of power and water consumption, but also that these companies make no conceivable financial sense.
@brunopostle could you share more infos about AI unsustainability and water consumption?
@Massimo the energy consumption is huge, but this requires an equivalent amount of cooling. These compute data centres no longer rely on AC, they use freshwater evaporative cooling: https://harvardsciencereview.org/2026/02/28/re-architecting-the-ai-server-the-hidden-water-cost-of-data-centers-part-ii/
@Massimo
Sabine is my favorite scientist on the tube :)
but the master of the skeptics is Ed Zitron, his position may be considered controversial but for what I understand is mostly based.
No one is against AI in principle (much less myself) but the way it has been promoted is quite frankly unrealistic, there is simply no power to generate the 'compute' needed, there is no business model to make it profitable unless as side project of major corporations, the capital expenditure is insane.
On the coding side I know very little, but I understand the concept of cognitive debt , by many considered as an overload for companies rather than time or cost saving. So the story that you can buy tokens to have agents to replace your juniors doesn't stand on its own legs, at least at the moment or in a short term.
Lastly, misusing AI turns us into the humans from the movie Wall-E — lazy, dependent, and so out of shape that we can’t even walk by ourselves. But it's an excellent tool to automate certain processes already, so I am positive ;)
@brunopostle
good read, thanks
on cooling the situation is apparently less dire than power consumption, closed-loop technology now being introduced, saves a lot of direct use of fresh water, the elephant in the room in my opinion is power consumption, after years (decades) of promoting energy saving procedures AI received a blank check with little to none public consideration.
If put together with the current state of power grids and the massive electrification exercise run in many countries, including increased use of EV, it's not difficult to see the limits of what proposed as "the solution"
Generally speaking, a rule i use to decide if i can use AI generated stuff is asking myself "Am i understanding what it is done and i can modify/manage it?"
Some observations about the practicalities of open source development (ignoring the environment for the moment): I am utterly convinced that right now, left by itself, AI will create broken software. The more it produces without guidance, the higher the risk of killing the software.
The skill for developers is now I feel to "catch up" to AI. I feel like I've got an eager hound. It's so easy to press the enter key and have it "just do things I tell it". And it can write faster than I can read. It can give the illusion of understanding faster than I can I can truly understand. So that's the real challenge ... can I read fast enough to keep up, and not just skim-read, but truly digest-read to see the software patterns and connections in my head. Because I'm not so much worried about syntax (or nitpicking opinionated dev stuff anymore), I'm worried about the "shape" of the code. The "smell" of the code. The "intent" of the code.
I now spend time just ... reading code. Reading and talking to myself and AI. I use AI as my reading companion - I share my screen and talk live with Gemini and read the code aloud in my own words and the AI either confirms or argues my point, and I argue back at it. It's like continuous rubber ducking with a PR generator. It's these reading sessions (and the AI debates that happen after I reach some conclusion) that steer the next round of code generation. I'm still struggling with it, and it's a constant balance between steering and losing control.
... but let's imagine it's doable. Let's imagine we are able to "work" in this strange new form of "just enough" fact-checking and micromanagement of AI output.
The next question is how on earth does this work in FOSS? The datamodel branch with the work @aothms has been doing (which recently merged in the IfcViewer stuff up to a checkpoint) is ... believe it or not ... +411,393 -365,768. Read that number again.
Four. Hundred. Thousand. Lines. Changed.
... and this isn't by a non-coder burning tokens without knowing what's under the hood. This is (mostly) by Thomas. He knows what he's doing. We all trust him to do right by IfcOpenShell.
And in the past three days, another. Eight. Thousand. Lines. Modified.
I'm sure those lines of code fit in Thomas' head. And the thousands of lines that I've AI-produced fit in my head. Probably not as deeply digested as the Bonsai code I wrote by hand, but I know what each file does and contains (apart from some of the more mathematical algorithms which I just trusted AI).
How can the rest of the community keep up? In the past, if a company "released" waterfall-style code to open-source without the slow drip-feeding of commits, we would never consider it truly open-source because there was no way for the community to comprehend it. But such is the nature of ultra-fast code generation now. What does comprehension mean? My best guess right now is that the volume of code is "shaped" and verified by the developer to reflect the true developer's intent, rather than "oh the AI generated it and I assume the AI is on the same page as what I envisioned but I never checked". Because now the code is the source of truth, and the only way to make sense is to read the code (perhaps with AI assistance). But is that collaboration? If you want to collaborate, you do the whole read and mass build ... but how do the core devs build trust with you as a contributor? In the past, they'd have to review (small) PRs ... but maybe that makes no sense anymore? Maybe we need to go back to actual voice calls and discussions and talks about code structure. Little "show and tells" about this is the shape of the code and why it is that shape.
... dunno just braindumping.
@aothms be like...
;)
So just regarding that number. That also includes generated code for the schemas. And I also asked AI to apply consistent naming conventions.
I think I would like to work towards clear levels of understanding: use cases, modules, data models, function calls. I'm beginning to feel ok if I don't read every function body given there are three things: (a) clear data models with explicit intent (I prefer to write them myself, AI is either pretty bad or my taste is atypical) and (b) clear invariants, pre and post conditions to what the function needs and can be expected after completion (c) tests. With this structure in mind I feel that I can still work top-down instead of fully open-ended AI slop. If, on the other hand, I start to actually read all function bodies I start fiddling with the horrendous verbosity and probably could have better done it myself from scratch.
And the same applies to the community, for most their level of understanding will stop at use cases and that's fine. How many lines of curl, grep or inkscape have I read? For review it'd be good if we can focus on data models and function invariants, pre- and post conditions - and the intended scope of the feature narrated by tests.
Completely agree - I think the single biggest thing that will keep projects alive in the AI age is strong tests and clear data models. Really strong definitions and naming and code structure.
Also just to emphasize for others reading along 400k shouldn't be taken at face value - but even if it was 95% generated, 20k is still significant, and even if that was all non-logic, simple changes, like spelling errors for example, 20k is still a lot of mental throughput or sheer scrolling and mental categorisation.
As for the financial part it is worth to read some pieces by tech & financial covering journalists, one example here: https://bsky.app/profile/edzitron.com/post/3mlvw4gyh4c2v.
If you have lived through the DotCom bubble and the Subprime crisis, you will see the same patterns repeatedly.