Re: Is AI-generated code changing free software?

[Danny Spitzberg]

   I've been doing applied research on auditing black-box algorithms and
   AI systems with a team at Princeton and MIT and CU Boulder, backed by
   Mozilla (say what you will about them and their 2024 layoffs,
   restructuring, and turn towards AI -- and in general). One of my
   teammates, Samantha Dalal (CU Boulder), interviewed Sayash Kapoor
   (Princeton) for a podcast about his popular writing in AISnakeOil.com.
   You might enjoy listening -
   [1]https://kgnu.org/looks-like-new-what-is-artificial-intelligence-capa
   ble-of/

   And to get at the heart of the question here about AI and code, Sayash
   cites a presentation (see slides at
   [2]https://www.cs.princeton.edu/~arvindn/talks/MIT-STS-AI-snakeoil.pdf)
   from Nov. 2019 (almost six years ago!), three categories of use for AI
   their relative accuracy:
   1. Perception, e.g. image or music recognition like Shazam — genuine
   progress
   2. Automating judgement, e.g. spam filtering — imperfect, but improving
   3. Predicting social outcomes, e.g. prison recidivism — fundamentally
   dubious, "no matter how much data you throw at it"
   The presentation concluded with the following takeaway (with my
   comments):
   * AI excels at some tasks, but can't predict social outcomes
   (especially for racial and economic justice, let alone technical skill
   and ability in the form of writing and reviewing code).
   * We must resist the enormous commercial interests that aim to
   obfuscate this fact (hence, FLOSS - but more, of course, as evidence by
   this list).
   * In most cases, manual scoring rules are just as accurate, far more
   transparent, and worth considering (yay for the enduring human spirit
   and dignified work).
   I hope this helps!
   FLOSS is undead, long live FLOSS!

   On Thu, Jul 17, 2025 at 2:58 PM Jean Louis <bugs@gnu.support> wrote:

     * Akira Urushibata <[3]afu@wta.att.ne.jp> [2025-07-09 18:36]:
     > Maybe I'm not looking hard enough.  If anybody can point out to
     good
     > examples of AI usage in free software development, I'd like to
     know.
     Genki. You are definitely not looking hard enough.
     Many open-source AI tools like TensorFlow and PyTorch are widely
     used
     for machine learning projects, offering powerful frameworks for
     model
     training and deployment.
     If you're looking for real-world applications, AI is used in
     automated
     documentation, bug detection, and even deployment optimization—all
     with freely available tools.
     > I believe that if AI is capable of code generation, it should be
     up
     > to other, related tasks.  Code inspection is one of them.  In many
     > regards inspecting existing code is easier than creating code from
     > scratch.  It is strange that we hear a lot about AI code
     generation
     > and little about AI code inspection.
     Because you are not asking people, and this mailing list is
     deserted. NVIDIA GeForce RTX 3090 on my computer helps me to do
     that,
     it is daily work, daily reviews, reports, many people involved, I am
     helped and assisted with new Large Language Model (LLM)
     technologies. I am doing code reviews daily.
     Keep asking people. Maybe you ask on Reddit if people use LLMs for
     code reviews.
     Code generation is often followed by code reviews, it is natural.
     Let's not call it "AI" as there is no intelligence, and let us not
     anthropomorphize computers.
     ┌────────────┬─────────────┬──────────────────────┐
     │    Date    │ Total Words │ Average Spoken Pages │
     ├────────────┼─────────────┼──────────────────────┤
     │ 2025-07-17 │         889 │                 1.78 │
     │ 2025-07-16 │        3460 │                 6.92 │
     │ 2025-07-15 │        3329 │                 6.66 │
     │ 2025-07-14 │        1209 │                 2.42 │
     │ 2025-07-13 │         537 │                 1.07 │
     │ 2025-07-12 │        6161 │                12.32 │
     Large Language Model (LLM) may be used for image recognition,
     analysis, talking and transcription like in this case above,
     designing
     and many other uses.
     Before anything get yourself a GPU and install some well working
     LLM,
     and then we can be on the subject. Start using it.
     Jean Louis
     _______________________________________________
     libreplanet-discuss mailing list
     [4]libreplanet-discuss@libreplanet.org
     [5]https://lists.libreplanet.org/mailman/listinfo/libreplanet-discus
     s

References

   1. 
https://kgnu.org/looks-like-new-what-is-artificial-intelligence-capable-of/
   2. https://www.cs.princeton.edu/~arvindn/talks/MIT-STS-AI-snakeoil.pdf
   3. mailto:afu@wta.att.ne.jp
   4. mailto:libreplanet-discuss@libreplanet.org
   5. https://lists.libreplanet.org/mailman/listinfo/libreplanet-discuss