In this context, AI is just a compiler that turns human-readable source code into machine-readable code. I mean, gcc does that too. Bison writes C code for you. IDEs help you edit. Spell checkers do your "look in the dictionary" job for you. We've had code-generating tools for a long time now, AI is just another one. Today's programmers are no longer involved in writing assembler, tomorrow's programmers will no longer be involved in writing C. How many of you can do low-level programming *now* ? How many people *today* write low quality programs with the existing tools?

What's different now is that AI disconnects you from the resulting logic in a way that previous tools didn't. Then again, C disconnects you from what the assembler does, just not as much. Python even more so. ESPHome "programs" are just YAML files. This trait enables more people to "write code", which is good, but increases the amount of bad code as much as it increases the good. We as a coding community need to figure out how to map our ideals and ethics to this new tool. What is "open source" if the compiler is an LLM instead of gcc? What is the source code? How do you reproduce or validate LLM-generated programs? Who will be providing the "core intelligence" that LLMs train on? There aren't many people who can write or maintain compilers any more, will we have the same problem with AI in the future? How do we recognize the effort, creativity, and authorship that goes in to "programming" with an LLM?

IMHO taking a general "for" or "against" position wrt AI/LLMs is not useful. We should take positions for or against *how* it's is used, and how we can "use it for good, not evil." Past generations thought computers were unobtainable magic. Today's generation has the Internet in its pocket. Future generations will wonder how we luddites survived without AI. It's up to us to figure out how to get there.

Besides, who doesn't want a compiler that reads the documentation and helps you write code? ;-)

I am just using a more restricted definition of distributed that you are. In a more generic sense git is also distributed. I was requiring online synchronization in my definition.

Git is not a decentralized database. It is completely independent copies of the database which can communicate to each other. That is not the same thing. A sharded database is decentralized. Git is replicated.

Differentiate decentralized development from git itself being decentralized.

This centralized vs decentralized make no sense. Run multiple copies of Scion and it will behave just like multiple copies of git. Data in git is not spread out in as decentralized way, everyone has a complete clone if they asked for it; just like running multiple copies of Scion. The way you control this is with who has push permissions. Git is only decentralized in the sense that it allows everyone to share in a controlled way.

The kernel development flow under Linus is different. It does not use 'git push'. Instead Linus pulls from his lieutenants, locally tests and inspects. And then he finally pushes to master himself. AFAIK only Linus and GregKH can push to the master Linux repo. Allowing AI to do uncontrolled pushes into a master repository is asking for trouble. There are likely stages in their development process which should be converted from push to pull steps.

My interpretation of their problem is that their CI system was doing git clone and cloning the entire repository, then checking out the head. I agree, the performance of doing that sucks. But they don't need to do that, instead clone with --depth == 1. Doing that copies over the compressed file and then the local check out decompresses them. This is a common problem in CI systems done by people unfamiliar with git details.

And good luck beating git's performance.

They don't know how to use git because they are letting claude do everything. Just use '--depth 1' in the CI clones and now git does exactly what they built.

Only if that NFT, itself, is delivered by carrier pigeon.

Visited a few days ago (wearing a "ceci n'est pas une lune" t-shirt, if course). The building compares well to the recently updated LA County Museum of Art, and has an oddly similar concept: a pod floating over the ground. There is a lot of good stuff packed into that building; we only saw half of it, if that. The ground floor gallery was underwhelming, though it had beautiful pieces. There was little descriptive material; evidently you need to bring a smartphone and... I don't know what, or I would have done it. The ground floor theater, showing a whole bench of interesting shirt films (we saw an episode of Cowboy Bebop), itself is likely to be worth visiting repeatedly. Trust me, both star wars fans and just plain arts folk will enjoy a visit.

I think the problem is that the first step is actually "Make an expensive system that's free to use, that we don't know how to sustainably pay for."

The publishing industry is demanding this via lawsuits. Their position: you can not make copies. So if the book is turned into electronic form, then the physical copy must be destroyed. Only one instance of the text can exist. This has been a fight now in multiple lawsuits where AI companies were forced into paying for a physical copy of every book they scanned.

AI can easily write the lean for any proofs they do. The problem is that lean is missing large parts of established mathematics.
My current lean project has to cite all of these externally due to missing lean support.

    Functional equations / means
    - lit_aczel_1948 — Aczél: symmetric + homogeneous means power/quasi-arithmetic mean (the CES forcing)
    - lit_aczel_1966_weighted — weighted Aczél characterization (weighted means)

    Fixed-point / topology
    - lit_brouwer_1911 — Brouwer fixed point
    - lit_cellina_approximate_selection_1969 — approximate selection (closed-graph correspondences); with Brouwer
      Kakutani
    - lit_glicksberg_1952 — Glicksberg fixed point (infinite/Bayesian games)
    - lit_berge_maximum_theorem_1959 — Berge maximum theorem (upper-hemicontinuity)

    Probability / large deviations
    - lit_sanov_1957 — Sanov / method of types (large-deviation rate = KL)
    - lit_fisher_tippett_gnedenko_1928 — extreme-value theorem (GEV limit laws)
    - lit_kolmogorov_1931_fokker_planck — Fokker–Planck diffusion equation

    Optimal transport / matching
    - lit_sinkhorn_1967 — Sinkhorn matrix scaling (entropic OT)
    - lit_lp_strong_duality_1951 — LP / transportation strong duality (Gale–Kuhn–Tucker)
    - lit_entropic_penalty_cominetti_sanmartin_1994 — entropic-penalty -convergence (T0)
    - lit_gale_shapley_1962 — deferred acceptance produces a stable matching
    - lit_gale_shapley_proposer_optimal_1962 — proposer-optimality of deferred acceptance

    Stochastic calculus / PDE
    - lit_ito_1944 — Itô's lemma
    - lit_black_scholes_pde_solution_1973 — closed-form solution of the Black–Scholes PDE
    - lit_liouville_dirichlet — Liouville/Dirichlet (harmonic-function / PDE result)

    Dynamical systems
    - lit_saddlenode_passage_time — saddle-node "bottleneck" passage time / (Strogatz/Fenichel)

... but they have separate sets of engineers, and this clarifies that engineers from both organizations will be included.

"from upstream development through production environments."
"to secure open source software at its source and across the entire supply chain."
"Share fixes upstream so that open source communities can include them in long-term maintenance."
"Upstream maintenance alongside open source community leaders;"

Please read the whole thing.

You do understand that the freedom and licenses you're defending, specifically allow others to use your work for purposes you don't agree with, right?
If you're opposed to "evil" (but legal) uses of FLOSS, you're opposed to the core values of FLOSS.

Don't ban from arvix for not proofreading something like citations on a hot result. Where do you draw the line? Drawing the line is hard to do so just push them into the secondary queue. Then they have to redeem themselves to get back into the main one.

This is so obvious, if they start banning people a second competing service will emerge.

A better idea is to simply have two submission queues. If you get banned from the high quality one you have to use the low quality queue. If they don't do that someone else is going to do it. It is a fact of life that papers are going to have AI generated writing in them going forward. Just accept this fact and make two queues.