Conversation

Notices

1. Embed this notice
   Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 00:39:18 JST Christine Lemmer-Webber

   RE: https://mastodon.online/@mastodonmigration/116312883173526888

   While there are AI enthusiasts and skeptics on both networks, I do think it's the case that the general vibes and views on AI tend towards the opposite on both networks, particularly in terms of leadership, and especially in terms of a sense of whether or not AI tooling is empowering or not.

   While my views are more complicated than just good/bad, I'm happier and more comfortable with the fediverse's challenging of the current hype personally.

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 00:43:07 JST Christine Lemmer-Webber

     in reply to

     I think something that's worth highlighting is that both communities are concerned with empowerment and disempowerment. And I tend to think these tools tend to *appear* empowering, but are actually disempowering, in their current configuration.

     I don't believe LLMs are fundamentally disempowering, and could be part of an empowering future, but the present *industrial deployment* of AI tech within our *socio-economic environment* is net-disempowering. And I worry that there is a big rush to adopt with so little settled about the legal implications on the one side, and with *well known* problems for AI generated code on the other.

     Not all AI coding usage is necessarily doomed to be a problem: using local models to "lint" or discover vulnerabilities/bugs is actually probably very good, in the way that having fuzzers is good. But there is so much pressure to adopt beyond the space of what's good and to dismiss real concerns that I am worried it is going to take a long time to undo damage

   • Embed this notice
     Evan Prodromou (evan@cosocial.ca)'s status on Monday, 30-Mar-2026 00:46:02 JST Evan Prodromou

     in reply to

     @cwebber this is a much more nuanced position than I thought you had.

     Christine Lemmer-Webber repeated this.
   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 00:48:31 JST Christine Lemmer-Webber

     in reply to

     Let's make the "empowerment" claim, that is somewhat true, here: it's true that most people experience computers as black boxes, and for a long time it has been the case that only the "hacker class" has access to "sieze the means of computation", as it were.

     I think the "everyone can instruct their computer to do the kinds of things they would *like* their computer to help them with" is actually a good goal.

     That said:

     - Current tools appear to do so, but there are so many gotchas in terms of how they operate, and when things go badly, users are stuck.
     - The approach is incomplete. I believe to successfully achieve the goal requires a *combination* of machine learning and constraint solver techniques. Without this, contemporary AI tech tends towards high amounts of dangerous failure cases.
     - We need to be very concerned about deskilling and where expertise develops from.
     - There are a lot of opportunities to improve that *aren't just LLMs*!
     - AI industry power dynamics are very bad.

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 00:49:17 JST Christine Lemmer-Webber

     in reply to

     • Evan Prodromou

     @evan Part of the reason I have a cassandra complex is mY taKeS aRe vEry nUanCeD

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 00:52:18 JST Christine Lemmer-Webber

     in reply to

     • Evan Prodromou

     @evan and that means it takes a ton of effort usually to get to what I am actually arguing, and the hot takes version is usually much easier to absorb, and I am not good at distilling, so you often have to wade through a 30 page blogpost until you get to what I am actually arguing

     or maybe I am just full of myself, that too

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 01:08:28 JST Christine Lemmer-Webber

     in reply to

     • morgan

     @fay Yeah, a real concern!

   • Embed this notice
     morgan (fay@lingo.lol)'s status on Monday, 30-Mar-2026 01:08:29 JST morgan

     in reply to

     @cwebber sure, but it's not possible to get at a point where you can stop training!

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 01:08:31 JST Christine Lemmer-Webber

     in reply to

     • morgan

     @fay It's a good criticism. However it's also true that the energy costs and computational power concentration risks of training a model once are very different vs using them once you have one which is trained.

   • Embed this notice
     morgan (fay@lingo.lol)'s status on Monday, 30-Mar-2026 01:08:32 JST morgan

     in reply to

     @cwebber imo "local models" are a red herring: they still require the industrial machine for pretraining (including licence laundering, power consumption etc) even if they're less wasteful at inference time, and the need for pretraining will never stop. It's far cheaper and efficient to develop deterministic tooling and good documentation (which llms would also need anyway)

   • Embed this notice
     Evan Prodromou (evan@cosocial.ca)'s status on Monday, 30-Mar-2026 01:13:09 JST Evan Prodromou

     in reply to

     @cwebber no shame in any of those.

     I really enjoy your posts on AI, btw! Always interesting reading.

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 02:41:57 JST Christine Lemmer-Webber

     in reply to

     • Erin 💽✨
     • morgan

     @erincandescent @fay Yes, that's the problem. The big AI companies are *hoping* that the output is considered a copyright dervative of its inputs because *they are the only ones who can play that game*. They can make license deals with the big media companies AND everywhere you've been coerced into signing your rights over to with a ToS. And in fact THEY HAVE BEEN making these deals, in the hope this will be the result!

     Trying to fight AI models on copyright grounds is a losing battle for us and a winning battle for big AI corporations, unfortunately.

     That said *we simply don't know* how things will go, so *don't incorporate AI output into your projects today*!

   • Embed this notice
     Erin 💽✨ (erincandescent@akko.erincandescent.net)'s status on Monday, 30-Mar-2026 02:42:01 JST Erin 💽✨

     in reply to

     • morgan

     @fay @cwebber I think most of the opposition I’ve seen will only result in net negatives.

     I have no issues with the idea that the output of a model can be a derivative work of the training data. On the other hand, if a model is a derivative work of the training data (and this is not fair use or local legal equiv.), the only models will be closed models, and that’s a significantly worse world.

   • Embed this notice
     morgan (fay@lingo.lol)'s status on Monday, 30-Mar-2026 02:42:02 JST morgan

     in reply to

     • Erin 💽✨

     @erincandescent @cwebber in principle, in an ideal society yeah sure. We're not there and resistance to these companies and respect for the consent of the authors of the ingested data will always come first

   • Embed this notice
     Erin 💽✨ (erincandescent@akko.erincandescent.net)'s status on Monday, 30-Mar-2026 02:42:03 JST Erin 💽✨

     in reply to

     • morgan
     @fay @cwebber we'll have to see how courts decide upon this, but ethically... I lean more towards copyright abolition than anything else.
     
     Compiling the sum of all human knowledge into a big but surprisingly not that big file is kinda cool and it's a shame that the US companies doing so don't release their versions of said file.
   • Embed this notice
     morgan (fay@lingo.lol)'s status on Monday, 30-Mar-2026 02:42:05 JST morgan

     in reply to

     • Erin 💽✨

     @erincandescent maybe! even then it doesn't solve the issue of data misuse/licence laundering, deskilling etc etc
     @cwebber

   • Embed this notice
     Erin 💽✨ (erincandescent@akko.erincandescent.net)'s status on Monday, 30-Mar-2026 02:42:06 JST Erin 💽✨

     in reply to

     • morgan
     @cwebber @fay I feel like you could dramatically reduce training intensity and a lot of it right now is driven by the fact that the leading labs have an existential need to stay in the lead without algorithmic breakthroughs
   • Embed this notice
     Erin 💽✨ (erincandescent@akko.erincandescent.net)'s status on Monday, 30-Mar-2026 03:02:05 JST Erin 💽✨

     in reply to

     • morgan

     @cwebber @fay Right. If you are OpenAI/Anthropic/XAI/et al, the ideal outcome of everything for you is

     • Models are definitively derivative works without fair use exemption, because then all models will be closed and everyone will be forced to pay you to use theirs.
     • Outputs can be (but are not necessarily) derivative works, because you can do deals with the people most likely to sue and with the information brokers you purchase training data from

     I think we all prefer a world where, if you want to use a model (LLM or otherwise; I’m sure we’ll see better systems than LLMs evolve over time), you have the choice of running it on your own hardware, where all your inputs and outputs can be private and not exposed to a handful of massive corporations.

   • Embed this notice
     Christine Lemmer-Webber (cwebber@social.coop)'s status on Monday, 30-Mar-2026 03:03:24 JST Christine Lemmer-Webber

     in reply to

     • Erin 💽✨
     • morgan

     @erincandescent @fay See that's the thing, I agree with you, and yet I'm still not advocating that people use these models right now!

     I think it's hard to talk about these things in this kind of detail because it's much easier to just pin someone in a boolean camp. And no wonder! We all have too much information to synthesize.