Clojurians Log v2

Better but I dislike the co-worker analogy. With a co-worker you get an idea if their work is likely to be solid or not, where gaps may be etc. With LLMs you get well formatted code with hidden bugs, misunderstanding of spec and so on. It’s more than occasional disagreement or a bit of handholding. You have to assume it will do something wrong and be prepared to interrogate it, prove it works as you intended. Imagine having a co-worker who regularly tells you: “this is definitely the answer” … “you’re absolutely right! I made that up based on probabilities”.

Yeah, completely agree

Oh, and coworkers are human beings

I... can't agree? > Programming has barely been a professional career for a generation Like... WHAT???. What the hell does that even mean? > barely five minutes ago that programmers were being treated like precious priests Okay, DHH, decide yourself: are we "barely a career for a generation" or "precious priests five minutes ago"?

'whose choices you might occasionally disagree with' - this is terribly stupid take. LLMs don't make choices, they do not have agency. it's like saying google and stackoverflow, or technical books make technical choices. they don't. you ask for something, they provide some information, valuable or not. you choose whether to use this information or not.

People do all kinds of strange things, all the time, which for others might seem odd or even crazy. But thanks for your reply, I'm buying your arguments.

and while I often use this information handling consequences is on me. without responsibility there's no agency

I did originally read it as "DHH is just a tool" which is probably closer to reality

I think both Linus and DHH have the wrong take. I also used to explain it to people as a well-read and enthusiastic but inexperienced intern. I like Rich Hickey's "automation" better.

I could see an argument that if I care about the concerns, I should be addressing them. I currently feel strongly enough about the concerns that I've chosen not to participate. That LLM/agent advocates aren't addressing them or solving them seems analogous to a factory dumping hazardous chemicals and saying, "well, we could do something about this (but we're not actually doing anything about it)" and that it's left to those who do care to clean it up. And the factory owners getting upset when those who care try to shut the factory down or at least enforce environmental regulations.

Many of these concerns can be addressed by using an open weight model on the cheap end of the pareto curve. The cost will be low. The cost won't be subsidized. You'll likely be able to use it, or a better open weight model indefinitely even if big ai firms go out of business. Some of the other concerns about data centers existing or about fossil fuels are not new concerns if you're working in tech. If you're using an efficient model the scale of the concerns should not bother you any more than before llms existed.

> open weight model on the cheap end of the pareto curve How did these models get created? How are those open weight models going to be improved in the future?

I think it's useful to separate conversations about the personal impacts of you using llms from the societal impacts of many other people using llms, btw.

> Some of the other concerns about data centers existing or about fossil fuels are not new concerns if you're working in tech. They're new in the sense of scale. Often attributed to Napoleon, "quantity has a quality all its own".

I'll throw two more issues on the fire • Anthropic and OpenAI are largely run/staffed by TESCREAL folks. I don't like to discuss religion (even a defacto religion) but it seems that their beliefs are leading to problems (ie HuggingFace hacks). • Jobs: If LLMs get better - Ben will likely lose his job. If they get worse, he might have been deskilled enough that he'll also lose his job. (As someone who somewhat agrees with Schumpeter's creative destruction, I admit this is one is confounding, but I dislike that we'd going all in on LLM companies without a plan.)

> think it's useful to separate conversations about the personal impacts of you using llms from the societal impacts of many other people using llms, btw If by that you mean separating what I can do personally by what needs to be done at the societal level, I agree with that only in as far as there's only so much one can do on one's own. In no way to do I accept that as an excuse to ignore the societal implications of my use or to throw up one's hands and say "well, what can one person do?"

> Anthropic and OpenAI are largely run/staffed by TESCREAL folks Yeah, that's pretty screwy. I do find it problematic when that fact is used to discount some of the issues raised by so-called "doomers" (such as the alignment problem). That, combined with conflation of AGI/ASI with LLMs (both due to ignorance and due to marketing) makes such discussions fraught. And regardless of their beliefs, their actions and impacts are what concern me first and foremost.

@ben.grabow - I'm sure you know that the open weight models were created by Chinese labs and (perhaps heavily) subsidized. On top of this, they are at least partially trained based on Anthropic and OpenAI's models. When the bubble bursts, we'll still be left with these open weight models, and (at least for a while) they could probably be retrained by enthusiasts or companies. It's not clear how long these will last.

@grzm - yeah. I'm happy to let them believe whatever. But I'm concerned when they let open-ended looping agents run forever with unmonitored access to the internet run hacking scripts because (allegedly) they think that's a way to create a machine-god.

(We're moving away from code - but using the models implicitly supports this stuff)

they think that's a way to create a machine-god

> but using the models implicitly supports this stuff) 100% agreed. Thinking of use of LLMs for code and ignoring the rest ignores the fact that this is all happening now in the real world. It's not some late night sophomore dorm room conversation.

Design systems which works with low end & local models

I'm not convinced by deskilling personally. The more I use LLMs the more I think about programming

> fossil fuels Also, it was only a few years ago that the hyperscalers (AWS, Google, etc...) were falling all over each other to say how they were moving to renewables, wanted to be carbon neutral, etc... They moved away from this quickly. If you were in tech back then, you might have thought that you were on the glide path to a better world.

> Design systems which works with low end & local models @sia.mohammady66 Can you elaborate? What does this mean? Is it being doing now?

It is simple, your AI system should work with local-ai models which run with just normal GPU, like 28b or 13b models

@rough.train4155 You saw the study in the talk? There was another one that came out recently too. The people involved believe they are getting better, but when tested were getting worse.

If you use LLMs to plug into your way of working it doesn’t matter if the provider changes / even goes away. I’d be doing the same stuff after even if I go back to writing more code.

@rough.train4155 I've seen in myself deskilling when not forced to actually think through an implementation, and have seen how it's harder to think critically about something I'm presented with as opposed to something I've written (see https://www.laws-of-software.com/laws/kernighan/), so I'm on the side of "deskilling is a real concern until shown otherwise". Or maybe I'm just a bad software developer.

@grzm You're not a bad software developer. You're more self-aware than many. Metacognition is also a skill that must be practiced.

> see https://www.laws-of-software.com/laws/kernighan/ I think this applies to code reviews too, something which is compounded by the volume of code coming from some LLM PRs

@sia.mohammady66 Again, is this happening now? Do you see the current AI companies doing the type of work necessary to be able to run low-power models locally? How are these models developed? Are you working on this? Do you see other groups working on this? Do you think they'll succeed? (As for "It is simple", what appears to be simple to one person may not appear simple to others.)

@grzm I do not get your point, local model already exists and have different vendors both big companies and small companies & i'm using them as well. Also converting an AI workflow which is working with high end model to work with low end model is not simple, it is really hard and i faced it, how ever the overall concept is simple

Yet another point that I forgot earlier (US centric) • the previous job I was at, I was forbidden to use LLMs to generate production code. Why? Because legally, LLM output can't be copywritten. "Creative" output from a machine might be a legal minefield in the future. If you work on open-source that might be less of a problem (depending on your licence).

One category of concern is the "rug pull" problem, where one invests in a way of working now that is taken away later. If you use open weight models and open source tools this problem is a non-issue. You will be able to continue to rely on your methods even if model development screeches to a halt. It's a different category of concern to wonder about the trajectory of model development and open access in the future.

@sia.mohammady66 - are you talking about local models generating code for you? Or are they running as part of an application's workflow? For example, I've used local vision language models to identify wildlife in photos. Works OK. I think local models work OK with Immich. The frontier models work much better but the local models are livable. IMO those uses of local models are part of the silver lining. We may all agree here? I found that smaller language models (ie Qwen) are OK with generating boilerplate code, but make many more mistakes than the large models do. (I've only used them through ollama though - maybe other paths are better somehow?)

I did not use any local model for code generation about code generation there are tips and tricks, whether your model is local or not, some kind of projects are good to be code generated, some others are not Wether your project is simple crud or a complex system? if your project is easy kind project & you have guideline of writing code there is no problem on using coding agents

One category of concern is the "rug pull" problem,

It addresses problems of the form "If I use AI personally today, and it's not available tomorrow, it's a problem". Deskilling is one example. If you use a workflow that relies on open resources you can expect they, or something better, will be available indefinitely, so I don't see a problem with deskilling. (I'm skeptical that deskilling is inevitable but even if it happens, the capabilities of AI that stand in their place will remain accessible.)

> "rug pull" problem You're right that relying on the Chinese models is better than OpenAI and Anthropic. But... consider that 1) they are partially trained from OpenAI and Anthropic models, 2) they are subsidized by the Chinese government, 3) if all companies stopped training models and we were left with only the existing open-weight models, those models would eventually drift and become less and less useful. For example - older models won't know anything about the new clojure flow library. So as new libraries are brought out, the LLM user won't be able to use them.

I understand "Deskilling" to be a decline in personal ability, not one's access to tools, though I do see in the https://en.wikipedia.org/wiki/Deskillingthat the removal of access to tools could be an interpretation of "Deskilling is the decline in working positions through the machinery or technology introduced to separate workers from the production process." The second paragraph of that entry is what I generally think of: > The term refers to a person becoming less proficient over time. I'll have to update my understanding of the term.

FWIW: deskilling as I'm using it is what was discussed around minute 15 of the talk. See also Stuart Halloway's reference to the paper: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6868618

Maybe there are two cases: ai replaces some personal capability and the skills you used before decay but overall you're as productive or more so; and using ai gives an illusion of productivity and you lose both skills and overall productivity in the process.

> https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6868618 > I can certainly understand and agree how AI use in education may be counterproductive. I'm not sure how that is related to software development by experienced professionals. The point of professional software development is to create a product. Acquiring skills along the way is a great side benefit when it happens but it's not the ultimate point. Both before AI and after, I expect to need to spend some time outside work to do intentional study and practice to build my skills since my work activity doesn't always give me the opportunity for it. When I am in a learning mode, my attitude and approach to learning is much different than what it was when I was in secondary school. I can't say for sure but I suspect the pitfalls that secondary school students fall prey to with AI use are less of a problem for me now, with the benefit of experience and discipline to endure the hard parts of learning instead of taking short cuts.

I also think there is a fundamental problem with taking conclusions from aggregate population studies or experiences and applying them to one's personal life. One of the main reasons why I engage with the clojure community and choose it for my work is because I think there are ways to execute dramatically better than the average if I find the right insights. Even if people generally make overcomplicated, slow, hard to maintain programs, that doesn't prevent me from finding a better way to do it. In the same spirit, I'm fundamentally curious about the problems people experience with AI, and curious to know if there are ways to improve upon that experience by applying the right insights and principles. I suspect this community in particular would be capable of figuring out solutions to those problems, which is why I'm interested in having this conversation instead of ignoring it.

@ben.grabow As you've alluded to a number of times in this discussion, you're interested in skepticism in a narrow scope: issues people have with using LLMs in their work as software developers. There are a number of us in #C0B7Q4XRM1U that have concerns much broader than that, and question whether the trade-offs in those larger scopes justify any gains in the narrow scope of the day-to-day software developer. I don't think there are many that believe there are no productive uses of LLMs in software development. For me personally, given the larger scopes that are currently necessary/involved with/enable the use of LLMs in software development, I'm not convinced it makes sense to further discuss how to use LLMs in the narrow scope without addressing the larger. I think you might find #C068E9L5M2Q more amenable to the discussions you're looking for?

@U0BGC50RM26 Please do share what you're doing! There are no institutions without individuals. I know I'm looking for more things to actively do.

And thank you for sharing the link to the Tech Workers Coalition. I'm hopeful that will gain enough traction to make a difference.

> But mere opinions/agreement on a Clojure slack don't affect anything but Clojureland I take your point. From my perspective, those of us who have chosen Clojure (and I think most of us have chosen Clojure as opposed to Clojure being something we were dropped into) have at least some shared values. Discussing these issues with those who already share some values can be useful (and perhaps an easier lift?) than discussing them with a random individual at large. This isn't to say discussing them only with those in the Clojure community, but that it's useful to do so here, too. (And not only as these issues affect Clojure.)

Yeah, to do this justice, it does take work. I feel ya. But I think it's worth it. (While slack has its uses, I miss the days of email groups, which encourage longer, more thoughtful discussion, in my experience. While slack is async, I still feel pressure to respond quickly, potentially at the expense of responding well.)

1. DeepSeek is currently generating revenue of around $38 per GPU hour https://wukong123.substack.com/p/reflections-on-deepseek-open-sourcing and they're already basically breaking even overall https://isaiprofitable.com/ 2. American closed models aren't the only game in town, and open models from China are performing just as well on most tasks 3. Chinese models are released in the open and can be built on using community efforts, DeepSeek even released their inference engine as open source now 4. Local models like Qwen 3.8 27b are quite capable now, Qwen performs as well as Opus 4.6 from less than a year ago. I think it's almost certain that we'll get local models that can do most tasks just fine within a few years. 5. The power consumption has now dropped dramatically, that's how DeepSeek is able to be profitable in the first place. Meanwhile, China is massively developing renewable energy, and that's where power used by data centres in China is largely coming from. 6. Psychological effects are definitely something people should consider. But LLMs aren't a unique problem in that regard either. Just look at all the negative effects of social media for example.

But LLMs aren't a unique problem in that regard either. Just look at all the negative effects of social media for example.

Agreed. Just saying this isn't the only issue in the world doesn't invalidate this particular one. The EXPONENTIAL curve of energy use coming from data centers dedicated to training and running these models is far and away the biggest signal that this is not the same thing as social media addiction. Not to mention the known cognitive decline and drug-like addiction potential. Some of my coworkers don't have an original thought of their own any more — system 2 almost completely disengaged. They are like zombies.

> DeepSeek is currently generating revenue of around $38 per GPU hour https://wukong123.substack.com/p/reflections-on-deepseek-open-sourcing and they're already basically breaking even overall That's encouraging to hear, from a financial point of view. That said, DeepSeek is a drop in the bucket compared to the overall economic picture.

> American closed models aren't the only game in town, and open models from China are performing just as well on most tasks @yogthos Could you provide references so I can learn more about this myself?

Oh this was a fun site to find… https://isaiprofitable.com/

https://openrouter.ai/rankings#top-models Deepseek accounts for around 25% of all token usage on openrouter. It's hard to compare that against people using anthropic subscriptions directly but I wouldn't call it a drop in the bucket. Open models are very popular among people paying directly for their tokens.

Looking at https://isaiprofitable.com/, DeepSeek has spend/revenue of $0.4B/$0.2B (which, at 50%, doesn't actually look that close to "break even" to me, but whatever). Compare that with OpenAI ($71.8B/$32.7B), xAI ($21B/$8.8B), or Anthropic ($33B/$22B). From an economic perspective, I'll stand by "drop in the bucket". And one of my concerns is exactly this economic perspective. This amount of investment on so little return threatens not only these businesses themselves, but a large part of the economy.

Sort of like https://www.reuters.com/world/asia-pacific/imf-chief-warns-energy-shock-growing-debt-ai-risks-threaten-global-growth-2026-10-07/

@grzm it's not that I only have interest in one aspect of AI skepticism. There are several different aspects that have essentially orthogonal problem spaces. Whether AI is useful on a personal level is very decoupled from things like whether models are open or closed, and decoupled from what energy modality is used to power the hardware. There are a myriad of interesting questions to consider independently. The unfortunate pattern I see in this channel and other "skeptic" spaces is that conversation about one particular aspect will get derailed by interjections about some other concern that has no direct bearing on the topic at hand. Lots of people seem more interested in litigating the question of whatever AI is bad/evil in general than they are in answering questions or making productive progress on any one particular aspect of AI. The discourse is just terribly disordered and unproductive. The main thing it's accomplishing is drawing battle lines to determine who is on which "side" of AI use. I have no interest in that kind of tribalism, on either side.

@grzm I also want to take a small aside from the conversation and say I believe we met at the conj last year. I hope you're doing well and I appreciate your effort to engage and move the conversation forward.

@grzm the official page is a good overview https://www.kimi.ai/blog/kimi-k3-1-pro and here's a fun read on DeepSeek beating Sol https://flowtivity.ai/blog/deepseek-v4-1-flash-benchmarks/

And whether DeepSeek is a drop in a buck in the US economy isn't really that interesting of a question to me. They're used massively in Chinese ecosystem, and they are actually focusing both on efficiency and being actually profitable. Since they do their stuff in the open that also paves the way for community efforts and open models. There are already efforts like this one https://github.com/antirez/ds4 to partially quantize large models keeping reasoning parts at high quants, and routing ones at lower ones. So, there's a very good chance we'll see very capable models that can run locally in the near future. I think that the AI bubble in the US will pop within a year or two, and there is going to be massive fallout from that. But the really interesting tech will always be developed in the open, and that's the stuff I'm interested in. It's also crazy to see the amount spending on this tech in the US compared to China given that the difference in outcomes is negilible. https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report

The mere fact that Chinese companies spent a tiny fraction of what companies like OpenAI and Anthropic spent means they have a far better chance of actually becoming profitable.

I will just add that, you're in a tribe whether you like it or not. Those of us in the anti-ai tribe are absolutely enemies to those who are pro-ai, as they stand in the class of people that are attempting to literally end us. And so, the ability to ignore the fact that this is happening comes from a place of privilege. Those of us who don't have the privilege are literally fighting for our lives. This will make any kind of “pure” conversation difficult, as per maslow’s hierarchy. Sorry about that, but .. that is, as they say, where we are at.

@stephen676 I'm not sure I follow. Was this in response into something in particular?

I was responding to Ben’s message above about tribalism. I have several friends who have been unemployed for over a year because they don't embrace AI. My current boss has threatened everyone that they will be left behind, and that layoffs are imminent next year, even as he is increasing the AI budget — it's an insane percentage of costs — 10% of salaries. If I don't let myself get infected I will likely be forced to retire very early. Probably next year. We likely won't be able to keep our house. And I support my husband’s parents and my Mom. Unless I can figure something out that is. It feels quite scary.

Thanks. I missed the explicit tribalism reference.

If you feel like you're in a fight for survival I can understand why you'd see everything as a tribal war. I wish more people were so forthright in describing their perspective in those terms. I don't think it's very healthy for discourse though.

I don't think it's very healthy for discourse though.

I expect them not to insist I'm in an enemy tribe, despite me making an effort to engage productively, as a start.

This is another reason why I think it's healthy to segment the conversation somewhat. A thread about difficult job situations and how to cope with them, both immediately and as a political topic long term wrt AI use is a good, healthy thing to have. Having a conversation about difficulties getting something personally useful out of AI, and how to improve the process is also a good conversation to have. I don't think it's very useful for either goal to mix these conversations at every opportunity.

Also worth noting that a lot of social problems around AI are a product of power dynamics in western societies. For example, the issue of AI related layoffs can be easily addressed as has been done in China https://globalnews.ca/news/11840683/ai-china-layoffs-court-ruling-canada/

how this new form of automation is used is merely a symptom of the underlying problem, and it seems to me that's what would be more productive to focus on

@yogthos I agree that a lot of the issues with LLMs currently in the US are a result of the current economic structure. I would encourage you to try to convince those who are pro-LLM or lean-LLM to think similarly, and actually act on it. (Though, I don't think this is a topic @ben.grabow would like to discuss, at least not in this thread.)

Well here's one thing I'll say having thought about this a bit over the past year. It's pretty clear this tech is now mainstream, and boycotting isn't going to make it go away. There's just too much money being poured into it for that to work. And empirical research shows that when public opinion goes against moneyed interests, the latter tend to win: https://www.cambridge.org/core/journals/perspectives-on-politics/article/abs/testing-theories-of-american-politics-elites-interest-groups-and-average-citizens/62327F513959D0A304D4893B382B992Bhttps://www.cambridge.org/core/journals/perspectives-on-politics/article/abs/testing-theories-of-american-politics-elites-interest-groups-and-average-citizens/62327F513959D0A304D4893B382B992B The question is then what are these companies are actually afraid of? And they're telling us this openly. They fear open models and models that people can run on locally on their machines. That's the actual threat to their business model. So, to me the solution is to develop this tech in the open, to have community efforts to train and optimize models for local use, and to wrestle control of this tech away from these companies. We already have precedent for this with open source and stuff like Linux displacing proprietary systems like Windows.

@ben.grabow on the topic of LLMS in professional software development, when people have raised deskilling, you've dismissed it. When they've raised how we develop new software developers (i.e., education), you've dismissed it as not part of the scope of the discussion you're interested in having (and dismissed the paper as inappropriate because it deals with aggregates as opposed to individual experience). I'm not sure where to go from there, or what discussion you're interested in having.

And again, I don't think there are people here arguing that there are no benefits of LLMs for software development. It's which tradeoffs people are willing to make.

Quotes? My goal here is mostly on the meta conversation. If I'm convinced that the non-meta conversation is possible I'm happy to engage. This isn't the thread for it imo.

The paper is about secondary school students. It has its place, especially if we're talking about education. I don't think it's sufficient evidence that deskilling is inevitable for professional software developers. I think it's a real problem to consider and I think there are ways to address it. However again I think this is more of a meta conversation about the discourse around AI. We should have a productive conversation about deskilling and how to avoid it if possible in another thread.

https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792944591299619?thread_ts=1792937638.074069&cid=C0B7Q4XRM1U: https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792944591299619?thread_ts=1792937638.074069&cid=C0B7Q4XRM1U https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792942434857339?thread_ts=1792937638.074069&cid=C0B7Q4XRM1U: https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792942434857339?thread_ts=1792937638.074069&cid=C0B7Q4XRM1U

Mostly the goal I have is to encourage conversations to target a specific aspect of AI and stay on topic with the goal of an actionable outcome that improves our lives.

The above quotes are hopefully invitations for further discussion. I don't intend to dismiss the conversation, only to suggest that there may be a good path forward that doesn't involve avoiding AI for fear of deskilling.

If/when I have a conversation about the political/societal aspects of AI, with respect, I think this forum has too narrow a scope (clojure) to be worth my engagement on that topic. I'd rather spend my breath on topics related to the specific relationship of clojure and associated design sensibilities with AI. How to use functional programming, data orientation, hammock-informed design, etc to get better results with AI.

@ben.grabow I'll reiterate: I think you're probably looking for #C068E9L5M2Q ("A general channel for coding with the help of AI. Things like ChatGPT, own LLMs, AI in the editor, Copilot, etc and whatever.") The description of #C0B7Q4XRM1U is "A place to discuss concerns and skepticism regarding the use of AI/LLMs, and AI assisted coding. Code of Conduct is in effect, keep it civil. For discussions about workflows and techniques, use #C068E9L5M2Q."

I'm well aware that channel exists. I'm here because you at-mentioned me, and I'm originally here because this is where Stu's talk was posted, which affects me via the clojure community much more than other discussions in this channel do. I don't quite know what your intent is by trying to direct my voice away from this channel (or this thread?).

My only purpose of mentioning #C068E9L5M2Q is that you've repeatedly mentioned you're interested in discussions about how to use AI for professional software development. I've repeated it because you said you don't think it would be useful for you to engage in https://clojurians.slack.com/archives/C0B7Q4XRM1U/p1792984453096719?thread_ts=1792937638.074069&cid=C0B7Q4XRM1U("If/when I have a conversation about the political/societal aspects of AI, with respect, I think this forum has too narrow a scope (clojure) to be worth my engagement on that topic.")

This channel (#C0B7Q4XRM1U) has been used for discussions much broader than just LLMs with respect to Clojure.

As for starting this particular thread, Slack won't let me continue the original thread, helpfully letting me know it's still "loading replies". I just checked, and it is still "loading replies" 🤔

I apologize if you feel you were dragged into a discussion you weren't interested in having. In the original thread, @john.t.richardson.dev raised a number of points, and I was interested in hearing you expand on why you thought they weren't actually issues, or at least not ones really worth being too concerned about.

@yogthos > So, to me the solution is to develop this tech in the open, to have community efforts to train and optimize models for local use, and to wrestle control of this tech away from these companies. I'm not a fan of LLMs. But, I'm glad there are pro LLM people that fight the open source fight. LLMs have been incredibly destructive to open source. But, if open source can get its revenge with open source models that would be something.

It does look like there are unintended consequences of this tech that corps didn't anticipate with people reverse engineering stuff using LLMs, rebuilding commercial apps, hacking stuff like smart TVs to get rid of ads, making custom firmware for home security cams that doesn't use company servers, and so on. So we might definitely get some revenge here in the long run. And this is also where open models will be important because it's almost certain that companies start to censors rental models. Since this tech exists, we might as well make the best of it by wrestling control away from the corps.

@yogthos Didn't I see that people publicizing that that "reverse engineer anything" skills had their Claude accounts disabled? I think both openai and anthropic disallow it now. Sigh. I thought the right to repair might get a boost. Even with open models there's the DMCA to be concerned about, if you live in the US.

@john.t.richardson.dev good news that DeepSeek, GLM, or Kimi will happily do this stuff and if you have a beefy enough machine you can run Qwen 3.8 locally as well. This project in particular works quite well https://github.com/morluto/rea Hosting in the US is definitely problematic, but it seems like projects like https://github.com/storytold/photocraft aren't getting taken down since they're clean room reimplementations, even though they mimic the UI/UX of the original apps. Maybe next step is to use https://radicle.dev/ for hosting. :)

@yogthos True enough. And I do suspect that Adobe's lawyers will look for any way to take out PhotoCraft, etc... if it gets popular enough - considering Photoshop is an "app" now - they might be able to. For those living in the US, I'd be concerned about an Aaron Swartz-type event. (Yes, I'm paranoid. 🙂 )

It's going to be interesting to see how this plays out. On the one hand you have the copyright industry which with all the media and software companies that have insane amounts of money, and on the other you've got AI companies that are moving into too big to fail territory. Their whole business model is built on ignoring copyright.

If companies set a precedent against small time projects, that can then be used to go after big players as well, so I suspect they would fight that pretty aggressively.

@yogthos I think it's different, legally. LLM companies claim that since they are only doing numeric analysis of material (somewhat similar to search engine indexing), their access is covered under fair use. My understanding is that this has held up in court. (And without this fair use, we wouldn't have the internet archive.) The small guys decompiling their HP printers, TVs, ... are in violation of Intellectual Property rights (DCMA) law (different than copyright law). I'm not a lawyer, and I think most related lawsuits have been settled out of court, but it's considered a criminal offense in the US to distribute such a tool. (Big deal compared with copyright violation - as per Enshitification.) And yeah, the small guys are easy targets. If the big companies want to fight, I wouldn't be surprised if the US took legal action stopping access to the Chinese frontier LLMs (see also: tiktok) that didn't put at least some IP guardrails up. I agree, the small models will be harder to stop.

I keep forgetting how utterly insane DMCA is. Given that you can't reverse engineer or build clean room implementations of stuff covered by it, I can definitely see how that will be problematic going forward.

Also, the Chinese announced upcoming (maybe current?) export restrictions on their frontier models. IIRC, they kinda bring them in line with the US frontier models. I wouldn't be surprised to find that it would include the stuff we're talking about, but who knows. I was actually surprised that they would bother with this, since you can download abliterated models (even Gemma, etc...). I suppose modified models aren't as good as the originals, but they might well be good enough.

As I recall, this was still just public consultation, but it could happen as well. My view is that this stuff would be ideally developed completely in the open and outside of corporate or state control. Something like this idea could be used to do distributed training for example https://github.com/bigscience-workshop/petals And existing open models could be used as a base to build on.

Ultimately, Chinese companies aren't releasing models out of goodness of their hearts. They just realize that there's no good way to differentiate your model from the competition, so it's necessarily a race to the bottom. Anything one lab can do, another figures out in a few months, and it's very easy for customers to switch, so LLMs are a generic commodity which means the pricing will go down to the cost of inference.

Companies that figure out how to make profit will either be at the top end of the stack where they build highly customized services, or hardware vendors at the bottom, like Nvidia making a killing right now selling shovels.

inference + amortized training, but yeah

It's also going to be interesting to see what happens once the bubble finally pops. It seems that investors are starting to realize these companies will never be profitable. I wonder if Anthropic or OpenAI will end up getting a government bailout in the end.

Finance people I've talked to say that if Anthropic gets a successful IPO out of this then OpenAI will do theirs earlyish next year. At that point, enough private equity and private credit (and perhaps the stock market, what with the crazy class C offerings!) will be wrapped up in this. And that might be a 2008 event, perhaps not quite as big. And that would require some sort of govt bailout to keep the financial system going. Not supposed to be quite as bad though, since the public banks won't be in too deep. My opinion is that the govt can't keep the frontier companies running, just unwind things more gracefully. The maintenance cost is too high. I want to be (somewhat) humble though. It's possible that in 2028 we'll all be hanging out in our hammocks without a care in the world because AGI will have solved enough problems to justify the expense. hammock

I'm expecting it will be a lot worse actually, and closer to doc-com bubble than 2008. The data centre buildout is a lot like the dark fiber laying that happened during the internet boom. And then when it popped a lot of the infrastructure ended up being unused. What makes this time around worse is that the economy is a lot less diversified now, so tech sector is a lot more central overall. And other sectors of the economy are also tied into it indirectly. For example, there is a big boom in construction building data centres, so when these projects get cancelled, there will be downstream effects to supporting industries. Personally, I feel very comfortable betting against any sort of AGI in the near future. It looks like LLMs are already starting to hit diminishing returns in terms of functionality. We're not seeing nearly as dramatic improvements with new releases as what was happening a few years ago. I'm also highly skeptical that LLMs alone are path to any sort of general intelligence. At best they're a piece of a much bigger puzzle.

And far more interesting things are happening in robotics because those systems have to interact with the physical world, and that requires building a model of the environment that's compatible with ours. Having a machine intelligence learn language on that basis seems like a far better approach than teaching it to string tokens together in a purely stochastic fashion.

Pretty much agree on all points. I look at things like PE10 and get worried. https://www.multpl.com/shiller-pe I'm kinda with Rodney Brooks on robotics. I think we could get some useful but focused behaviors in robotics using AI like DreamerV3. (I was in robotics back in grad school, my takeaway was that the real world is really hard!)

Yeah, real world is much trickier, but also what makes it so much more interesting. I can see how some systems will evolve to be effectively like a nervous system and reflexes, and others will do higher level planning.