Clojurians Log v2

I like he mentioning "figuring out how to use this Automation in our workflows". Also the "Cognitive surrender to AI", it is easy/tempting to delegate our thinking and loose the benefits of the flow of making something.

I think this talk is going to make dialogue about AI worse, not better. Lots of snide, unsubstantiated comments. I'm disappointed.

@aldosolorzanop thanks for sharing the specifics that you are reacting to. One of the challenges in a wide ranging talk is that we can all end up touching a different part of the elephant and thinking we have having a conversation while talking past each other.

I thought this was a clear-eyed view of the challenges developers face from LLMs. It was backed up with data*. I'm particularly concerned about deskilling and the impact on community - which he covered. Yeah, he had touch of sardonic humor, but that seemed in line with Halloway's impactful style. Listen to Ed Zitron's financial analysis if you want to watch a someone bring out the fire and brimstone! Halloway only touched on that. And he didn't mention enshitification, but I suppose there was only so much to do in half an hour. (* I've read to Cal Newport, Cory Doctorow, and Adam Becker's views so it may be that I'm familiar enough with the subject to nod along.)

I agree with @ben.grabow's implied point that substantiation is important. To that end, most of the slides for the talk have links to further information. https://github.com/stuarthalloway/presentations/blob/master/2026_Conj_Making_A_Living.pdf.

There's an incredible variety in: • models • prices • design approaches • feedback mechanisms • use cases • hosting options • freedom (as in speech) In some cases I find AI to be very good and beneficial, whereas in others it's not so helpful, or it's expensive, or it takes power out of my hands. What I would find simulating is a discussion of this variety, with curiosity for the cases where AI might be beneficial. I liked Stu's mentioning of the option to host models on one's own hardware. Much of the rest of the talk painted AI with too broad a brush. I've seen many examples of that kind of commentary and one more video saying the same thing doesn't do much to advance the conversation.

A ~year ago when I was much more pessimistic about AI, when I heard claims about wild productivity with AI my response was always "show me what you're doing". Lots of people claim wild results but either don't show their methods or results or their results are ugly and bad. I think claims of positive benefit of AI should be backed up by concrete evidence of it. I think it would be helpful if people now who find AI to not be useful to show what their experience is, what they have tried to do, what their methods were, and what their results were. That would advance the conversation and help prevent us talking past each other.

I have nothing but respect for giving a talk that

I have nothing but respect for giving a talk that will be vilified and intentionally misunderstood. But the message is important. I'm hopeful this encourages a little more independent thinking and intentionality in how we allow automation into our lives and craft. I enjoyed the heck out of it live and it was worth the re-watch.

@ben.grabow: are you saying that you find vibe-coding, from head to tail, to work well for you? Or do you use LLMs for the dull stuff - maybe creating config files and boilerplate?

The answer is "it depends". It varies based on what part of the code I care about, how long lived the code will be, who is going to interact with the code, etc. Lots of things I have tried have not worked well. Many others have worked well. I would like to share my experiences when possible, and I would like to see the same from people who don't seem to find much positive use for it.

@ben.grabow It sounds like you are in the "caution" or "trial" from the Thoughtworks Tech Radar mentioned in the talk. I don't think Halloway was saying absolutely not to use it. He's a smart guy, I'm sure he understands what it can do. But he's saying to consider the tradeoffs.

That's not how I interpreted the talk. Stu can correct me if I'm misinterpreting but he presents the "Slopfree Software Index" (with clojure as a member) with members who reject AI/generated code entirely. He says we should treat it as rude to show ai output to people. I think there are legitimate criticisms of some of the ways some people use AI but it's a leap too far to apply these criticisms to AI itself. I would love to see dialogue that wrestles more directly with the concrete problems of that work output rather than focusing just on the means by which it was produced.

@ben.grabow Thus the tradeoffs. Someone with an LLM can generate a huge PR in minutes. There is a huge push from upper management to do this because management sees more, bigger PRs as better. ("Build the heaviest airplane") How do you push back against this? You have to send backpressure back up the system. He could have said to reply to the PR sender with a more polite, "I'm sorry, but this is just too much code to review thoughtfully" (which I might do, at least at first). But his approach of using shame will stick in the sender's mind. There absolutely are problems with LLMs themselves, beyond the quality of the generated code. Some are mentioned in the talk and are completely reasonable, IMO. Those are the tradeoffs. Did you see some of these as incorrect?

Trade-off to me implies both positives and negatives to consider, but I heard nothing positive in the talk. Of course there's no obligation to say something positive but the tone was as if the overall topic had already been ligated and it's now just a matter of figuring out practical ways to avoid/undermine AI. I definitely have sympathy for people stuck in situations where AI is forced by management, either beyond its point of usefulness or regardless of its usefulness. I hope each person has the autonomy to decide for themselves exactly how much they want to use AI or any other means to accomplish their work, provided the output is good.

@ben.grabow I also hope that each person would have the autonomy to decide how they would deploy an LLM. You might pick up the book The Reverse Centaur's Guide to Life After AI to get more of an idea of what Halloway is talking about. In order to make the financials work, your boss has to fire most of your coworkers and have you constantly reviewing PRs to be sure an LLM didn't hallucinate. I see the talk more as a countervailing force. It's not like we haven't heard the benefits of LLMs. Virtually no one in business is talking about the tradeoffs, and those need to be thought through. I do use LLMs. I use them to do code reviews*. I sometimes use them to do research, both in a codebase and online. (I don't use them to generate production code anymore.) When I use them, here are some tradeoffs that weigh on me but weren't in the talk: • Tokens are heavily subsidized now. At some point they won't be. Where will I be then? • OpenAI and Anthropic (and the neoclouds) are on shaky financial ground. Where will I be if I've deskilled and they go under? • Frontier models are under the control of billionaires - many of whom have questionable political ideas/affiliations. What will that mean for my friends who might be affected? • LLMs run in data centers. They are controversial. • LLMs use quite a bit of power. Most of that power comes from burning fossil fuels. • Some people get chatbot psychosis. I don't think I'd be vulnerable, but who knows? I could go on, but some of these are a little far afield of code, and you get the idea. I hope that if the frontier model business fails that we'll still have the local models to play with. (* I have found myself arguing with the LLM about code, which might be amusing or might be concerning.)

Those are all real problems, and there are ways to minimize, address, or solve them. I wish we could talk more about that instead of just stating the problems and not exploring the best ways to make use of the technology. As an example, you can run open weight models on consumer hardware in your own home and have control over it forever without influence from any company. If you don't want to buy the hardware you can pay an independent hosting company to run an open weight model, spending less than 0.1% of your salary, and get significant benefit from it. Advances in models, advances in harnesses, and intelligent design and verification choices make hallucination insignificant. I would love to see more conversation about these aspects. If skeptics have explored all these aspects deeply and still are pessimistic about AI I would love to learn from their experience.

The Reverse Centaur's Guide to Life After AI (written by Doctorow, from the talk) agrees with you. Local models under your control, running on your computer, and used thoughtfully can be a positive force. I largely agree too. Maybe Halloway does as well? Frontier LLMs forced on you by your boss, run in the cloud, and used to generate code at a reckless pace seem likely to be a negative force. Does this seem logical?

Yeah absolutely. Frankly any time pressure imposed from above tends to worsen the outcome, AI or not. Non practitioners telling practitioners what tools/automations/methods they need to use also worsens the outcome. Although I would say where the LLM runs, cloud or local, matters less. For an equivalent workload, cloud will be more efficient in terms of energy and capital expenditure than running locally. It's less of a libre approach but probably a drop in the bucket if you're using a low cost model, compared to the non-libre aspects of running a saas in the cloud.

@ben.grabow this conversation happens to have landed while I am vacation. I will try to respond in some detail but it may be a few days. I appreciate your contributions on this thread.

Over time for me the biggest thing. Is all models have ultimately been trained on stolen information. In a world where theft is normalised the big corporations win. That combined with what LLM use does to communities, people spend more time talking to their LLMs than each other. There's no more mentoring. It's an incredibly destructive force. This is before we get into the fact that people stop sharing their code because instant copies happen without attribution, even MIT. Again theft normalised, not attributing others work normalised, etc. Are they useful. Yes. Is it worth it? For me personally no. Interestingly from a self awareness perspective, I found LLMs/agentic workflows removed constraints that I actually crave. I found I was going down crazy rabbit holes that were technically interesting but not high impact, or adding features that were useful but ultimately undermining the product. I've found I've become much more productive (less output but way higher impact) since removing them completely from my workflow. Turns out attention is what you need.

@andersmurphy I really like your point about those features. That’s probably the biggest thing I'm struggling with. Far too often I catch myself thinking "I wouldn't have thought of this, it makes sense, so it must be good." The bigger and more complex the codebase, the easier it is to fall for that. I think these are exactly the problems we have to solve if AI coding is our daily reality whether we like it or not. (About the theft stuff, I feel like that's a really complex question. Kind of like the piracy of European books in the US back in the 19th century. Or public libraries: publishers have to give them copies by law, anyone can go in and copy things, and the author gets basically nothing from each reader. But as a society we decided open access to knowledge is worth it. Ofc, big corps making billions of public data adds a whole new layer to it, but it feels more like a distribution problem?)

Yeah they are very prone to over-engineering and adding more stuff. They also love doing big co codebase stuff. Let's add five guards here, incase someone changes this code. The problem in isolation (or if you're disengaged) those changes can seem reasonable and that's assuming you're even reading the code. By the time your swarm of review agents have had a noodle with it, it will subtly be blended into the spec (and/or crafted into a perfectly convincing PR/overview comment) That's another thing I noticed, these problems affect all usage of LLMs, from simple chat to complex agent set ups.

Just want to chip in from the “other side”. I work in a company where the use of agentic coding is encouraged. I think it’s unfair to expect of me (or others in that situation) to either 1) change the company or 2) “vote with my feet” and change jobs. To be honest, I’m not sure I would want to. What troubles me though is that the discourse about agentic coding seems to me to be extremely polarized, and I find myself in the middle with almost no-one to talk to about it. I think all of the points that @andersmurphy are valid points, as are those of @ben.grabow. And I’d love to be able to have a fruitful discussion about how to solve eg the mentorship problem without throwing agentic coding out the window. I’d love to know if someone is successful in harnessing the agents such that they don’t lead you down the rabbit holes and don’t make you consider the edge cases that are not ever going to appear. Unfortunately I haven’t found a space for these discussions, and I’m fearing that @stuarthalloway’s talk doesn’t contribute to less polarization, but rather more.

See @slipset, I really don't think you're "in the middle". This thing has ramped up dramatically in, what, the last 12 months mostly? If already your attitude is that it's inevitable, we have to live with it almost everywhere, there's no room for push-back, you're not in the middle. You're pretty far in one end of the spectrum. This is precisely the time for push-back. In some cases to limit the use of LLMs, and yes in some cases to completely rule it out.

@gentkr fair I guess, because I don't really have an opinion on whether or not agentic coding is here to stay or not, nor do I see it as inherently evil. I see it as a thing (won't use the word tool) that is available to me, which seems likely to be available to me for some time, and I would therefore like to have a discussion around how to make the best of it. The reason I consider myself in the middle is that I don't think that agentic coding is all roses and unicorns nor do I think that it will replace programmers at large, just like compilers, ide’s and 4GL haven't done either.

I took a bit of a time to see this talk. It's... wow... honestly, inspiring. It's amazing to see the leaders of Clojure showing what I feel others are failing to do... As someone currently working on a tech stack that feels "corrupted" by multiple political and social problems, this really gives me hope. Honestly, I seriously needed to see this talk - more than I though.

My copy is sitting on the shelf, waiting for me to finish Giblin and Doctorow’s https://craphound.com/category/chokepoint/

@grzm What do you think of Chokepoint Capitalism so far? Is it like Enshittification (which might cover the same theme) but focused on the creative folks?

I haven't read Enshittification.

As for Chokepoint Capitalism, I've found it enlightening. It's focused on creative arts/content. The first part of the book describes issues, and the second part of the book addresses potential solutions. I find a lot of it persuasive. As an aside, I've increasingly wished for books to present the steel-manned arguments against their positions, as I know it can be difficult to see the best arguments against while being presented with one side.

@john.t.richardson.dev I finished Chokepoint Capitalism, and I’m glad I did. The last chapter (which I considered skipping because I was getting anxious to move onto another book) helped place the particulars presented in the previous chapters in a larger context. Also, the focus on content creators I found resonates with us specifically as software developers. I recommend it.

@grzm Thanks! I’m adding Chokepoint Capitalism to my list. If you find the time, you might appreciate his newer books too. Enshittification was an overview of the problems in the modern economy, Reverse Centaur is similar, but focused down on AI. The ideas have entered the zeitgeist now though, so you get enough from just skimming one of them.

BTW: The Promise of Artificial Intelligence is a slow roll. I don’t have a background in philosophy so I’m only making it through a few pages a day

@john.t.richardson.dev Enshittification is on the shelf behind me. The Reverse Centaurs Guide is on the desk next to me. https://pluralistic.net is in my RSS reader. (Yes, I’ve gone back to RSS. I miss the days where most internet discourse was blog posts.)

I’ve been planning to go back to RSS too!

Also on the short list is Elinor Ostrom’s https://www.cambridge.org/core/books/governing-the-commons/A8BB63BC4A1433A50A3FB92EDBBB97D5. I had purchased a while back, I think as part of diving deeper into Gene Kim’s devops-related work, but hadn’t read it. It came up in a conversation I had at the Conj discussing alternatives to free-market capitalism/centrally-planned Marxist socialism.

Will check it out, thanks!

Ostrom won a Nobel prize? Is this digestible by mere mortals? 🙂

Good question. I’ll consider it when I’ve worked through more of my backlog of books. 😆