Clojurians Log v2

@anthonykfranco I like your takes here: Prototyping vs. engineering is a much more balanced and holistic framing without elevating one over the other. Also there is already evidence that smaller, possibly self-hosted LLMs are becoming more powerful than larger ones of previous generations. That's a very welcome diversification and decentralization.

The problem with prototyping vs. engineering is that people will choose to evolve the prototype, dragging bad decisions. Regarding NL interfaces, I kind of agree, but for many tasks it’s much simpler and quicker to use a graphic UI, especially for many people that don’t like to read or write (a huge percentage of the population). Having an agent create a program for you to book a hotel seems an overkill to me. A total waste of energy too.

The tech sector doesn't always make sense from an economic point of view; it's often driven by hype. For those who remember the NoSQL boom, at the time, you "needed " to have at least one NoSQL database. Yet, it died down; most of these projects were wasted money.

Here's a good piece of speculative fiction on the topic: https://www.citriniresearch.com/p/2028gic

How about the economics (warning, long read): https://www.wheresyoured.at/the-subprime-ai-crisis-is-here/

I think all of your points are valid, but perhaps also a bit pessimistic. I think we're currently in the Wild West phase of LLM adoption, but that doesn't mean we all have to behave like cowboys. There are a lot of loud and irresponsible voices out there, so don't let them drive you crazy or doubt your good judgement. To address a few points directly: Consultancies producing code without deeper technical skills always existed to some degree. If coding agents open up more opportunities like that, then that's not necessarily a bad thing. In the short term, the economy might feel like it's a fixed sized pie, but what will likely happen is that it grows into different directions. Software that couldn't have been written before, because of monetary and time constraints might get written now. That includes the whole quality spectrum. Also the winners are not chosen yet. I think it's highly unlikely that it will only be the ones who choose to reduce headcount by churning out slop. Much more likely the ones that look for new opportunities and niches to grow into.

Fair points. I am just saying that I don't see money to be made with Gen AI and it seems to me that is zero sum game business.

One funny thing I've seen coming up is people providing their anti-slopification services. It's weirdly both hopeful and cynical at the same time.

Better dead than doing that.

Btw one thing that helped me get a better sense of it all and be more optimistic and curious about LLMs was to start experimenting with building custom harnesses/agents and using local LLMs. There are a bunch of different approaches. You can directly depend on llama.cpp, or use ollama (more convenient). Also github copilot seems to have a decent SDK and Pi is said to be quite extensible. Once I started to think about what I can do with these things, how they work and possibly how to improve their usage, I started to both get a bit more realistic sense of what they are capable of and more ideas for what could be achieved with them. I don't have any concretely useful answers as of yet though 😅

There are two markets. The first is the AI producers. In that segment you have the model labs, the hosting providers, the hardware makers, the data hoarding/aggregating companies, the tooling/plug-in makers, the wrappers that add compliance, permissions, cost management, you have the knowledge base makers, and so on. There's a lot of money to be made here, because the demand for genAI hosted models, or even on-prem/on-device is looking to be large. I think that's mostly where investment is going. The users, including end products that are looking to add genAI features or those who will leverage genAI for their work or personal use are going to pay for the above segment. The idea is they have no choice but to pay or risk losing to their competitors who do. Their margins will thin as they will pay the genAI tax. That said, this second segment is expected to be able to produce more due to genAI usage, and therefore sell more units at lower margins but still make more growth. And it's likely that some will grow bigger as if they manage to gain a lead due to their early wins from genAI they'll eat their competitors. There's also the hole that you can raise prices once that plays out, and users will be hooked on the genAI features, so your margins will go back up.

Lastly, that first segment is dominated by the US with China behind, and almost no one else to be seen. That means the second segment might be the rest of the world, and the first segment might be US and China. Therefore, it is expected that genAI will be a boost and major growth area to US and Chinese GDP.

people point out that railroads were a bubble back in the day as a way of arguing that "bubble" and "useful technology" aren't mutually exclusive, but it seems that after the bubbles popped, railroads and locomotives had a much longer useful life than individual AI models or GPUs are anticipated to have

This reddit poster has an opinion (worthwhile, IMO), and links to their references: If the AI bubble pops don't worry the banks are covered. https://www.reddit.com/r/economy/comments/1soent7/if_the_ai_bubble_pops_dont_worry_the_banks_are/ > Moody's put out a report in February 2026 specifically warning about this. They found that across just five tech companies including Meta there is $662 billion in future lease commitments that are completely invisible in their financial reporting right now because the leases have not started yet. The truth is, honestly probably impossible to tell the true scale of what these companies owe as it is being hidden with accounting magic™. [5] And, in the posts' thread, a commentator shared this: > Vanderbilt Policy Accelerator that discusses the AI financing structure(s) and what might happen after an AI crash: https://cdn.vanderbilt.edu/vu-URL/wp-content/uploads/sites/412/2026/03/23144242/After-the-AI-Crash.pdf

it’s so weird that you can actually have the AI simplify and break it down to proper commit sizes, but people just dont do it

I've often said that our industry/sector needs to bifurcate into software prototypers and software engineers. @denis.baudinot anti-slopification is really just a business’ shift from attempting to quickly find product-market fit to being able to consolidate around better architectural decisions to entrench an advantage

To me, LLM’s are a revolution in user interface. And what we are racing to is cheaper, faster, more ubiquitous models. The default interface will move to natural language for everything over time, and that will allow many more people to actually use the computers in their pockets and on their desks. After all, AI is not inventing automation; it's just making more people be able to use what machines could already do

A financial bubble is often created because no one knows what company will win in the long term in capitalizing on the new technology. Investors need to take bets, and there is even the possibility that none of the current companies that you can invest in are the ones which will capitalize, as new ones could be created later. That said, if you pick right, you're looking at huge returns. Think Amazon during the .com bubble for example. On top of that, bubbles often underestimate the time horizon for the technology to be transformative. Being right about the technology doesn't mean you'll be right about the investment.

In this particular case it seems there is no path to profitability. Too much money pouring in and not enough coming out.

@didibus, interesting comments, I just don’t see many successful use cases implemented so far, I see the potential but I also see lots of marketing/hype and, I suspect, many companies will decide to get rid of GenAI in the future because it will always have 10% of errors/inaccuracies and high costs of fixes, evolution and maintenance. Dealing with that 10% is a real pain in the ass. Some/many companies cannot afford that. I don't think we can compare to railroads and locomotives. Hosting models on-prem/on-device/your-cloud is interesting, creating your own LLM model is also doable, but you need lots of hardware resources, which are getting prohibitive. Regarding programming, if you program in other languages I see the need to use LLMs, but with Clojure you can do a lot with small/strong teams. Maybe a better approach for the long term due to maintenance issues and developer happiness (humans like solving problems). Real life is not like chess with a set of deterministic rules.

1. ultimately the moat is in proprietary data and reliability. There will definitely be deflationary pressure of SaaS overall due to the raising of the skill floor from AI, since software engineers can now maintain peripheral code for their own internal use. 2. I don’t think AI will be much useful for the feature side. There are not many use cases for something that is maybe 70-80% correct. On the operations side, there’s definitely much more leverage where human+AI can be much more efficient. However operation costs are capped (which is why people like to focus on features), the benefits are on unit economics where you bring your cost down to sell to more customers. 3. I don’t think people are factoring reliability aka ops as a major moat for AI-powered or AI-leaning companies or services. You cannot just generate massive amounts of code without expecting either the architecture or the infrastructure to break. Meeting nontechnical folk writing code with AI is very interesting because they have not internalised all the many edge cases software engineering cover when providing solutions.

https://substack.com/@kakashiii111/note/c-233603767

I'll have to disagree. It doesn't matter if the last 10% is inaccurate and difficult to fix over top what AI did. At that point you're already locked into the new development flow. Same for business use cases, business users will still be locked into having their doc initially drafted, proof read and so on with AI. Schools have already changed, so new grads only know how to operate using these tools. That AI gives you actual productivity increases or better decision making, or new found abilities it's irrelevant, only that it appears to be is what matters. You cannot risk that it does for real and you didn't benefit, so if it appears to be, you'll jump in to not miss the boat. Once you've jumped in, your entire company processes, your way of working, all will adapt to working with AI, there will not be a retrospective to compare if the path without you adopting AI would have been better, it just becomes the new way to operate. New grads learn this new way, and everyone is hooked on that it allows you to be lazier at work, and it praises you constantly. There's so many things that follow this cycle I feel. Do you really need Slack ? Notion ? Did they materially change anything about how efficient/effective your company operates? Not really, they just become the new way of doing things 🤣

And don't forget, the use cases go beyond text generation. The models are multi-modal now. You have to factor in use for generated art, marketing materials, assets and so on. As well as the longer bet on robotics and voice. The idea that vision models will allow more robotics use cases, and also create new voice paradigms for computer interaction, which would enable fully airpod driven or glasses driven interfaces, and so on.

Yes, you are right @didibus, we are mimetic apes. Unfortunately, "they just become the new way of doing things". My main issue is that people talk about how things will be (massive unemployment, huge productivity gains, dark factories,...) very lightly, without really believing it. And because of that stupid hype and fear, many individuals and companies are losing money when the reality is totally different. The current AIs are almost intelligent, almost useful... we are living in an eternal almost. This planet still needs human intelligence. More than ever I would say.

I admit I probably also fall in the worried camp. Not because AI is so intelligent, but because so many jobs don't require much of it. So many jobs are just data entry or bookkeeping, or involve such things. Some dev jobs are just making a UI over a CRUD MySQL DB all running in a single inctance. And so on. Even if you manage to automate 10% of the work of white collars, and 90% remains, that's a huge disruption that causes chaos.

I fall in the "how others automate stuff we can't automate?" camp. I read the news (Mythos lately), I read Linkedin posts (Mercadona Tech developing an internal RAG in a weekend), but I only see AI slop in real life. And no money to be made with this new way of doing things. Meanwhile humans have been devalued. It's funny to me seeing people bragging about doing things with the help of AI.

Yeah, I think if you're looking for full, mature, complete applications, start to finish, fully vibecoded. Those do seem to be mostly bragging about it, but not showing anything working. But if you start digging more into the AI assisted or just use of AI in general, I think there's quite a lot of ways people are leveraging it that aren't immediately as impressive because it's not like they told the AI they wanted this amazing app and it built it from scratch.

There's also been a noticeable increase in slop, and it's just becoming something you have to live with on a daily basis. People send you PRs that are monstrous and I've basically given up on trying to properly review them. You also get documents now where they're not even trying to hide it. All the tells are right there, nothing's been done to make it look like a human wrote it. It's wordy, it's long, and it doesn't distill the key information at all. So I often can't be bothered to read it in depth either. Plus, it always has that strange quality of being mostly correct but not quite, which ends up being more confusing than helpful.

Spot on.