[mmargenot]

At least within tech, there seem to have been explosive changes and development of new products. While many of these fail, things like agents and other approaches for handling foundation models are only expanding in use cases. Agents themselves are hardly a year old as part of common discourse on AI, though technologists have been building POCs for longer. I've been very impressed with the wave of tools along the lines of Claude Code and friends.

Maybe this will end up relegated to a single field, but from where I'm standing (from within ML / AI), the way in which greenfield projects develop now is fundamentally different as a result of these foundation models. Even if development on these models froze today, MLEs would still likely be prompted to start with feeding something to a LLM, just because it's lightning fast to stand up.

[andy99]

Its probably cliche but I think it's both overhyped and under hyped, and for the same reason. They hype comes from "leadership" types that don't understand what LLMs actually do and so imagine all sorts of nonsense (replacing vast swaths of jobs or autonomously writing code) but don't understand how valuable a productivity enhancer and automation tool to can be. Eventually hype and reality will converge, but unlike e.g. blockchain or even some of the less bullshit "big data" and similar trends, there's no doubt that access to an LLM is a clear productivity enhancer for many jobs.

[mallowdram]

AI was a colossal mistake. A lazy primate's total failure of imagination. It conflated the "conduit metaphor paradox" from animal behavior with "the illusion of prediction/error prediction/error minimization" from spatiotemporal dynamical neuroscience with complete ignorance of the "arbitrary/specific" dichotomy in signaling from coordination dynamics. AI is a short cut to nowhere. It's an abrogation of responsibility in progress of signaling that required we evolve our lax signals that instead doubles down on them. CS destroys society as a way of pretend efficiency to extract value from signals. It's deeply inferior thinking.

[joquarky]

Let me use AI to translate this into plain English:

> AI was a huge mistake. It shows a lack of imagination and confuses ideas from different sciences. Instead of helping us improve how we communicate, it reinforces our weakest habits. Computer science pretends to make things more efficient, but really it just extracts value in shallow ways. This is poor, second-rate thinking.

[mallowdram]

It lacks references, it's garbage, advertising, cliff-notes for apes uninterested, devolving, asleep, bored, and needing to be told what to think without knowing why or how. The inertia in CS, and the inertia and entropy CS unleashed on the gen public will take years to cleanse from the system before we get back to imaginative progress and invention.

[kragen]

What new non-AI products do you think wouldn't have existed without current AI? Because I don't see the "explosive changes and development of new products" you'd expect if things like Claude Code were a major advance.

[spicyusername]

At the moment, LLM products are like Microsoft Office, they primarily serve as a tool to help solve other problems more efficiently. They do not themselves solve problems directly.

Nobody would ask, "What new Office-based products have been created lately?", but that doesn't mean that Office products aren't a permanent, and critical, foundation of all white collar work. I suspect it will be the same with LLMs as they mature, they will become tightly integrated into certain categories of work and remain forever.

Whether the current pricing models or stock market valuations will survive the transition to boring technology is another question.

[kragen]

Where are the other problems that are being solved more efficiently? If there's an "explosive change" in that, we should be able to see some shrapnel.

Let's take one component of Microsoft Office. Microsoft Word is seen as a tool for people to write nicely formatted documents, such as books. Reports produced with Microsoft Word are easy to find, and I've even read books written in it. Comparing reports written before the advent of WYSIWYG word processing software like Microsoft Word with reports written afterwards, the difference is easy to see; average typewriter formatting is really abysmal compared to average Microsoft Word formatting, even if the latter doesn't rise to the level of a properly typeset book or LaTeX. It's easy to point at things in our world that wouldn't exist without WYSIWYG word processors, and that's been the case since Bravo.

LLMs are seen as, among other things, a tool for people to write software with.

Where is the software that wouldn't exist without LLMs? If we can't point to it, maybe they don't actually work for that yet. The claim I'm questioning is that, "within tech, there seem to have been explosive changes and development of new products."

What new products?

I do see explosive changes and development of new spam, new YouTube videos, new memes (especially in Italian), but those aren't "within tech" as I understand the term.

[mmargenot]

I do agree that there's a lot of garbage and navel-gazing that is directly downstream from the creation of LLMs. Because it's easier to task and evaluate an LLM [or network of LLMs] with generation of code, most of these products end up directly related to the production of software. The professional production of software has definitely changed, but sticky impact outside of the tech sector is still brewing.

I think there is a lot of potential, outside of the direct generation of software but still maybe software-adjacent, for products that make use of AI agents. It's hard to "generate" real world impact or expertise in an AI system, but if you can encapsulate that into a function that an AI can use, there's a lot of room to run. It's hard to get the feedback loop to verify this and most of these early products will likely die out, but as I mentioned, agents are still new on the timeline.

As an example of something that I mean that is software-adjacent, have a look at Square AI, specifically the "ask anything" parts: https://squareup.com/us/en/ai

I worked on this and I think that it's genuinely a good product. An arbitrary seller on the Square platform _can_ do aggregation, dashboarding, and analytics for their business, but that takes time and energy, and if you're running a business it can be hard to find that time. Putting an agent system in the backend that has access to your data, can aggregate and build modular plotting widgets for you, and can execute whenever you ask it a question is something that objectively saves a seller's time. You could have made such a thing without modern LLMs, but it would be substantially more expensive in terms of engineering research, time, and effort to put together a POC and bring it production, making it a non-starter before [let's say] two years ago.

AI here is fundamental to the product functioning, but the outcome is a human being saving time while making decisions about their business. It is a useful product that uses AI as a means to a productive end, which, to me, should be the goal of such technologies.

[kragen]

Yes, but I'm asking about new non-AI products. I agree that lots of people are integrating AI into products, which makes products that wouldn't have existed otherwise. But if the answer to "where's the explosive changes and development of new products?" is 100% composed of integrating AI into their products, that means current AI isn't actually helping people write software, much. It's just giving them more software to write.

That doesn't entail that current AI is useless! Or even non-revolutionary! But it's a different kind of software development revolution than what I thought you were claiming. You seem to be saying that the relationship of AI to software development is similar to the relationship of the Japanese language, or raytracing, or early microcomputers to software development. And I thought you were saying that the relationship of AI to software development was similar to the relationship of compilers, or open source, or interactive development environments to software development.

It also doesn't entail that six months from now AI will still be only that revolutionary.

[mmargenot]

For better or for worse, AI enables more, faster software development. A lot of that is garbage, but quantity has a quality all its own.

If you look at, e.g. this clearly vibe-coded app about vibe coding [https://www.viberank.app/], ~280 people generated 444.8B tokens within the block of time where people were paying attention to it. If 1000 tokens is 100 lines of code, that's ~444M lines of code that would not exist otherwise. Maybe those lines of code are new products, maybe they're not, maybe those people would have written a bunch of code otherwise, maybe not. I'd call that an explosion either way.

[spicyusername]

    Where is the software that wouldn't exist without LLMs?

Where are the books that wouldn't exist without Microsoft Word?

[kragen]

I've definitely read a lot of books that wouldn't exist without WYSIWYG word processors, although MacWrite would have done just as well. Heck, NaNoWriMo probably wouldn't.

I've been reading Darwen & Date lately, and they seem to have done the typesetting for the whole damn book in Word—which suggests they couldn't get anyone else to do it for them and didn't know how to do a good job of it. But they almost certainly couldn't have gotten a major publisher to publish it as a mimeographed typewriter manuscript.

Your turn.

[spicyusername]

My point is that these are accelerating technologies.

    maybe they don't actually work for that yet.

So you're not going to see code that wouldn't exist without LLMs (or books that wouldn't exist without Word), you're going to see more code (or more books).

There is no direct way to track "written code" or "people who learned more about their hobbies" or "teachers who saved time lesson planning", etc.

[apwell23]

> What new non-AI products do you think wouldn't have existed without current AI?

AI slop is a product

[kragen]

You mean, like, SEO? It's a product in the same sense that perchloroethylene-contaminated groundwater is a product of dry-cleaning plants.

[boringg]

I think the payment model is still not there which is making everything blurry. Until we figure out how much people have to pay to use it and all the services built on its back it will remain challenging to figure out full value prop. That and a lot of company are going to go belly up when they have to start paying the real cost instead of growth acquisition phase.

[jebarker]

I don’t think a payment model can be figured out until the utility of the technology justifies the true cost of training and running the models. As you say, right now it’s all subsidized based on the belief it will become drastically more useful. If that happens I think the payment model becomes simple.

[mmargenot]

There's enough solid FOSS tooling out there between vLLM and Qwen3 Apache 2.0 models that you can get a pretty good assistant system running locally. That's still in the software creation domain rather than worldwide impact, but that's valuable and useful right now.

[mallowdram]

The immaterial units are arbitrary, so 'agents' are themselves arbitrary, ie illusory. They will not arrive except as being wet nursed infinitely. The developers neglected to notice the fatal flaw, there are specific targets but automating the arbitrary never reaches them, never. It's an egregious monumental fly in the ointment.