LLM Inevitabilism

The author seems to imply that the "framing" of an argument is done so in bad faith in order to win an argument but only provides one-line quotes where there is no contextual argument.

This tactic by the author is a straw-man argument - he's framing the position of tech leaders and our acceptance of it as the reason AI exists, instead of being honest, which is that they were simply right in their predictions: AI was inevitable.

The IT industry is full of pride and arrogance. We deny the power of AI and LLMs. I think that's fair, I welcome the pushback. But the real word the IT crowd needs to learn is "denialism" - if you still don't see how LLMs is changing our entire industry, you haven't been paying attention.

Edit: Lots of denialists using false dichotomy arguments that my opinion is invalid because I'm not producing examples and proof. I guess I'll just leave this: https://tools.simonwillison.net/

The IT industry is also full of salesmen and con men, both enjoy unrealistic exaggeration. Your statements would not be out of place 20 years ago when the iPhone dropped. Your statements would not be out of place 3 years ago before every NFT went to 0. LLMs could hit an unsolvably hard wall next year and settle into a niche of utility. AI could solve a lengthy list of outstanding architectural and technical problems and go full AGI next year.

If we're talking about changing the industry, we should see some clear evidence of that. But despite extensive searching myself and after asking many proponents (feel free to jump in here), I can't find a single open source codebase, actively used in production, and primarily maintained and developed with AI. If this is so foundationally groundbreaking, that should be a clear signal. Personally, I would expect to see an explosion of this even if the hype is taken extremely conservatively. But I can't even track down a few solid examples. So far my searching only reveals one-off pull requests that had to be laboriously massaged into acceptability.

> LLMs could hit an unsolvably hard wall next year and settle into a niche of utility

LLMs in their current state have integrated into the workflows for many, many IT roles. They'll never be niche, unless governing bodies come together to kill them.

> I can't find a single open source codebase, actively used in production, and primarily maintained and developed with AI

Straw man argument - this is in no way a metric for validating the power of LLMs as a tool for IT roles. Can you not find open source code bases that leverage LLMS because you haven't looked, or because you can't tell the difference between human and LLM code?

> If this is so foundationally groundbreaking, that should be a clear signal.

As I said, you haven't been paying attention.

Denialism - the practice of denying the existence, truth, or validity of something despite proof or strong evidence that it is real, true, or valid

> Can you not find open source code bases that leverage LLMS because you haven't looked, or because you can't tell the difference between human and LLM code?

The money and the burden of proof are on the side of the pushers. If LLM code is as good as you say it is, we won't be able to tell that it's merged. So, you need to show us lots of examples of real world LLM code that we know is generated, a priori, to compare

So far most of us have seen ONE example, and it was that OAuth experiment from Cloudflare. Do you have more examples? Who pays your bills?

What are you talking about? I have multiple open-source projects where I've generated multiple PRs with 90+% AI tools. I don't care that the code isn't as good, because I have people using these features and the features work.

1) https://github.com/domino14/Webolith/pull/523/files (Yes, the CSS file sucks. I tried multiple times to add dark mode to this legacy app and I wasn't able to. This works, and is fine, and people are using it, and I'm not going to touch it again for a while)

2) https://github.com/domino14/macondo/pull/399 - A neural net for playing Scrabble. Has not been done before, in at least an open-source way, and this is a full-fledged CNN built using techniques from Alpha Zero, and almost entirely generated by ChatGPT o3. I have no idea how to do it myself. I've gotten the net to win 52.6% of its games against a purely static bot, which is a big edge (trust me) and it will continue to increase as I train it on better data. And that is before I use it as an actual evaluator for a Monte Carlo bot.

I would _never_ have been able to put this together in 1-2 weeks when I am still working during the day. I would have had to take NN classes / read books / try many different network topologies and probably fail and give up. Would have taken months of full-time work.

3) https://github.com/woogles-io/liwords/pull/1498/files - simple, but one of many bug fixes that was diagnosed and fixed largely by an AI model.