I hope to keep this fairly short. I have been in a position to work with LLMs recently (as in, attempt to employ them for useful tasks), and I am struck by how little they have actually progressed since I last earnestly played with them (when GPT-3 was released in early 2023). Though my LinkedIn feed has been inundated with a relentless deluge of content suggesting the opposite, I was shocked to learn through repeated underperformance that even the latest and greatest Claude Apotheosis 4.9 million headass models to date still have the exact same problems that GPT-3 did. There's a bit more lipstick on the pig, but purely out respect for engineering as a craft I'd have hoped someone would try something more fundamental to address what are clearly architectural problems. Things like easily falsifiable hallucinations, blatantly absent models of both textual constructs and the physical world, and an abysmal error rate across mathematical tasks, are flaws inherent to the stochastic nature of Large Larguage Models, which are still fundamentally "token predictors". All observable improvements appear to be attempts to simply circumvent these problems with duct tape and scaffolding. So, "the models will get better", as I always hear, seems to be less directionally accurate than "we'll superglue more pig shit and mud on top of them and call it a roof".

But, there's actually a better question to ask than "will it/has it gotten better?"

Why would you want it to?

There's a line from Dune that I like to give to AI psychosis-ites:

Thou shalt not make a machine in the likeness of a human mind.

While this is a fun quip and/or irritating-conversation-ender, the in-universe history behind it is extremely prescient for having been written in the '60s:

Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.

Perhaps this is not some kind of gigabrained end-state to imagine if you understand how the capitalist class operates. Frank Herbert certainly would have had a keen understanding of this, having witnessed a long string of "banana republic" scandals shortly prior to writing Dune. Nevertheless, it's still quite eerie that it perfectly predicts the "best case scenario" that all the leaders and governments of the Western world are rooting for, not to mention everyone on LinkedIn. I cannot overstate how many times I've heard "you have X months to escape the permanent underclass", even from quite intelligent friends of mine. Like yes, "the permanent underclass", if only there were some convenient word in English to describe such a predicament. One starting with an "s" and ending with a "y", perhaps. Too bad there isn't, anyway why is this the outcome all of tech is trying to build toward now?

I do understand that there's some aspect of a prisoner's dilemma here. It's the thought that "someone will be evil enough to build it, so why shouldn't I make a bag doing it?" I'm gonna be so real with you, though, that thinking is completely asinine. If this were something that a trained monkey with a hammer could cobble together then it's basically an inevitability, sure. But, as evidenced by GenAI not getting meaningfully better over the past 3 years or so, this is a genuinely difficult problem that only a handful of people in the world, if any, have the capacity to solve. And they all congregate. Just have the stones to say "In literally all of the fields where this could be transformative, it does equal if not more harm than good. Let's opt out of building it." For instance, my dad enjoys espousing the protein folding angle in particular, so for the sake of being explicit with that counterargument we need to ask ourselves, why do we want John Schizophrenia to be able to shit out a bioweapon with nothing but $20 and a little sneaky wording again?

It's also worth noting that big tech CEOs are huge nerds and likely have that same Dune passage mounted on their wall as some sort of perverse aspirational quote. I would bet a lot of money on Elon, specifically.

And financially, it's all nonsense.

All these companies are running at a colossal deficit. OpenAI had their dirty laundry aired recently, but the bogus accounting isn't limited to their shop. When you bet on a technology that doesn't currently work, doesn't fundamentally change, and your only strategy to improve it is "throw more computers at it lul", pretending there's some kind of a sound investment there is simply fraud. A breakthrough would have to come from the foundational technology, not building more infrastructure with borrowed funds, and not splattering more layers of error correcting "if" statements on top of a model that possesses no actual knowledge.

But, even with that aside, AI companies are betting that this technology will eventually materialize, you the employee are betting that it won't (because you'd like to enjoy the money you made in a world that isn't ruled by Mark Zuckerberg, Sam Altman, Dario Amodei, Sundar Pichai, or Satya Nadella), and you are in a position to influence the outcome. Why not cheat a little?

One last thing.

If anyone ever tells you "let's not boil the ocean here" in a meeting, tell them "so we're turning off the AI summary?" Send me a note if you successfully execute this one off at work, I want to hear about it.