Fuck AI

There are a few, yeah. Whether it's enough to balance the massive weight of all the externalities, I have no idea. Currently leaning no. Could be wrong about that, who knows.

Basically: we now have the tech to make sense of language and language semantics, and use language as a universal interface. You and I are fine clicking buttons in programs, sure, but you and I are also having this discussion on an obscure federated social media platform the general public has never heard about. Interacting usefully with a computer system through language alone becomes possible in a way that it wasn't before. I'm not quite sure how valuable this is going to be in the long term, but then, I'm also a tech nerd who is used to clicking buttons and writing command lines.

We can now process large amounts of text fast for data extraction, which is a deceptively hard problem. You can do things like importing itemized PDF bills into an accounting database with no prior knowledge of how those bills are formatted. This extends beyond text. We can now generate a textual description of arbitrary images and videos. That too is a very hard problem. It can now be done on a regular desktop computer using a small local LLM.

It's an even harder problem when the text is computer code and the data being looked for is the cause for a specific behavior. The process of debugging an obscure issue can now be massively accelerated.

Given a reliable corpus of knowledge, that corpus can be queried more or less instantly using natural language. That's also something we could not do before.

LLMs suck at designing software but can produce code to spec faster than a human, which means they can be used to increase throughput where a skilled human does the design and is limited only by the speed of implementing it. Given the prevalence of software in the economy, the impact of that alone will be significant.

All of these come with major drawbacks and sometimes intractable problems. Language is squishy and ambiguous. LLMs don't THINK, they extrude statistically probable continuation tokens. AI content sucks, be it writing, images, videos, because the probable tokens there are the median of the training corpus, and median is a cognate of mediocre for a reason. I hope AI slop goes away. But I don't think it will. The ability to generate custom porn on demand alone will likely sustain a market.

And I didn't think we can go back to the world of before. But personally, I wish we could. Because the externalities here are, and remain, enormous.