AI, LLMs, and What Is Worth My Attention

• Mar 04, 2024
• ai,
• large language models,
• thinking

"True counterculture is always offline." —Daftsocial

Is AI a bubble? It certainly feels like the bubbles of yore, all of which had in common a viable technology that took center stage (e-commerce, peer-to-peer, Web 2.0, social media, crypto), a number of companies trying to conjure a market into existence, and scores of investors seeking to catch the next wave.

Most of these bubbles didn't really pop so much as slowly deflate without causing a industry crash on the scale of the dotcom bust of 2001. Of these, I feel like we collectively dodged an asteroid in crypto, which gave us the courtesy of losing steam before it could be jammed into every facet of daily life. That could have been bad!

Which is exactly what AI feels like now. It's got those vibes.

But I also feel like there's something there there. I'm interested in the technology, even if don't feel the need to delve into gritty details yet^[1]. I tend to "lazy-load my learning" and that's worked for me over a 25-year career^[2] It's always useful to be able to have a conversation about something that's transforming your industry! Or destroying it, maybe!

What do I think, anyway? Probably nothing that others haven't already said, but it's nice to write it down:

• The NLP capabilities are astounding and I agree that no one seems to care very much about this! Ever since I first typed "hello" into my Commodore VIC-20 and got ?SYNTAX ERROR? in return I've been wondering if we'll ever really be able to say "Computer! Analyze atmosphere!" with Picardian authority and be understood.
• Accuracy is going to be key to successful adoption. I think LLMs will get a lot more accurate, but the non-zero chance of hallucination is going to constrain where they can be used safely.
• Which is why I think there will be an interesting market for APIs that shunt requests away to specialized subsystems. Let the LLM do the talking, not make decisions.
• Get ready for a lot more very public humiliations as companies race to prematurely integrate AI into their offerings. I hope someone doesn't die as a result.
• Efforts to integrate tone and facial expression into NLP are probably already underway. There's so much more that goes into human communication.
• Chatbots are boring and hackable.
• Seriously, how can you trust tech that can be subverted by telling it to ignore its programming?
• None of the above concerns will stop the FOMO. Get ready for more and more companies offering up their (possibly your) data to AI in exchange for income to offset falling advertising revenue. They won't be able to resist. Everyone from Wikipedia to Urban Dictionary to Everything² is fighting an internal battle, probably.
• We should prepare for a future court decision declaring AI web scraping legal, due to some concepts of common public good and "to foster creativity" much like Fair Use but maybe without compulsory licensing like we have in the US with music.
• I haven't even touched on image generators! Slurping up original art from blogs and webcomics to feed your AI feels like an even grosser ethical violation than text. That said, if these tools are to exist, they are best deployed by the hands of artists.

Finally, I wonder how these models will stay relevant? So much cool stuff happens completely offline. I wonder if LLMs will always be a few years behind certain trends because they'll have to wait for someone to write about them.

• Previous: Links: Technical Debt