LLMs Compared to the Internet
The more I think about LLMs and understand their limitations, the more I wonder what sort of lasting effect it will really have. It’s clear that this will be transformational, probably as significant to the computing industry as the internet. But maybe comparing it to previous major impacts is obfuscating the case.
Compared to the internet
The internet was obviously transformational…but was it the internet per se or was it the inevitable connecting of zillions of computers. There’s no particular reason why the Internet Protocol had to win. Many other alternative networking schemes were in wide use at the time, but TCP/IP had several features that made it attractive to everyone: it had grown out of universities and was therefore well-understood by many potential employees. Its open protocol, funded by the government, meant no pesky license or copyright issues. I’m sure there were plenty of good technical reasons people might have resisted it, but those were all overcome by the advantages. In other words, it wasn’t the internet that was transformational – it was the idea of networking, especially wide area networking. Once computers started to talk to one another, a whole bunch of useful applications became possible. WWW was the big obvious one, but there could have been any number of alternatives.
LLMs are a class of software solutions driven by the combination of easy large-scale data (text) scraping combined with algorithms and hardware that can process that text into something useful.
Government and regulation is naturally slow-moving, so the new possibilities enabled by networking happened too quickly for regulators to keep up.
But the internet has changed the speed at which regulators can now operate, especially when combined with a powerful incumbent tech industry with the money and intellectual horsepower able to influence the government.
A networked world is full of at least as much danger as anything you can associate with AI: think about malicious hackers and ransomware, illicit contraband traders, or for that matter all the ills people want to blame on social media. If a powerful enough lobby existed in the 1990s when all of the networking technology began to appear, couldn’t they have scared politicians into all sorts of ill-conceived regulations “to protect” us.
AI is no more dangerous than networking. For that matter, it may be less dangerous than Google searches. In fact, if you discount the sci-fi Terminator style scenarios, AI is less dangerous than these.
In fact, thinking more closely, what’s really scary about AI?
It’s a clever compression of all the data humans have compiled. (and incidentally we’re probably closer than we think to exhausting the useful data that can be added to that compression).
Chinese Room, Turing Test: Solved
Now that both of these thought experiments have been realized, what lessons do we have?
The AI doesn’t understand anything.
What do humans do when we understand something?
But Heidegger didn’t emphasize “being and nothingness”. That was Sartre.
I wonder how much the influence went the other way, back to Heidegger.
And were Chinese philosophers and metaphysicians interested as well?