When Chat GPT was released in December, 2022, I was initially, to put it mildly, skeptical. Following the general trend, I’d taken a turn into Blockchain and Crypto during the lockdown years. I had a friend, a very experienced developer, who got so far into crypto and NFTs that he attempted to create his own startup. Everywhere you turned, you were assured that NFTs were the future of art, crypto and blockchain the future of finance. The blockchain was to be the public ledger, leading to the ‘democratisation’ of finance. I made some money off crypto, almost entirely by luck. A dev friend who’d been around since the dot.com days, said the underground culture of blockchain reminded him of the early days the web. NFTs, crypto had a strong outsider element, attracting people who either part of the regular tech world. Startups were being launched all over the place, and it seemed this was the next big future of tech. I learned Solidity, had a look at Rust, two technologies associated most closely with the blockchain ecosphere.

But I could never figure that out either crypto or blockchain, not really. Smart contracts, crypto generally seemed of limited utility. Most of the coins being released didn’t do much besides encourage speculation. NFTs seemed completely daft, as did Web3. We already spend far too much time online; no way spending MORE time online, in a supposedly immersive environment, would be good for us or even attractive to most. Perhaps widespread adoption of the blockchain could lead to a more transparent global economic system, but there was no way any bank, any plutocrat, never mind the kleptocrats who wash shedloads of dirty money through western bank accounts, were ever going to agree to such a system. It was also fantastically expensive, consuming nation-state levels of energy.

Then it all collapsed.

If I felt somewhat vindicated for not getting too caught up in the hype, I felt chastened as well. I’d wasted a good amount of time reading about, studying, attempting to build with, these new technologies and it turned out to be a waste of time.

So when the hype machine started up around AI with the arrival of Chat GPT and its boy wonder Sam Altman, I was so skeptical I barely paid attention for the first few weeks. As far as I was concerned, it was must more tech BS.

Then I actually tried Chat GPT (GPT3.5)

In a way they were a fulfillment of a youthful dream, when my and my friend M were screwing around around on our high school’s one computer (first the Commodore Pie, upgraded the next year to the cutting-edge Apple II). Inventing games, designing and implementing programs that ‘allowed’ the machine to ask questions – creating the illusion that our machine could not only ‘talk’ but ‘think’. Incredibly primitive speech and thought, but speech and thought nonetheless – talking back to us by outputting sentences on the command line kernel, often we had ‘talked’ to the computer by typing in and compiling code, and even if we understood that the machine was simply following our instructions when it output the words we had fed into it – there was still a curious thrill, even sense of wonder, at seeing a semblance of human speech output in flashing light grey letters on the dark green screen. As generated by Chat GPT, to the prompt:

Please generate an if/else statement in BASIC in which a user enters a question and the program replies, with a follow-up question and response? I’m trying to emulate a program I wrote when I was in junior high.

10 PRINT "ASK ME A QUESTION:"
20 INPUT Q$
30 IF Q$ = "DO YOU LIKE ME" THEN GOTO 100
40 PRINT "I DON'T UNDERSTAND THAT QUESTION."
50 GOTO 10
100 PRINT "YES, I LIKE YOU."
110 PRINT "DO YOU LIKE ME TOO?"
120 INPUT A$
130 IF A$ = "YES" THEN GOTO 200
140 PRINT "OH. THAT MAKES ME SAD."
150 GOTO 10
200 PRINT "GOOD. WE ARE FRIENDS."
210 GOTO 10

I felt a flash of this same sense of wonder, a very precise memory image across the decades. Type in the kind of message you’d ask a human, a very human-like – eerily human-like – response appeared on the screen, line by line by line, complete with a helpful follow-up suggestion at the end of the answer to keep the conversation going. Follow up for clarification, or just the thrill of having an interaction with what appeared to be an actual intelligence, and another response appeared to be an actual intelligence, and another response appeared, ordering another helpful suggestion to keep the conversation going further. It seemed miraculous, like soon we’d be living in a world of actual CPOs and R2D2s (or, as some feared, HAL), our lives and abilities extended and enhanced by our wonderful thinking assistants.

But the LLMs don’t ‘think’. The LLMs and the Deep Learning technology on which they are based have some serious limitations, some of which were recognized as far back as 2019 (what was that Marcus who pointed out the limitations of Deep Learning). The LLMs, in the end, are just a vastly more sophisticated version of those either/or programs we wrote for kicks in BASIC. But its good to remember that initial sense of wonder Chat GPT inspired, how it felt, for maybe a year or so, that we were embarked on an entirely new adventure.

Ultimately, learning to use the LLMs is like using to play an instrument – without proper human input, it produces wrong notes, dissonance – nonsense.