Comment by ssivark
5 hours ago
> 4 years ago, a program that could [...] we would have called it AGI
If you had told someone in the 1800s that a machine could instantly multiply 100 digit numbers, that would have been considered dazzlingly intelligent. And yet we are not that dazzled by our calculators today (despite how useful they might be!).
Are you trying to explain how things once considered dazzling get normalized over time? Because otherwise this is a non-sequitur and has no bearing on the trivially verifiable, exponential explosion of capabilities we have seen in the last 4 years.
I keep saying this, until ChatGPT came out 4 years ago it was basically unimaginable that a single model could do any of, let alone all, the things they are doing today. Like, seriously, go take a look at the state of the art in NLP and NLU, the very first challenge in getting computers to even “understand” natural language, let alone other things like reasoning. Everything it does automatically was once a heavily experimental deep research field with long glorious careers for the researchers.
And now it’s all gone because the Bitter Lesson won again. If that’s not general enough to qualify for the G in AGI I don’t know what it is. And we’re sitting here going, “But it sometimes writes bad code though.”
Speak for yourself, I am dazzled by calculators!
In any case, I think this misses OP's point that LLM capabilities have rapidly made progress towards being more generally intelligent and capable, which is not true of most tech advances.
This is a motte & bailey moment. Parent comment stated something much sharper, that I responded to:
> I do not understand how you can see what has happened in the last 3 years and say "it will not get us there" no matter what "there" is.
--
Your statement is something much weaker, and I would still question what exactly "general" means when AI capabilities are commonly accepted to be so "jagged".
Maybe my phrasing is too weak but if the parent comment is the 'bailey', I fully agree with it.
The last 3 years of progress have been so explosive and, yes, general that it seems crazy to fully rule out dramatic future progress.
When people have a very narrow 'confidence interval' about their AI predictions, in either direction, it's difficult to trust them.
>I think this misses OP's point that LLM capabilities have rapidly made progress towards being more generally intelligent and capable, which is not true of most tech advances.
A PC of today can accomplish many more "general" tasks than one of 40 years ago. Much of the "why" is because of the huge infrastructure built up around them in the meantime. The abilities of LLMs to accomplish those same tasks through the PC is heavily piggybacking on that (both in the specific, with the existence of all the APIs and tools; and in the generic, using search engines to find specific sources and using that for instruction or troubleshooting).
In the world of "agents" much of the improvement appears to have been on a specific set of skills: impersonation of an 'I' that wants to accomplish a goal, and synthesizing existing information from documents with trial-and-error execution loops to move rapidly toward a solution much faster and with less boredom than a human would. The quality of the output when there is not a rapid-evaluation-and-validation harness lags considerably.
It's incredibly powerful automation but doesn't appear to be trending towards Matrix-style conscious AIs. The quality of an individual method written by the agent also is not particularly advanced compared to GPT-4 in early 2023, as far as I can tell—I was dabbling with trying to make such harnesses back then, where a major challenge was that the model itself was bad at staying on-track in a conversation, so instead much of that logic was moved to deterministic code, which was much more limited as it was super-tedious to enumerate all the necessary tool calls/etc to find its way out of corners. Staying on task is much better now, as is "read compiler error, fix try next thing" harness loop-handling. But the output remains—across Fable, Astra, whatever else I've tried—"iffy" in terms of the actual code structure on the first pass output. You can set it then on a different task to review and clean up the code, and it can do that well too, but it is a curious gap of generality where the "create" focus is much more limited than the "review" one (and conversely the "review" focus can make suggestions, but if it goes deep down the well of implementing them, loses that big-picture again).
If it kills us all, it will because someone decided to give the trial-and-error-loop-machine access to nukes or similar. The blame for that is on the "someone" not on some sort of "rogue" AI.
(I wonder if re-watching Terminator/Terminator 2 would support this sort of interpretation of it. Unlike in the Matrix, I don't think we get much sentient-AI POV/infodumping. Is it a plausible universe for "someone made ChatGPT control a fleet of soldier robots and gave it a bad harness with an insufficient sandbox"?)
Actually, check out the story of Horizon Zero Dawn.