One popular idea is that these systems will asymptotically approximate human intelligence because they're trained on mostly human-written texts. Not only is that untrue, it's also directly contradicted by our experience with previous RL-systems, where they seem to breeze right by human ability without even the slightest hiccup.
Most human systems are much, much, much more complicated than most closed world games (which is where RL approaches have seen massive success, mostly through self-play).
Like LLMs are great, but I honestly can't see us getting actual general intelligence out of them.
Can you elaborate on this? I am clearly not aware of this line of thinking and the related contradiction.
What priors should be updated?
Yet it is generating billions in revenue which Eliza did not.
Perhaps all we need is scale and some refinement techniques to eat a big fraction of the economy.
If unimpressive inputs lead to impressive outputs, that should make you more worried, not less.
It can hold many complex and partially contradictory thoughts in its head at once, in a way that feels significantly superior to Opus (for example). And then can make reasonable syntheses across these.
In a couple rounds of back and forth, with relatively low effort (but strategic) prompting, it produces complex, accurate analyses in 5-10 minutes that would take me multiple hours of hard, very focused work.
I still need to remain tightly in the loop, providing frequent course correction, clarification, high level reframing, nudging, and grounding.
It incorporates my feedback incredibly well.
It’s honestly staggering. Fable has changed my assessment of the current trajectory more than any model since possibly gpt-4. Opus 4.5 of last year might be a close second.
———
My advice for anyone who wants to get more value out of these tools:
When a model does something idiotic, don’t throw your hands up in the air. Be curious. Try to turn it into a puzzle to be solved.
It know it’s hard sometimes, especially if you are drowning in slop from other people… or generated by yourself, heh.
It can be exhausting. I struggle with this also. I have thoughts on how to make it better. We shall see.
As for "updating priors", that goes both ways. There's plenty more reason to think "hey, transformers and RLHF might actually make some killer products" but certainly no reason to think the few people who didn't realise that "GPT3 is too dangerous to release" and "all software engineers will be replaced within 6-12 months" were marketing rather than prophecy have some kind of special insight into how it's all going to pan out. Clock's ticking to the promised 2027 reckoning too...