22 comments

[ 2.8 ms ] story [ 66.0 ms ] thread
the sheer volume of content creator speculation based on a "leak" of a single letter of the alphabet from openai connected to "clues" all self referentially feeding themselves into a media frenzy based on absolutely no other confirmatory information from openai... could provide enough training data for GPT5.

as a writer this scares me bc of how un-self-aware every q* anon seems to be right now. how many times have i myself been in this position?

ah, so your comment leads me to conclude that this is part of some hype marketing campaign to get high-level investors to throw money at openAI

...which faintly smells rotten because of them "lack of a moat" arguments

if you genuinely think either of your sentences is an accurate world model, i think you are miscalibrated.
I'm most certainly calibrated differntly than you

what I actually think is way crazier than that. if you asked me (which you didn't) I'd say that I think that behind the scenes what all these llm and ai companies are doing is licensing chinese tech

i regard the timing for the openai fiasco right after the apec meeting in sf as too good for this all to be true

> How to think about the OpenAI Q* rumors

Are rumors worth thinking about?

Rumors are weak evidence. Weak evidence is still evidence.
I hear andreflnr is a giraffe.
You don't actually, is the thing. No humans exist who think that. If they did, that would imply something, wouldn't it, if only that there are some really weird people obsessed with me?
Yes, the person lying to create a rumor doesn't think it's true, but the people who believe the lie do. They treat it as credible, just as you suggested.

Who wouldn't be interested in a giraffe that can type? You deserve your fame.

You're being obtuse. Not all rumors are deliberate lies. Those that are are evidence of something else. Remember that the only question was "are rumors worth thinking about", which includes meta-level concerns.
Not sure how you jumped from "rumor" to "evidence". Prefixing it with "weak" does nothing when you are trying to link "evidence" with "rumor".

A rumor is a rumor. If you want to talk about it, have fun, you might be lucky and it's true, but more often than not when marketing is involved, it's far from being the truth.

The fact that the rumor exists is a different state of the world than if that rumor didn't exist, or a different rumor existed. Sometimes that implies things. If it wasn't made up from whole cloth, someone closer to the truth than you actually thinks that's true. They could still be full of crap, yes, but you don't necessarily want to discount them, either. If it was made from whole cloth, sometimes you can learn things from what you adversary wants you to believe. I don't think this is difficult logic.
> They could still be full of crap, yes, but you don't necessarily want to discount them, either.

I agree. But going from this, to calling a rumor "weak" evidence is a long stretch.

Some people apparently don't understand what evidence is.
I think if you've seen this section of https://youtu.be/zjkBMFhNj_g?si=GPUxt1eHbQJ8HFj-&t=2113 this talk you'll probably generally agree that Q* is something like what's described in this article.
The video hints at why achieving artificial general intelligence (AGI) is so difficult. As an example, it points out that AlphaGo was able to surpass human Go players by playing against itself, which worked because winning Go games provides a clear reward function: winning the game.

Similarly, the rumored Q* breakthrough involves solving simple math problems, another closed domain with a well-defined reward function. This has likely gotten people overly excited about the prospects for rapid advances in domains like open-ended text generation. Defining, much less evaluating, a reward function that maps to the "quality" of generated text in general for use during the inference phase is hard to imagine. So while grafting planning onto a language model works for constrained use cases like math, it is hard to see how it could work for generating high quality, coherent text in general.

Why is this blog called Understanding AI when every single post is an opinion piece about anything but AI?
The density of advances considered "breakthroughs" (whether they are or not) is high right now, and whatever Q may or may not be, it is very likely to be eclipsed shortly, and almost certain to be eclipsed before long, and is otherwise just being used for clickbait, so I find it easy to ignore