This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
I dunno, I've read the case for hitting AGI on a short timeline just based on foreseeable advances and I find it... credible.
And If we go back 10 years ago, most people would NOT have expected Machine Learning to have made as many swift jumps as it has. Hard to overstate how 'surprising' it was that we got LLMs that work as well as they do.
And so I'm not ruling out future 'surprises.'
That said, Sam Altman would be one of the people most in the know, and if he himself isn't acting like we're about to hit the singularity well, I notice I am confused.
Aschenbrenner is a smart charlatan, he's probably going to do very well in the politics of AI.
My opinion is that the way he has everyone fooled and the way he has zeroed in on the superpower competition aspect makes it clear what he is after. Power. Has he gotten as US citizenship yet? He'll need that.
There's going to be an enormous growth in computing power, possible hardware improvements (e.g. the Beff Jezos guy has some miniaturised parallel analog computer that's supposedly going to be great for AI stuff.. ). But iirc, the models can't really improve easily because there's not the best data to pretrain them, so now everyone is trying to figure out how to automatically generate good synthetic data and use that to train better models, combine different modalities (text/ images etc). All stuff that's hardly comprehensible to outsiders, so people like Leopold can go around and say stuff with confidence.
Likely, yes, but how computationally and energy expensive it's going to be matters a whole lot. Like e.g. aren't they basically near hitting physical limits pretty soon? That'd cap lowering power costs, right?
And scaling up chip production to 1000x isn't as easy as it sounds either. Especially if Chinese get scared and start engaging in sabotage.
It'd make me feel better if someone could muster a rebuttal that explained with specificity why further improvements aren't going to be sufficient to breach the "smarter than human" barrier.
There's an existence proof in the sense that human intelligence exists and if they can figure out how to combine hardware improvements, algorithm improvements, and possibly better data to get to human level, even if the power demands are absurd, that's a real turning point.
A lot of smart people and smart orgs are throwing mountains of money at the tech. In what ways are they wrong?
To sum it up, to train superhuman performance you need superhumanly good data. Now, I'd be all okay for the patient, obvious approach there - eugenics, creating better future generations.
I'll quote twitter again
I'd not say they're wrong. Even present day polished applications with a lot of new compute could do a lot of stuff. They're betting they'll be able to make use of that compute even if AGI is late.
And remember, the money is essentially free for them. Those power stations will be profitable even if datacenters aren't, the datacenters will generate money even if taking over the world isn't a ready option. & There's no punishing interest rates for the big boys. That's for chumps with credit cards.
It isn't clear we need superhumanly good data. Humans can make novel discoveries if they have a sufficiently good understanding of existing data and sufficiently good mental horsepower to use that data, i.e. extrapolate from their set of 'training data' and accurately test those extrapolations to discover new, useful data.
It seems like we just need to get an AI to approximately Von Neumann level and if it starts making good contributions to various fields at that point we can have it solve problems that hold up AI development. We're seeing hints of this now with Alphafold 3 and AlphaProteo.
Right now, the one thing that appears to be a hard hurdle for AIs are navigating real world environments, where there is far more chaos and variables that don't interact with each other linearly.
It can be difficult to see a new true innovation coming when every single company starts slapping "AI Powered!" as a feature on their products, but I think the case that AI will make surprising leaps in the next few years is stronger than it will inexplicably stagnate.
It is.
LLMs and similar systems aren't human, not in the slightest.
They're nowhere near that. People are happy they can count letters correctly.
As stated, be really nice if there was a sound case for why this won't change in the near future.
The jump to where we are was sudden and surprising, the next one could be as well.
If a cat surprises you with a sudden jump to a wardrobe top just under a ceiling, how likely is it to then surprise you with jumping through the ceiling ?
There's no superhuman data, this isn't as easy. What happened until know was just adding more scale and it led to increasing returns until it got to the point there was no new data to add. And this is still a system that's basically a glorified retrieval tool because it has trouble with basic logic unless it uses 'chain of thought' which is pretty expensive and not much better.
So do I, but they gave me the PhD anyway. Applied math, so there were a lot of computers rather than just the traditional coffee maker and a chalkboard, but you'd be surprised at how heavily even the more-respectable pure mathematicians rely on that chalkboard (or whiteboard, these days) to keep track of a chain of thought.
And on that subject, math is one field where you can add new perfect-quality synthetic data with no obvious bound. Generating a proof may be NP-complete, but verifying a rigorous proof can be done by non-AI computer programs. Both OpenAI and Meta have started using Lean to train models to generate new proofs. It's not quite as good a target as Go self-play, since "play against yourself" is a naturally good difficulty-level for self-improvement, whereas some theorems are deceptively complex to state and easy to prove or vice-versa, but it's not a slow-growing data set in the same way as "mimic our recorded human language" is, and yet it's natural to translate back and forth between the formal proof language and the basic-logic-plus-much-more subset of human languages.
I'd still say it's possible that there's some other "spark" of out-of-sample creativity that humans have and models trained under current techniques just can't acquire, but if that's what you're shooting for, at least lead your target. Even if current progress does stall out, we may end up going from "has trouble with basic logic" to "(dis?)proved the Riemann hypothesis" before things plateau again.
More options
Context Copy link
The entire course of technological progress was us surprising ourselves by jumping through what we thought for millenia was the ceiling. Either we've been achieving superhuman data every time, or there is much more human data than we think.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link