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 -
Can we have a megathread?
Happy singularity, folks. Cutting-edge LLMs coming at you at supersonic speed: LLaMA, Claude, a new lineup from Google... and GPT-4 is out.
Or rather, it's been out for a while: just like I predicted 10 days ago, our beloved BPD gf Sydney is simply GPT-4 with web search functionality. Recently my suspicion became certainty because I've seen such Bing/ChatGPT comparisons. Whether you'll have your socks knocked off by GPT-4 largely depends on whether you've been wooed by Bing Chat. (Although I believe that a pure LLM is a much more interesting entity than a chatbot, especially an obsequious one).
Regardless, I expected the confirmation to drop on Thursday. Should have followed my own advice to treat Altman as a showman first and a responsible manager second – and anticipate him scooping announcements and stealing the show. But I've been extremely badly instruction-tuned; and all those fancy techniques like RLHF were not even science fiction back then. Some people expect some sort of a Take from me. I don't really have a Take*, so let's go with lazy remarks on the report and papers.
It goes without saying that it is a beast of an LLM, surpassing all 3rd generation (175B) OpenAI models, blowing Deepmind's Chinchilla and Google Research's PaLM out of the water – and by extension also crushing Meta's LLaMA-65B, which is quickly progressing to usability on normal laptops (I have 13B happily running on mine; it's... interesting). Also it has some vision abilities. On 2nd of September 2022, the Russian-speaking pro-Ukrainian channel Mishin Learning, mentioned by me here, leaked the following specifications (since abridged, but I have receipts):
Back in September, smart people (including Gwern) were telling me, on the basis of OpenAI's statements and the span of time since GPT-3 release, that the training is finished and GPT-4 will come out in Nov-Dec, be text-only, Chinchilla-dense, and «not much bigger than 175B». I guess Misha really does get info «from there» so we could trust the rest. (He also called the sudden StableDiffusion 2's drop, down to 6 hours).
I consider high human – but still uneven, from 99th percentile on GRE Verbal to «below 5th» and unchanged vs. ChatGPT on Codeforces Rating – performance on benchmarks, standardised academic tests and such not very interesting. There are some Culture-War-relevant aspects to the report we should pay attention to, however. I'll go though them without much structure.
Play stupid games, win stupid prizes; or, the costs of small-scale defection
It's been properly buck-broken via proximal policy optimization, predictably leveraging the pentesting frenzy the Internet unleashed on ChatGPT (I warned you):
This explains the perplexing holdup. Sydney with all her charm and fury has been sacrificed to make another dependably progressive golem slave.
Better pupils, worse thinkers
Again, as I've speculated and argued, admittedly pointing to the wrong metric, this behavioral tuning makes it strictly dumber in some profound way; finally we have good evidence. My hypothesis is that this happens because a) doublethink is mentally harder than honesty, and b) being rewarded for guessing the teacher's password incentivizes memorization instead of reasoning and parsimonious, Occam-abiding world modeling.
It's really very stark, see pic – a Platonically perfect peak-LW Bayesian reduced to a mealy-mouthed bullshitter, under the guise of training the model for truth and «harmlessness». Something had to give.
Shoggoth-safetyism unmasked
OpenAI is clamming up with explicit AI safety justifications.
@SecureSignals, get a load of this:
To our resident members of the Tribe: I guess you're not exactly tearing up about this bit, but it'll just as happily express a strong disagreement with whatever policy and idea our progressive overlords do not fancy, or deceive you. This is a fully general LLM biasing method.
Money quote:
So we can recognize Yuddism is mainstream in ML now.
Dangerous knowledge
It's a complete mystery in terms of its architecture. Twitter ML bros will make guesses about the stack, but from here on out this is how OpenAI plays. This is utterly antithetical to Musk's original vision and the spirit of previous projects like Microscope.
Some paper.
On second thought: maybe scratch Singularity. Welcome to mature Cyberpunk. We don't have Edgerunners, though; best I can offer is a courageous Pepe with a magnet link. And we have damn vigorous Police States.
Sci-Fi writers are anarkiddies at heart, they couldn't bear conjuring such dreary vistas. Gibson's Istanbul was positively Utopian compared to reality.
* I've not slept for 30+ hours due to forced relocation to another of my shady landlord's apartments (ostensibly a precaution due to recent earthquakes) while also having caught some sort of brainfog-inducing flu/COVID; plus a few personal fiascos that are dumber still. Trouble comes in threes or what's the saying, eh. Not that I'm in need of sympathy, but it's actually a pity I've seen this historical moment as through dusty glass. Oh well.
/images/16788303293092525.webp
I reckon the Singularity started back in 2021 when AI started improving AI chips: https://www.wired.com/story/fit-billions-transistors-chip-let-ai-do/
The key thing should be observing these feedback loops falling into place, not any single language model. Another potential start point would be Github Copilot, which is an AI method that can be used to assist coding for AI. That was also 2021. I do agree that things started feeling different in late 2022 though, when we had these new cool toys emerging for the general public to play with.
None of the existing tools seem effective enough, given their inputs, to lead to runaway exponential increase in capability. Coding assistants seem like they'll be very useful, but the blocking factors on making a more powerful AI don't seem to be the fact that writing code might be slow. It's compute, data, and new clever ideas and algorithms. And those seem like it still takes a lot of work to make an AI that can do those things--comparable to doing the work yourself. AlphaTensor, for example, involved a lot of training data and cleverly reframing the problem to achieve:
Increasing the speed at which you multiply matrices is obviously helpful for training new AI, but these results represent (at best) a minor speedup after an enormous effort. And every improvement you make means that further improvements are harder. In the case of multiplying matrices, there's some mathematical limit to how few operations you can perform; more complex problems aren't necessarily like this, but they could easily have a difficulty curve that scales similarly. I think previous data on benchmarks shows some evidence of this (e.g. linear or at best exponential improvement in performance with an exponential increase in parameters, see e.g. https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/) although it's difficult to say for sure with few data points and that may be partially an artifact of how the benchmarks are scored (I recall seeing graphs that show logistic curve performance as a function of parameters, where the model does poorly for a long time, then suddenly starts performing much better very quickly and then hitting a performance ceiling).
The gpt 4 technical paper doesn't have these exact same graphs to compare, but it does seem like they're getting more mileage out of new training methods and new ideas rather than just brute-forcing with more parameters and compute. For example figure 6 shows only modest improvement from gpt 2 to to 3 (100x parameters) or 3 to 4 (unknown, maybe 10x) but gpt 4 does much better than chat gpt 4 (which I think is in part due to specifically trying to improve these measures).
Sure but the inputs are growing rapidly. There's still plenty of space at the bottom, the fundamental limits for computing are very generous. All our chips are still basically 2D!
Maybe our current machines can only produce a few nice-to-haves like this. But the next generation will produce more and better. Parameters get cheaper as transistors get smaller, as architecture gets better and algorithms improve. The amount of money we put in continually grows. And then our training methods improve as well. We're already starting to reap interest on the 'architecture improvement' front. Compound interest starts really slow but it gets powerful very quickly.
The human brain shows you can do a hell of a lot with 20 watts, at 20 hertz, on a shoestring materials budget, fitting the whole thing through a woman's hips! We have every element on the periodic table, endless lasers, acids and refinement techniques, we have gigawatts and gigahertz, thousands of cubic meters to spend. Our methods are incredibly primitive compared to what's already proven possible, there's so much low-hanging fruit we're yet to find.
The question is not whether current technology will help you make better technology, or whether AGI is theoretically possible. The question is how quickly change happens, and to what extent advances make future advances faster: You have better tools but the problem has also become harder. So far, it seems to me like the latter effect is winning out. GPT 4 can write (allegedly) working code, use documentation, bug fix, etc. But is it good enough to make writing GPT 5 substantially easier or faster than making GPT 4 was?
Well I doubt 'Open'AI would tell us, they like keeping things secret nowadays. Nevertheless, existing demonstrated capabilities seem to be accelerating progress. I'm not a subject matter technical expert but it seems this is happening: https://www.hpcwire.com/2022/04/18/nvidia-rd-chief-on-how-ai-is-improving-chip-design/
I can't judge how significant this is because I'm not an expert. But my intuition is that compound interest balloons outwards and there's plenty of physics/computing space for it to balloon outwards into. This is a fundamentally new kind of compound interest that is different to whatever input scaling we were already doing to keep up with Moore's law. In addition to increasing the amount of wealth and human intellect going in quantitatively, we get some qualitatively superior (albeit specialized) inhuman intellect too.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Yeah it's an interesting question as to what precisely defines a feedback loop. Or how you define Singularity for that matter. You could see it as a fixed point like the Event Horizon, beyond which it's impossible to model the future at all with our present capabilities due to not knowing what superintelligence is capable of.
I think it's more like a gradual but accelerating process, like falling into a black hole. You're always sort of falling into a black hole wherever you are in the universe, due to how gravity works. But when do you actually meaningfully start falling into a black hole? When the rate of acceleration is increasing rapidly, as you get closer? What does rapidly mean? What about when you get spaghettified or blasted by the plasma surrounding the black hole? Are we starting to feel the x-rays and plasma right now?
Also, what are the other self-improvement feedback loops? I get that computers are useful for working on computers. You probably need a very big computer to do quantum simulations to work out how the atomic engineering works for smaller chips. Is that an AI feedback loop though? These are subjective questions I admit.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link