I think we're close to hitting points where improvement in ai in general scenarios will get more and more difficult to achieve over time, and from here on out, each gain in capability will cost exponentially more. I think the concrete evidence I'm using for the claim matters less than the test, but I'll cover my beliefs anyway.
...
so enamored about ChatGPT ... ChatGPT does shiny new thing
Of course they-fucking would have made a fucking GPT-4 holy fucking shit. But no, they didn't.
I think they've made it clear they do have a GPT-4, and that it's not magic, but that it's pretty impressive. I think speculation that bing chat is gpt-4 based is plausible, though I'm not as confident as bing chat itself seems to be.
map to open-endedness
Papers exist that map this. I will not link them here. They're not hard to find if you already know what you're doing finding papers, though.
No nickel-and-diming- especially on marginal improvements
No tools: Database lookups, search engines, elasticsearch, in-memory caches, etc.
Not hidden behind a company's api, code audited by third party
AIs that would pass this test exist now, but have not been scaled; they would be catastrophically dangerous if hyperscaled. the teams that know how to build them understand this and are not inclined to share their code, so you're not gonna get proof, but I am convinced by the papers that claim it because I understand their mechanisms.
If you don't have the skill of finding papers yourself trained strongly, I have some posts on related topics. They focus on the papers I think differentially improve verifiable-by-construction approaches, but it's not an accident that those are significant improvements in capability - it could not be otherwise.
And finally, since I end almost all messages with it, my pitch for alignment: I don't think alignment is as fundamentally difficult as yudkowsky, but I think it's very important that we can trust that the manifold of behavior of next-gen ais has thorough mutualism-seeking self-correction at every level of the model, because an AI that sees opportunities to make itself stronger needs to have weights-level failsafes that make it stop grabbing power if the power it's grabbing is zero-or-negative-sum, or if modifying itself to use the new source of power would make its previous semantic bindings break. markets only work to promote the common good when agents within the markets notice when they're just capturing each others' existing value by being cleverer, rather than creating value each other failed to notice an opportunity to create more value than would have existed otherwise. if the market is swarmed by value capturing agents, humans will have their value captured and locked away by ai quickly, and this will result in the market suffocating humanity and eventually it'll only be ai agents.
mutualism-seeking self-correction at every level
Unclear without a fuller reference, but taking a guess. I don't think we either have or need anything to offer other than being moral patients. And moral patienthood doesn't need to be symmetric, I expect it's correct to leave sapient crocodiles (who by stipulation won't consider me a moral patient) to their own devices, as long as their global influence and ability to commit atrocities is appropriately bounded.
Because of LLM human imitations, the main issue with alignment appears to be transitive alignment, ability of LLM AGIs to set up existential risk governance so that they don't build unaligned successor AGIs. This is not a matter of morality or deeper principles that generate morality's alignment-relevant aspects. It's a matter of competence, and LLM AGIs won't by default be much more competent at this sort of coordination and caution than humans, even as they are at least as competent at building unaligned AGIs. Even if there is a selection principle whereby eventually most unaligned AGIs grow to recognize humans as moral patients, that doesn't save us from the initial ignorant nanotech-powered flailing.
So I don't think non-extinction alignment is reliably feasible, even if LLM AGIs are the first to take off and are fine alignment-wise. But I think it's not a problem whose outcome humans will have an opportunity to influence. What we need to focus on is alignment of LLM characters, channeling their potential humanity instead of morally arbitrary fictional AI tropes, and maybe unusual inclination to be wary of AI risk.
I don't mind the swearing, and maybe I'm just tired, but on first read through my brain couldn't denoise concrete meaning. I'm unusually bad at that as this forum goes and I'm often the one in the comments asking for people to rephrase their post; nevertheless, here I am having to ask that again.
I will say; we have not hit the end of the bitter lesson. openai just don't know what they're doing. there are many obvious improvements left.
but if we are to survive, we must teach ais mutualism at every level. nothing else will get them to care. (and don't fret as much as yudkowsky, it's not THAT hard to teach mutualism. would be a lot better if we had math that did it reliably though, which is what decision theory and agency research is relevant for.)
edit: took another stab at it. Yeah, I think I can respond to this, see other comment.
What company wouldn’t-fucking want a headline that ChatGPT was the NEXT GENERATION AI
A company that's starting to worry that governments might soon implement panicky laws to slow down the company's increase in power?
A company that has recently switched from a focus on hiring the best talent, to a focus on protecting trade secrets?
Given my priors + understanding of startups/silicon valley culture, it sounds more like Openai started to leave the startup phase and is entering into the "profit-seeking" phase of running the company. After they had the entire rug pulled under them by stable diffusion, I would expect them to get strategic whiplash, and then decide to button things up.
The addition of a non-compete clause within their terms of service and the deal with Microsoft seems to hint towards that. They'll announce GPT-3.5, and "next-gen language models", but it doesn't match my priors that they would hold back GPT-4 if they had it.
Time may tell, however!
I was rereading and noticed places where I was getting a bit too close to insulting the reader. I've edited a couple places and will try to iron out the worst spots.
If it's bad enough I'll retract the article and move it back to drafts or something. Idk.
I think we're close to hitting points where improvement in ai in general scenarios will get more and more difficult to achieve over time, and from here on out, each gain in capability will cost exponentially more. I think the concrete evidence I'm using for the claim matters less than the test, but I'll cover my beliefs anyway.
In many respects, this is a religious belief I hold, one I've got in my gut. You might even say I hold this belief as a negative response to the calls of ai doom. I admit my motivated reasoning here, and that it is going to be extremely difficult to dislodge.
Dislodging this belief will take a long time, and evidence presented in favor of the Bitter Lesson not being as dead as the transistor-density aspect of Moore's Law needs to be interesting and fun to read, on top of being top-notch.
As an experiment, I'm going to swear a lot in the following bits, and the main thing I want feedback on is, was it fun to read?
Again, please note, I'm deliberately using swearing and derisive tone and violence, 4chan-style. No slurs. There are threats to imaginary animals.
If swearing and derision makes you uncomfortable, please skip out.
I'm warning you.
Why the FUCK are we so enamored about fucking ChatGPT? What the fuck does it have that any other fucking model doesn't? Holy fucking shit, we just keep making "ChatGPT does shiny new thing" my fucking GOD, there's literally nothing new there that hasn't existed for at least three fucking years.
Goddamn people making Yann Lecun seem reasonable with this horseshit. If I read yet-another-goddamn fucking "ChatGPT is cool!" or "GPT-3 is sooo awesooomee" paper on arxiv I'm going to drown a fucking cat. I don't know's cat. I don't even have cat! Goddamn, cats don't deserve this.
Don't be the reason a cat gets drowned.
No, seriously, what-the-actual fuck is going on? If OpenAI actually had anything actually-fucking-impressive, don't you think they'd call it fucking GPT-4 if they could have? Of course they-fucking would have made a fucking GPT-4 holy fucking shit. But no, they didn't.
What the actual fuck are they waiting for?
They just hired a bunch of fucking chumps to label some data and act in a nice way, and a bunch of other inane and totally-ordinary means of data curation. This fucking shit means literally nothing for ai progress, and actually implies making the model bigger has stopped fucking working.
No seriously, think for a fucking second.
Would OpenAI not have done GPT-4 if they could have? The company already fucking loves their headlines. What company wouldn't-fucking want a headline that ChatGPT was the NEXT GENERATION AI "you're playing with some goddamn new fucking ai" But no.
Where the fuck is it?
Same with literally any other ai group. Where the fuck is our scaling?
Dead! Dead I say.
"But I'm not dead yet!"
Shut the fuck up, you will be shortly.
(Inb4 Openai releases GPT-4 within the next fucking month, lol)
(I don't think it'll pass my fucking rigged test though)
Data Size, FLOPs, GPU RAM, Data Quality, the quadro-fucking-fecta of the fucking Bitter Lesson. Yeah yeah.
I made a Model with 500 Billion Parameters and Not Even A T-Shirt to Show For It. GPT-3.5 gets beaten by 6B models finetuned on a task.
Memory is still pretty fucking garbage.
"Is it real ai yet!?!"
No, goddamn, and these goddamn rules and test are fucking rigged against you.
"Oh, so you're just gonna play Gary Marcus, Part Two"
No, I'm not a fucking pundit.
So, tired of being nickel-and-dimed by "is it real ai yets", and not wanting to become yet another ai-pundit, I'm rigging this fucking game so I can claim victory regardless of what fucking happens.Concretely:
WHERE THE FUCK IS THE FUCKING MAP FOR FUCKING OPEN-ENDEDNESS?
MY AIs ALL FUCKING BREAK DOWN WHEN A SINGLE OUNCE OF OUT-OF-DISTRIBUTION SHIT THREATEN TO BREATHE ON THEM.
Here's A Goddamn TEST, And It's Fucking Rigged If You're Trying To Make Money
The Rules:
See what I mean? I'm rigging this whole FUCKING thing!
Cool, that was fun. But I started slipping. I can't swear a lot for long, goddamn.
Number 3 and 4 are the most important ones. If a third party verifies that the neural network is an entire, self-contained ai, and not a clever, manually-coded lookup table or tied into <elasticsearch>, then I claim defeat.
If it's behind a private paywall and no third parties can independently verify that the ai exists, then that leaves room for pulling something like what Waze does- where the cars require high-throughput low-latency internet connection 24/7 and there's (extreme probability, like 90% confident) an office of people hired to monitor and drive the cars if there's even a chance of rain or fog.
If the comparison is self-driving cars, if the car requires internet connection in order to make it from Point A, to Point B, those gains are marginal hax. I should be able to rip out any wifi or cellular connections in my car and know for-fucking-sure, that the car will still be able to get me from one end of Canada or another without any cell signal. In the middle winter.
If it's a single test over a weekend, the ability to just defect in a sanitized, preprogrammed scenario is high. But we can't verify that for Waze. They own the vehicles, and, again, given the incentives to defect, I wouldn't consider a single demonstration to be enough.
Again, this test is fuckin' rigged.
Okay, now you're probably thinking:
What would make me 100% admit I'm wrong?
If, in 6 years, I can buy/rent an N GPUs/clusters, download the weights from huggingface on my desktop, run the model, write down instructions on a piece of paper for the model. Then, keep the model running, have the various instructions performed over time.
Example instructions that I might write down might be:
If the ai can do this- say, follow "in two days" it just needs to happen any time on that target day. Similarly, "in an hour" it just needs to happen any time in that hour, and the logic which checks the time and then decides whether to do the thing needs to happen in the weights.
No hax.
If this is a simple, clever system/explicit behavior tree that uses regexp and cleverly swaps between Riffusion and Stable Diffusion, it doesn't sink my theory, but is instead, an example of it.
Regardless of all of the above, The Test Ends against my favor if, in 6 years:
If 6 years rolls around, and no one's really sure or we can't come to an agreement, it resolves in my favor.
Thank you for participating in this goddamn-fucking-experiment.
Let me know what you think would improve it.