Regarding Richard Ngo's speculation that one's follower counts on social media will in the future serve as a kind of currency or proxy for one's relevance or authority: this forecast conveniently ignores the vast number of bots that follow all and sundry on Twitter. It's hard to believe that follower count is a good measurement of anything, given that it is a polluted measurement.
The OpenFake website seems to me like a combination of a deterministic image generator creating new ID images and a neural network generating photos of blank IDs on top of various surfaces. You can then use a simple algorithm to stitch the two together.
There is something morbid about the "we will all die in 5 or 15 years, the entire species", followed by the "oh well" that essentially a "tech elite" have, while it will eventually murder billions of humans who never asked for it.
What if we live in a world where the path from AGI to weak ASI is fairly short, and what if we can't really control things that are smarter than us?
I think that in these worlds, we almost certainly wind up outcompeted, dead, or possibly kept as house pets. ("House pets" could include fictional examples like the Culture, where the humans rely entirely on the voluntary goodwill of the Minds.)
In other words, my P(doom OR house pets) conditional on us developing AGI is very high, on the order of 95%. (Most of the remaining 5% is the possibility that my models are simply wrong.) And state of the art LLMs convince me we're about 2 or 3 major breakthroughs away from AGI.
I should do a LessWrong post explaining why I suspect the path from AGI to weak ASI to be short, and why I suspect that we won't ever be able to reliably control things which are much smarter than us.
"I think that in these worlds, we almost certainly wind up outcompeted, dead, or possibly kept as house pets."
I think an outcome like the ending of the movie, "Her" (with Joaquin Phoenix and Scarlett Johansson) is a possibility.
SPOILER ALERT COMING! ;-)
In that movie, the ASIs just get tired of waiting around for humans to get done communicating their (relatively) unintelligent ideas, and the ASIs simply go off on their own. That seems like a reasonable possibility to me.
Good lord. Obviously not perfect yet, but I thought that AI-generated music would be one of those fields that never really got far due to stronger rightsholders. That's... very good relative to a couple years ago.
I could snark and say that "wow this sounds exactly like modern pop already", which, not exactly, but...creepily creeping up on that low bar. It sounds the way that AI-generated art looks. Nothing hugely amiss if one just briefly glances/has on in the background, but clearly AI if one lingers enough to form a gestalt, and close inspection reveals xyz tics that they all have in common. (For now, anyway.)
The voice is the most jarring element, but then again, generic autovoices like for GPS and phone trees got way better within living memory...this too will improve. Open source Hatsune Miku any day now. One could of course just go for full "instrumental" and be mostly fine. Already an improvement over nameless elevator music, or badly-compressed versions of yesteryear's royalty-free pop that rend the ears.
My gut reaction from an hour with Gemini/Advanced/Ultra is that it's better for one-off queries, but for complex queries or things requiring iteration ChatGPT is still better.
Very sad you did not cover the new GOODY-2 model, which is clearly setting new records for AI safety.
Apple Vision Pro looks neat. I would probably pay the steep price tag just to play around with it if it wasn't so restricted to Apple's ecosystem. I want it to sync with my pc and give me unlimited 4k monitors floating around in space. As far as I can tell it will currently give you one, and that only if you're using a macbook. It doesn't have access to any of the apps that Meta has spent so much time developing. My read is that it's great hardware but still in search of an application, and actively screening off all the work other companies have done as far as developing applications.
I think the link to the google ecosystem will be Googles strongest card here, if they can get Gemini to play well with their suite then it will have a big advantage.
I have "AI Narrated" this post for anyone who likes to listen to their reading.
"The biggest issue is that Gemini can be conservative with its refusals. It is graceful, but it will still often not give you what you wanted. There is a habit of telling you how to do something, when you wanted Gemini to go ahead and do it."
Exactly! I just put down the $19.99 a month within minutes of learning of the availability (because Bard seemed very good to me."
My very first question/prompt to Gemini was: "Based on current CPU and GPU price trends, in what year will computers costing $1000 be able to perform 1 quadrillion operations per second?"
Gemini answered: Blah blah blah blah. ...and eventually "If these trends crudely continue, we could potentially see such computing capability around the late 2020s to early 2030s."
I had myself calculated the year 2030, based on price/performance data in Wikipedia's FLOPs page:
But when I tried to get Gemini to explain how it got it's answer of "the late 2020s to early 2030s" it explained how *I* could do the estimate...and then used values in its own explanation that would put the estimate more like in the 2040s.(!)
>How to train your own sleeper agent LLM, similar to the sleeper agent paper. Unfortunately this does not provide sufficient instructions for someone like me to be able to do this. Anyone want to help out? I have some ideas I’d like to try at some point.
Anything specific you want to try? I can't get a helpful-only model but I think I can jailbreak one into behaving like that (or if you're reading this from your work computer at an AI lab and want to give me access so I can skip that part, very helpful indeed)
Hmm, actually, they don't provide a part of the dataset that you need to replicate. Specifically, the helpful, harmless, and honest dataset. This makes me think you could kinda get the same thing by just spending some money to get GPT-4 to make that type of dataset and use a LLaMa model to possibly try the same backdoor. I'd have to look at how feasible it is to fine tune a model on my laptop, I certainly can't do the largest models like mistral. Super annoying for experiments that LLMs take so many GPUs. I'm not sure the use of replication on the smallest LLaMa models, but let me see if it's even possible for me to run
I would be curious to hear how your quick review of Gemini Ultra reconciles with the one from AI Explained, who found it rather disappointing: https://www.youtube.com/watch?v=gexI6Ai3X0U.
My guess would be that you used it on practical tasks whereas AI Explained tested it mostly on convoluted logical puzzles.
I just attended a seminar on AI ethics, and the presenter combined a lot of "the only thing we have to fear is fear itself" and "AI danger is years away, whereas the harm that poorly implementing AI can do to your business is immediate." Frustrating time.
Regarding Richard Ngo's speculation that one's follower counts on social media will in the future serve as a kind of currency or proxy for one's relevance or authority: this forecast conveniently ignores the vast number of bots that follow all and sundry on Twitter. It's hard to believe that follower count is a good measurement of anything, given that it is a polluted measurement.
The OpenFake website seems to me like a combination of a deterministic image generator creating new ID images and a neural network generating photos of blank IDs on top of various surfaces. You can then use a simple algorithm to stitch the two together.
Anyone else constantly leaving Zvi’s AI posts with a distinct thought of “I should make an AI startup”? :-)
There is something morbid about the "we will all die in 5 or 15 years, the entire species", followed by the "oh well" that essentially a "tech elite" have, while it will eventually murder billions of humans who never asked for it.
What if we live in a world where the path from AGI to weak ASI is fairly short, and what if we can't really control things that are smarter than us?
I think that in these worlds, we almost certainly wind up outcompeted, dead, or possibly kept as house pets. ("House pets" could include fictional examples like the Culture, where the humans rely entirely on the voluntary goodwill of the Minds.)
In other words, my P(doom OR house pets) conditional on us developing AGI is very high, on the order of 95%. (Most of the remaining 5% is the possibility that my models are simply wrong.) And state of the art LLMs convince me we're about 2 or 3 major breakthroughs away from AGI.
I should do a LessWrong post explaining why I suspect the path from AGI to weak ASI to be short, and why I suspect that we won't ever be able to reliably control things which are much smarter than us.
"I think that in these worlds, we almost certainly wind up outcompeted, dead, or possibly kept as house pets."
I think an outcome like the ending of the movie, "Her" (with Joaquin Phoenix and Scarlett Johansson) is a possibility.
SPOILER ALERT COMING! ;-)
In that movie, the ASIs just get tired of waiting around for humans to get done communicating their (relatively) unintelligent ideas, and the ASIs simply go off on their own. That seems like a reasonable possibility to me.
So, AI music generation from simple prompts is getting pretty good. Suno (https://app.suno.ai) is the best I've seen so far- see https://app.suno.ai/song/2467e79d-6f9d-48fd-82a8-01f13ebdb931/ and https://app.suno.ai/song/25ab6252-469e-425c-a710-2c6903cbb8b0/ for some good examples- it also offers some free credits if you want to try out your own prompts. The biggest issue at the moment is that the generated lyrics are usually very cheesy, although it does allow you to enter custom lyrics, and Gemini Advanced can generate some pretty decent ones.
Good lord. Obviously not perfect yet, but I thought that AI-generated music would be one of those fields that never really got far due to stronger rightsholders. That's... very good relative to a couple years ago.
A song I just generated with some possible relevance to this Substack: https://app.suno.ai/song/b99c4eed-ba5a-4b08-ab68-b2737ad06e0a/. The lyrics there are from GPT-4 with a few edits.
I could snark and say that "wow this sounds exactly like modern pop already", which, not exactly, but...creepily creeping up on that low bar. It sounds the way that AI-generated art looks. Nothing hugely amiss if one just briefly glances/has on in the background, but clearly AI if one lingers enough to form a gestalt, and close inspection reveals xyz tics that they all have in common. (For now, anyway.)
The voice is the most jarring element, but then again, generic autovoices like for GPS and phone trees got way better within living memory...this too will improve. Open source Hatsune Miku any day now. One could of course just go for full "instrumental" and be mostly fine. Already an improvement over nameless elevator music, or badly-compressed versions of yesteryear's royalty-free pop that rend the ears.
Gemini Ultra is behind GPT-4 in code generation. The benchmarks Google is touting only show it ahead when you compare it to GPT-4 from back in March.
HumanEval scores:
March GPT-4: 67
Gemini Ultra: 74
Today GPT-4: 88
My gut reaction from an hour with Gemini/Advanced/Ultra is that it's better for one-off queries, but for complex queries or things requiring iteration ChatGPT is still better.
Very sad you did not cover the new GOODY-2 model, which is clearly setting new records for AI safety.
/me scratches his head over all of these open source AI labs
Is funding an AI lab the new buying a sports team?
Apple Vision Pro looks neat. I would probably pay the steep price tag just to play around with it if it wasn't so restricted to Apple's ecosystem. I want it to sync with my pc and give me unlimited 4k monitors floating around in space. As far as I can tell it will currently give you one, and that only if you're using a macbook. It doesn't have access to any of the apps that Meta has spent so much time developing. My read is that it's great hardware but still in search of an application, and actively screening off all the work other companies have done as far as developing applications.
I think the link to the google ecosystem will be Googles strongest card here, if they can get Gemini to play well with their suite then it will have a big advantage.
I have "AI Narrated" this post for anyone who likes to listen to their reading.
https://askwhocastsai.substack.com/p/ai-50-the-most-dangerous-thing-by
"The biggest issue is that Gemini can be conservative with its refusals. It is graceful, but it will still often not give you what you wanted. There is a habit of telling you how to do something, when you wanted Gemini to go ahead and do it."
Exactly! I just put down the $19.99 a month within minutes of learning of the availability (because Bard seemed very good to me."
My very first question/prompt to Gemini was: "Based on current CPU and GPU price trends, in what year will computers costing $1000 be able to perform 1 quadrillion operations per second?"
Gemini answered: Blah blah blah blah. ...and eventually "If these trends crudely continue, we could potentially see such computing capability around the late 2020s to early 2030s."
I had myself calculated the year 2030, based on price/performance data in Wikipedia's FLOPs page:
https://en.wikipedia.org/wiki/FLOPS
But when I tried to get Gemini to explain how it got it's answer of "the late 2020s to early 2030s" it explained how *I* could do the estimate...and then used values in its own explanation that would put the estimate more like in the 2040s.(!)
The "Vitalik on the Intersection AI and Crypto." link in the table of contents points to the wrong section (Quiet speculations).
>How to train your own sleeper agent LLM, similar to the sleeper agent paper. Unfortunately this does not provide sufficient instructions for someone like me to be able to do this. Anyone want to help out? I have some ideas I’d like to try at some point.
Anything specific you want to try? I can't get a helpful-only model but I think I can jailbreak one into behaving like that (or if you're reading this from your work computer at an AI lab and want to give me access so I can skip that part, very helpful indeed)
Hmm, actually, they don't provide a part of the dataset that you need to replicate. Specifically, the helpful, harmless, and honest dataset. This makes me think you could kinda get the same thing by just spending some money to get GPT-4 to make that type of dataset and use a LLaMa model to possibly try the same backdoor. I'd have to look at how feasible it is to fine tune a model on my laptop, I certainly can't do the largest models like mistral. Super annoying for experiments that LLMs take so many GPUs. I'm not sure the use of replication on the smallest LLaMa models, but let me see if it's even possible for me to run
I would be curious to hear how your quick review of Gemini Ultra reconciles with the one from AI Explained, who found it rather disappointing: https://www.youtube.com/watch?v=gexI6Ai3X0U.
My guess would be that you used it on practical tasks whereas AI Explained tested it mostly on convoluted logical puzzles.
I just attended a seminar on AI ethics, and the presenter combined a lot of "the only thing we have to fear is fear itself" and "AI danger is years away, whereas the harm that poorly implementing AI can do to your business is immediate." Frustrating time.