ClaytonB
Member
- Joined
- Oct 30, 2011
- Messages
- 10,245
We are in the AI Singularity
- Thread starter ClaytonB
- Start date
ClaytonB
Member
- Joined
- Oct 30, 2011
- Messages
- 10,245
Yeah, I use it almost daily, my employer pays for a top-tier subscription for their employees, which is nice. It's a glorified RTFM but it saves me a lot of time.
I'm glad that AI is making it possible for non-techies to start accessing the power of computers. We need that. But AI doesn't magically make computers more powerful. They already had all that computing power before, it's just much easier to access now. I can do things I couldn't do before, just because it would have been time-prohibitive to go look up all the information in the manual to figure out all the little various function calls and syntax. AI cuts these kinds of tasks down from dozens or hundreds of hours, to an afternoon. And that's cool. But it's not magic. It really and truly cannot think in any meaningful sense of that word. It can collate, interpolate, remix, and so on, and it can do these things at a level that is above mere syntax. But it's not thinking.
And engineering is still hard, even with AI. Engineering isn't just about making something work. It's about making something that keeps working, even under adverse conditions. Just about anybody can build a rocket, it's not actually that hard. But making it do what you want, even when there are winds, even when you're lifting a payload, and so on, and so forth, is another matter altogether. Pre-training-based AI is simply not capable of thinking at that level, or in that way at all. It's simply not suited to task.
Will we have better AI systems in the future? Obviously. Pre-training-based LLMs have unlocked enormous potential for future AI development. However many years away Hollywood AI (e.g. HAL9000) was before, it's 10x closer now. So, LLMs have been a step-function in terms of what came before. I would compare Transformer-based AI to the Gutenberg Press. It's a really big deal. But Hollywood AI it is not.
Last edited:
ClaytonB
Member
- Joined
- Oct 30, 2011
- Messages
- 10,245
Kilcher is one of the best YT commentators on AI out there, he actually has the chops to get down into the details of the mathematics and does deep-dives into the papers on a regular basis. So, when he says AGI is not coming, it's not just some Random Internet Opinion. Here's more evidence exposing why he's saying this now, with the GPT-5 drop:
Gemini falls for the same trick:
As noted in one comment, it's a strange paradox that these LLMs can achieve high scores on the International Math Olympiad but are terrible at basic arithmetic. But the contradiction is not that difficult to understand... the IMO is testing the kinds of mathematical processes that a human can do quickly, which are relatively shallow operations. I'm not denigrating the Olympiad itself, it's impressive, it's just that the kind of skills being tested there are based on quick recall and relatively shallow mathematical constructs that can be memorized and quickly mentally configured on-the-fly. The Olympiad has absolutely nothing whatsoever to do with the kind of work done by someone like Grigori Perelman, for example, other than the word "Math". So, the IMO is a great way for the AI companies to show off flashy statistics like "XYZ model scored Silver in IMO!" but does not give any real insight into the actual thinking abilities of these systems. Their real capabilities turn out to be extremely brittle... even very slight deviations out-of-distribution can reduce GPT-4 or even GPT-5 to a stuttering moron. And this is precisely the opposite of what we really want in AI systems.
An AI that's about as intelligent as a high-functioning mentally-disabled person would be ideal for 90+% of use-cases I would want to use an AI for. I already believe in God, I don't need a digital phantasm to be a digital idol for me. Most of the AI hype is ultimately being driven by social and spiritual forces, not economic/technological ones. Someone who is mentally handicapped but knows how to make coffee, check the mail, tell an intruder to get lost or they'll call 911, and so on and so forth, is 10 million times more useful than GPT-5 could ever be for the simple reason that they truly know and understand the things they know and understand. They are robust, living people and so the simple jailbreaks by which you can collapse an AI into blubbering gibberish won't work on them. AI is like the Storm Troopers when Obi-wan waved his hand and said, "These are not the droids you are looking for". In most cases, it is not even slightly difficult to jailbreak the AI and get it to say all kinds of insane things. Just write a prompt-equivalent of the Jedi hand wave and boom, GPT will start blubbering gibberish like a babbling brook.
GPT-5 proves that the era of data scaling is over. That doesn't mean there won't be future progress. But the illusion of "exponential progress" has been shattered. The real question is whether they will even be able to achieve linear progress...
ClaytonB
Member
- Joined
- Oct 30, 2011
- Messages
- 10,245
Register: AWS CEO says using AI to replace junior staff is 'Dumbest thing I've ever heard' -- They're cheap and grew up with AI … so you're firing them why?
Voluntarist
Member
- Joined
- Sep 10, 2011
- Messages
- 3,834
I'm pretty sure we're the AI users and RPF is ChatGPT.
We'll know for certain if TheTexan replies with "That's very inciteful; would you like me to develop a business plan for that?"
We'll know for certain if TheTexan replies with "That's very inciteful; would you like me to develop a business plan for that?"
Last edited:
ClaytonB
Member
- Joined
- Oct 30, 2011
- Messages
- 10,245
AI is auto-tune for the mind
I grew up during the 1990's. This was during the rise of auto-tune. I didn't listen to a lot of pop-music but I did like heavy rock and a selection of metal (mostly nu-metal). These genres tended to use auto-tune quite liberally and, during this era, this resulted in a lot of objectively inferior talent rising to the top of the charts. "Sounds great on the CD, sounds terrible live" was a common phenomenon.
This still happens today, but it's less common than it used to be. And that's the point of this post. AI is going to do to a lot of knowledge-work and media-production work what auto-tune did to music in the 90's and early 00's. It's going to result in lots of very inferior talents getting boosted to the tops of the charts. We will see "rockstar" programmers, artists, etc. who, in terms of their actual personal talents, are very marginal. But they're good at making the AI bump the party. They will be more DJs of knowledge-work, more than actual knowledge-workers.
Lots of DJs have no actual musical creation ability at all... they cannot sing, they cannot compose, they cannot perform any instrument at all ... but they can mix. That's also a talent, don't get me wrong, even if it is just mixing at the desk (not live-mixing / performance DJing). All valid stuff. And so I acknowledge that "being good at AI", that is, being skilled at using AI to achieve certain tasks, rather than doing it manually, will also be its own skill, kind of like DJing is, in respect to music. But let's not get it twisted -- being a DJ doesn't make you a Mozart, it doesn't make you a Kirk Hammett. Being good at doing things using AI won't make you actually good at doing those things. You might be able to do both... both have a talent and use the AI to also perform, kind of like someone who is a keyboardist and a DJ at the same time.
Long run, my hope is that the AI faker phenomenon will start to fade out as people realize that faking is objectively inferior to doing the hard work of actually building a talent. Yeah, there is a certain "economic scaling" that fakers can take advantage of that blood-sweat-and-tears talent cannot match, because it takes years of invested effort to build real talent, whereas the fakers can just "plug and play" and start immediately feeding into whatever is viral at this second in history. But just like auto-tune leads to disappointed fans who discover that their CD music heroes suck at live performance when they buy tickets and go to the live concert, the "AI autotune" is going to have the same effect. People are going to be hired for their "rockstar engineer" resume, only for management to eventually realize they've hired a dud who is just a master of AI-autotune.
Don't let the transhumanist-depop psychos dictate the narrative around AI. They're lying their absolute pants off about AI. Ignore the hype, focus on what is real.
I grew up during the 1990's. This was during the rise of auto-tune. I didn't listen to a lot of pop-music but I did like heavy rock and a selection of metal (mostly nu-metal). These genres tended to use auto-tune quite liberally and, during this era, this resulted in a lot of objectively inferior talent rising to the top of the charts. "Sounds great on the CD, sounds terrible live" was a common phenomenon.
This still happens today, but it's less common than it used to be. And that's the point of this post. AI is going to do to a lot of knowledge-work and media-production work what auto-tune did to music in the 90's and early 00's. It's going to result in lots of very inferior talents getting boosted to the tops of the charts. We will see "rockstar" programmers, artists, etc. who, in terms of their actual personal talents, are very marginal. But they're good at making the AI bump the party. They will be more DJs of knowledge-work, more than actual knowledge-workers.
Lots of DJs have no actual musical creation ability at all... they cannot sing, they cannot compose, they cannot perform any instrument at all ... but they can mix. That's also a talent, don't get me wrong, even if it is just mixing at the desk (not live-mixing / performance DJing). All valid stuff. And so I acknowledge that "being good at AI", that is, being skilled at using AI to achieve certain tasks, rather than doing it manually, will also be its own skill, kind of like DJing is, in respect to music. But let's not get it twisted -- being a DJ doesn't make you a Mozart, it doesn't make you a Kirk Hammett. Being good at doing things using AI won't make you actually good at doing those things. You might be able to do both... both have a talent and use the AI to also perform, kind of like someone who is a keyboardist and a DJ at the same time.
Long run, my hope is that the AI faker phenomenon will start to fade out as people realize that faking is objectively inferior to doing the hard work of actually building a talent. Yeah, there is a certain "economic scaling" that fakers can take advantage of that blood-sweat-and-tears talent cannot match, because it takes years of invested effort to build real talent, whereas the fakers can just "plug and play" and start immediately feeding into whatever is viral at this second in history. But just like auto-tune leads to disappointed fans who discover that their CD music heroes suck at live performance when they buy tickets and go to the live concert, the "AI autotune" is going to have the same effect. People are going to be hired for their "rockstar engineer" resume, only for management to eventually realize they've hired a dud who is just a master of AI-autotune.
Don't let the transhumanist-depop psychos dictate the narrative around AI. They're lying their absolute pants off about AI. Ignore the hype, focus on what is real.
Similar threads
