The argument for current LLM AIs leading to AGI has always been that they would spontaneously develop independent reasoning, through an unknown emergent property that would appear as they scale. It hasn’t happened, and there’s no sign that it will.
That’s a dilemma for the big AI companies. They are burning through billions of dollars every month, and will need further hundreds of billions to scale further - but for what in return?
Current LLMs can still do a lot. They’ve provided Level 4 self-driving, and seem to be leading to general-purpose robots capable of much useful work. But the headwinds look ominous for the global economy, - tit-for-tat protectionist trade wars, inflation, and a global oil shock due to war with Iran all loom on the horizon for 2025.
If current AI players are about to get wrecked, I doubt it’s the end for AI development. Perhaps it will switch to the areas that can actually make money - like Level 4 vehicles and robotics.
Oh no, the technology that is literally just a glorified text prediction that gives you random guesses about what word comes next, based on what was in the text you trained it on, can not scale to have an independent reasoning?
Color me surprised, who would have thought?
But how can we be sure? Give us a couple more billion to waste on graphics cards first!
There are options, and a well-funded new player will burst on the scene just as ChatGPT made itself eponymous with language AI: https://open.substack.com/pub/qnfo/p/the-diminishing-returns-of-scaling
Pretty sure it was never going to become an AGI.
It’s just donkey-work machines to cut low level jobs.
I don’t think anyone in the industry thought LLMs were going to reach AGI. But LLMs will be useful as part of an AGI framework. That’s the current focus in the industry.
I mean, yeah, this is about people outside the industry, those who invested money.
To my knowledge, LLMs still don’t pay for themselves. When the hype dies down and investors aren’t willing to provide money anymore, then prices for LLMs will become prohibitive for many current use-cases. That will also shrink the industry.
What that means in effect, we’ll still have to see, but AGI was one path that investors hoped for to get towards profitability, so it doesn’t aid the hype when they’re slowly learning about reality.
I’m developing some human centric LLM frameworks at work. Every API request to OpenAI is currently subsidized by venture capital. I do worry about what the industry will look like once there is a big price adjustment. Locally run models are pretty decent now and the pace is still moving forward, especially with regards to context window sizes so as long as I keep the frameworks model agnostic it might not be a big impact.
Yep, based off public numbers regarding OpenAI finances, it was estimated that as of a few months ago they spent $2.35 for every $1 they made.
No they did for sure. Ofcourse that was conditional on a few things. Those things have yet to arrive and one major detractor is actually the lack of training data. Most of the public internet has been crawled by AI crawlers and half the new content is poisoned by AI making it worthless. Its gonna take a few more years to see if it keeps scaling or not.
Google, Microsoft and Amazon are all making heavy investments in nuclear power to run more GPUs. These aren’t the moves of companies who are about to taper off utilization.
They don’t expect to taper off. But they might just be burning money.
They might be, but that won’t slow them down for at least the next gen of NVIDIA hardware.
To be fair, most Americans don’t demonstrate independent thinking, regularly regurgitate entire phrases they’ve been fed without showing any cognitive understanding, and they also sometimes perform tasks useful to corporations.
Current LLMs can still do a lot. They’ve provided Level 4 self-driving, and seem to be leading to general-purpose robots capable of much useful work.
Really? I don’t think this has anything to do with LLMs. They are likely using reinforcement learning combined with traditional AI techniques, an approach which has been the foundation of these kinds of robotics and automation for decades at this point.
If other areas of AI and automation have seen a boost at the same time as LLMs came on the scene, it’s because the underlying hardware has become so much faster, cheaper and easily available, along with the massively increased interest in and funding for these types of research, and computer scientists re-skilling into a discipline that’s in the midst of a bubble.
Not entirely true, the big change was multi-headed attention and the transformer model.
It’s not just being used for language but anything where sequence and context patterns are really important. Some stuff is still using convolutional networks and RNNs etc. but transformers aren’t just for LLMs. There’s definitely a lot of algorithmic advances driving the wave of new ai implementations, not just hardware improvements.
Thanks for the clarification. The point remains that it’s not true to say that LLMs have “provided Level 4 self-driving and … general-purpose robots.”
Agreed. It’s a lot of the same tech that powers both, but it’s not like a self driving car contains a language model that’s fine tuned on the adventures of Steve MacQueen or something.
Self driving uses LLMs? Or a specific type of AI? If it’s LLM I wouldn’t trust it on a side street, since that’s not what a language model is designed for.
No, self-driving cars don’t use Large Language Models, they don’t need to process language at all.
Remember: the dot-com bubble didn’t end the internet.
This tech has cool uses, outside of venture-capitalist cult obsession. Mostly porn. I mean, Jesus Christ, so much porn.
But the general idea of image-to-image transformers, based on natural-language descriptions, is here, and it is witchcraft. Generating new images from scratch is a stupid demo gone feral. The real applications will be all-purpose “CGI” as an idiot-proof Photoshop filter. Select tin can on string, type “spaceship,” get decently plausible results for your no-budget sci-fi show. No roto, no modeling, no lighting. The machine makes excellent guesses. And if you disagree, well, run it again. If you want it to be a specific kind of spaceship, either build a little toy or drag a PNG across the screen, and the machine will try to fix the aspects which make that look stupid.
The applications that money-robots want will be what destroys their industry. Animators don’t want to type in “cartoon rabbit walking” and get a finished product the machine spat out - but they’d love to have their drawings tweened. They’re all busy mocking this “framerate upscaling” nonsense, and missing that it means they can put in an on-eights previs and have it come out as smooth or as choppy as they want. And then they can doodle over whichever parts they don’t like, and have a different model turn those sketches into on-model drawings. The ultimate outcome of which can look like any Pixar movie even if your process is entirely 2D. Or… it can look like live action. Starring real actors, living or dead. Or starring literal nobodies, as made-up as any animated character, but as plausible as any person on film.
We’re gonna see a repeat of the webcomic boom… for movies and shows. It simply will not cost one billion dollars to make a whole-ass media franchise. Expect this to completely surprise the lumbering giants who keep trying to get rid of the little people who made up the stories and the characters.
LLMs are not all that is ai
AI is synthetic cognition or synthetic knowledge. What we’re calling AGI must “know” the logic of our quantum universe first, which is all but incomprehensible to us.
It’s hard to reason with the inconsistent (sometime nonexistent) human written communication that LLMs know. There’s something deeper in other channels of communication that we use for our own logic.
They seriously thought IterationX would work?
Yeah, but we should let AI solve the problem of making itself better! Then it can solve everything from climate change to making fully self driving cars to figuring out the most efficient way to murder the whole planet!
/s
Maybe if the author wouldn’t write “AI did hit a wall” in 2022, when everything is just currently talking about diminishing return, then someone might habe taken him seriously a bit. However AI is complex and there are new approaches to speed up learning and result speed, different approaches to steer a model output. The tech is still too new to say what’s up next. So complex even, that we might have months or years with no significant upgrade until a break through. Other than that it just reads as if the author wants to get back their reputation after making himself look like a negative Nancy. People forget that even the brain has hallucinations, but also layers in place to correct them.
