6 ms·

LLMs have been improving exponentially for a few years. let's at least wait until exponential improvements slow down to make a judgement about their potential

bloppe 2y ago
They have been improving a lot, but that improvement is already plateauing and all the fundamental problems have not disappeared. AI needs another architectural breakthrough to keep up the pace of advancement.
- Animats 2y ago
  Yes. Anything on the horizon?
  - bloppe 2y ago
    I'm not as up-to-speed on the literature as I used to be (it's gotten a lot harder to keep up), but I certainly haven't heard of any breakthroughs. They tend to be pretty hard to predict and plan for. I don't think we can continue simply tweaking the transformer architecture to achieve meaningful gains. We will need new architectures, hopefully ones that more closely align with biological intelligence. In theory, the simplest way to real superhuman AGI would be to start by modeling a real human brain as a physical system at the neural level; a real neural network. What the AI community calls "neural networks" are only very loose approximations of biological neural networks. Real neurons are subject to complex interactions between many different neurotransmitters and neuromodulators and they grow and shift in ways that look nothing like backpropagation. There already exist decently accurate physical models for single neurons, but accurately modeling even C. elegans (as part of the OpenWorm project) is still a way's off. Modeling a full human brain may not be possible within our lifetime, but I also wouldn't rule that out. And once we can accurately model a real human brain, we can speed it up and make it bigger and apply evolutionary processes to it much faster than natural evolution. To me, that's still the only plausible path to real AGI, and we're really not even close.
    - segasaturn 2y ago
      I was holding out hope for Q*, which OAI talked about with hushed tones to make it seem revolutionary and maybe even dangerous, but that ended up being o1. o1 is neat, but its far from a breakthrough. It's just recycling the same engine behind GPT-4 and making it talk to itself before spitting out its response to your prompt. I'm quite sure they've hit a ceiling and are now using smoke-and-mirrors techniques to keep the hype and perceived pace-of-progress up.
  - schleck8 2y ago
    OpenAI's Orion (GPT 5/Next) is partially trained on synthetic data generated with a large version of o1. Which means if that works the data scarcity issue is more or less solved.
- amelius 2y ago
  If they were plateauing it would mean OpenAI would have lost its headstart wrt the competition, which is not the case I believe.
  - talldayo 2y ago
    It's pretty undeniable that OpenAI's lead has been diminished greatly from the GPT-3 days. Back then, they could rely on marketing their coherency and the "true power" of larger models. But today we're starting to see 1B models that are undistinguishable from OpenAI's most advanced chain-of-thought models. From a turing test perspective, I don't think the average person could distinguish between an OpenAI and a Llama 3.2 response.
  - bloppe 2y ago
    OpenAI has the biggest appetite for large models. GPT-4 is generally a bit better than Gemini, for example, but that's not because Google can't compete with it. Gemini is orders of magnitude smaller than GPT-4 because if Google were to run a GPT-4-sized model every time somebody searches on Google, they would literally cease to be a profitable company. That's how expensive inference on these ultra-large models is. OpenAI still doesn't really care about burning through hundreds of billions of dollars, but that cannot last forever.
    - bunderbunder 2y ago
      This, I think, is the crux of it. OpenAI is burning money at a furious rate. Perhaps this is due to a classic tech industry hypergrowth strategy, but the challenge with hypergrowth strategies is that they tend to involve skipping over the step where you figure out if the market will tolerate pricing your product appropriately instead of selling it at a loss. At least for the use cases I've been directly exposed to, I don't think that is the case. They need to keep being priced about where they are right now. It wouldn't take very much of a rate hike for their end users to largely decide that not using the product makes more financial sense.
  - hatthew 2y ago
    OpenAI had a noticeable head start with GPT-2 in 2019. They capitalized on that head start with ChatGPT in late 2022, and relatively speaking they plateaued from that point onwards. They lost that head start 2.5 months later with the announcement of Google Bard, and since then they've been only slightly ahead of the curve.
- famouswaffles 2y ago
  >but that improvement is already plateauing Based on what ? The gap between the release of GPT-3 and 4 is still much bigger than the time that has elapsed since 4 was already released so really, Based on what ?
  - riku_iki 2y ago
    there are no much reliable benchmarks which would measure what is gap really. I think currently corps compete in who will leak benchmarks to training data the most, hence o1 is world programming medalist, yet makes stupid mistakes.
COAGULOPATH 2y ago
In some domains (math and code), progress is still very fast. In others it has slowed or arguably stopped. We see little progress in "soft" skills like creative writing. EQBench is a benchmark that tests LLM ability to write stories, narratives, and poems. The winning models are mostly tiny Gemma finetunes with single-digit parameter counts. Huge foundation models with hundreds of billions of parameters (Claude 3 Opus, Llama 3.1 405B, GPT4) are nowhere near the top. (Yes, I know Gemma is a pruned Gemini). Fine-tuning > model size, which implies we don't have a path to "superhuman" creative writing (if that even exists). Unlike model size, fine-tuning can't be scaled indefinitely: once you've squeezed all the juice out of a model, what then? OpenAI's new o1 model exhibits amazing progress in reasoning, math, and coding. Yet its writing is worse than GPT4-o's (as backed by EQBench and OpenAI's own research). I'd also mention political persuasion (since people seem concerned about LLM-generated propaganda). In June, some researchers tested LLM ability to change the minds of human subjects on issues like privatization and assisted suicide. Tiny models are unpersuasive, as expected. But once a model is large enough to generate coherent sentences, persuasiveness kinda...stops. All large models are about equally persuasive. No runaway scaling laws are evident here. This picture is uncertain due to instruction tuning. We don't really know what abilities LLMs "truly" possess, because they've been crippled to act as harmless, helpful chatbots. But we now have an open-source GPT-4-sized pretrained model to play with (Llama-3.1 405B base). People are doing interesting things with it, but it's not setting the world on fire.
- fluoridation 2y ago
  >But once a model is large enough to generate coherent sentences, persuasiveness kinda...stops. All large models are about equally persuasive. No runaway scaling laws are evident here. Isn't that kind of obvious? Even human speakers and writers have problems changing people's minds, let alone reliably.
  - klipt 2y ago
    The only people who changed minds reliably were Age of Empires priests. Wololo, wololo!
  - lmm 2y ago
9cb14c1ec0 2y ago
I can't think of any exponential improvements that have happened recently.
bamboozled 2y ago
Do you know what exponential means? They might be getting getting but it hardly seems exponential at this stage.
rifty 2y ago
I don’t think you should expect exponential growth towards greater correctness past good enough for any given domain of knowledge it is able to mirror. It is reliant on human generated material, and so rate limited by the number of humans able to generate the quality increase you need - which decreases in availability as you expect higher quality. I also don’t believe greater correctness for any given thing is an open ended question that allows for experientially exponential improvements. Though maybe you are just using exponential figuratively in place of meaning rapid and significant development and investment.

bloppe 2y ago

They have been improving a lot, but that improvement is already plateauing and all the fundamental problems have not disappeared. AI needs another architectural breakthrough to keep up the pace of advancement.

Animats 2y ago
Yes. Anything on the horizon?
- bloppe 2y ago
  I'm not as up-to-speed on the literature as I used to be (it's gotten a lot harder to keep up), but I certainly haven't heard of any breakthroughs. They tend to be pretty hard to predict and plan for. I don't think we can continue simply tweaking the transformer architecture to achieve meaningful gains. We will need new architectures, hopefully ones that more closely align with biological intelligence. In theory, the simplest way to real superhuman AGI would be to start by modeling a real human brain as a physical system at the neural level; a real neural network. What the AI community calls "neural networks" are only very loose approximations of biological neural networks. Real neurons are subject to complex interactions between many different neurotransmitters and neuromodulators and they grow and shift in ways that look nothing like backpropagation. There already exist decently accurate physical models for single neurons, but accurately modeling even C. elegans (as part of the OpenWorm project) is still a way's off. Modeling a full human brain may not be possible within our lifetime, but I also wouldn't rule that out. And once we can accurately model a real human brain, we can speed it up and make it bigger and apply evolutionary processes to it much faster than natural evolution. To me, that's still the only plausible path to real AGI, and we're really not even close.
  - segasaturn 2y ago
    I was holding out hope for Q*, which OAI talked about with hushed tones to make it seem revolutionary and maybe even dangerous, but that ended up being o1. o1 is neat, but its far from a breakthrough. It's just recycling the same engine behind GPT-4 and making it talk to itself before spitting out its response to your prompt. I'm quite sure they've hit a ceiling and are now using smoke-and-mirrors techniques to keep the hype and perceived pace-of-progress up.
- schleck8 2y ago
  OpenAI's Orion (GPT 5/Next) is partially trained on synthetic data generated with a large version of o1. Which means if that works the data scarcity issue is more or less solved.
amelius 2y ago
If they were plateauing it would mean OpenAI would have lost its headstart wrt the competition, which is not the case I believe.
- talldayo 2y ago
  It's pretty undeniable that OpenAI's lead has been diminished greatly from the GPT-3 days. Back then, they could rely on marketing their coherency and the "true power" of larger models. But today we're starting to see 1B models that are undistinguishable from OpenAI's most advanced chain-of-thought models. From a turing test perspective, I don't think the average person could distinguish between an OpenAI and a Llama 3.2 response.
- bloppe 2y ago
  OpenAI has the biggest appetite for large models. GPT-4 is generally a bit better than Gemini, for example, but that's not because Google can't compete with it. Gemini is orders of magnitude smaller than GPT-4 because if Google were to run a GPT-4-sized model every time somebody searches on Google, they would literally cease to be a profitable company. That's how expensive inference on these ultra-large models is. OpenAI still doesn't really care about burning through hundreds of billions of dollars, but that cannot last forever.
  - bunderbunder 2y ago
    This, I think, is the crux of it. OpenAI is burning money at a furious rate. Perhaps this is due to a classic tech industry hypergrowth strategy, but the challenge with hypergrowth strategies is that they tend to involve skipping over the step where you figure out if the market will tolerate pricing your product appropriately instead of selling it at a loss. At least for the use cases I've been directly exposed to, I don't think that is the case. They need to keep being priced about where they are right now. It wouldn't take very much of a rate hike for their end users to largely decide that not using the product makes more financial sense.
- hatthew 2y ago
  OpenAI had a noticeable head start with GPT-2 in 2019. They capitalized on that head start with ChatGPT in late 2022, and relatively speaking they plateaued from that point onwards. They lost that head start 2.5 months later with the announcement of Google Bard, and since then they've been only slightly ahead of the curve.
famouswaffles 2y ago
>but that improvement is already plateauing Based on what ? The gap between the release of GPT-3 and 4 is still much bigger than the time that has elapsed since 4 was already released so really, Based on what ?
- riku_iki 2y ago
  there are no much reliable benchmarks which would measure what is gap really. I think currently corps compete in who will leak benchmarks to training data the most, hence o1 is world programming medalist, yet makes stupid mistakes.

COAGULOPATH 2y ago

In some domains (math and code), progress is still very fast. In others it has slowed or arguably stopped. We see little progress in "soft" skills like creative writing. EQBench is a benchmark that tests LLM ability to write stories, narratives, and poems. The winning models are mostly tiny Gemma finetunes with single-digit parameter counts. Huge foundation models with hundreds of billions of parameters (Claude 3 Opus, Llama 3.1 405B, GPT4) are nowhere near the top. (Yes, I know Gemma is a pruned Gemini). Fine-tuning > model size, which implies we don't have a path to "superhuman" creative writing (if that even exists). Unlike model size, fine-tuning can't be scaled indefinitely: once you've squeezed all the juice out of a model, what then? OpenAI's new o1 model exhibits amazing progress in reasoning, math, and coding. Yet its writing is worse than GPT4-o's (as backed by EQBench and OpenAI's own research). I'd also mention political persuasion (since people seem concerned about LLM-generated propaganda). In June, some researchers tested LLM ability to change the minds of human subjects on issues like privatization and assisted suicide. Tiny models are unpersuasive, as expected. But once a model is large enough to generate coherent sentences, persuasiveness kinda...stops. All large models are about equally persuasive. No runaway scaling laws are evident here. This picture is uncertain due to instruction tuning. We don't really know what abilities LLMs "truly" possess, because they've been crippled to act as harmless, helpful chatbots. But we now have an open-source GPT-4-sized pretrained model to play with (Llama-3.1 405B base). People are doing interesting things with it, but it's not setting the world on fire.

fluoridation 2y ago
>But once a model is large enough to generate coherent sentences, persuasiveness kinda...stops. All large models are about equally persuasive. No runaway scaling laws are evident here. Isn't that kind of obvious? Even human speakers and writers have problems changing people's minds, let alone reliably.
- klipt 2y ago
  The only people who changed minds reliably were Age of Empires priests. Wololo, wololo!
- lmm 2y ago

9cb14c1ec0 2y ago

I can't think of any exponential improvements that have happened recently.

bamboozled 2y ago

Do you know what exponential means? They might be getting getting but it hardly seems exponential at this stage.

rifty 2y ago

I don’t think you should expect exponential growth towards greater correctness past good enough for any given domain of knowledge it is able to mirror. It is reliant on human generated material, and so rate limited by the number of humans able to generate the quality increase you need - which decreases in availability as you expect higher quality. I also don’t believe greater correctness for any given thing is an open ended question that allows for experientially exponential improvements. Though maybe you are just using exponential figuratively in place of meaning rapid and significant development and investment.