Comment by nateb2022
10 hours ago
> AI is going to be a highly-competitive, extremely capital-intensive commodity market
It already is. In terms of competition, I don't think we've seen any groundbreaking new research or architecture since the introduction of inference time compute ("thinking") in late 2024/early 2025 circa GPT-o4.
The majority of the cost/innovation now is training this 1-2 year old technology on increasingly large amounts of content, and developing more hardware capable of running these larger models at more scale. I think it's fair to say the majority of capital is now being dumped into hardware, whether that's HBM and research related to that, or increasingly powerful GPUs and TPUs.
But these components are applicable to a lot of other places other than AI, and I think we'll probably stumble across some manufacturing techniques or physics discoveries that will have a positive impact on other industries.
> that ends up in a race to the bottom competing on cost and efficiency of delivering
One could say that the introduction of the personal computer became a "race to the bottom." But it was only the start of the dot-com bubble era, a bubble that brought about a lot of beneficial market expansion.
> models that have all reached the same asymptotic performance in the sense of intelligence, reasoning, etc.
I definitely agree with the asymptotic performance. But I think the more exciting fact is that we can probably expect LLMs to get a LOT cheaper in the next few years as the current investments in hardware begin to pay off, and I think it's safe to assume that in 5-10 years, most entry-level laptops will be able to manage a local 30B sized model while still being capable of multitasking. As it gets cheaper, more applications for it become more practical.
---
Regarding OpenAI, I think it definitely stands in a somewhat precarious spot, since basically the majority of its valuation is justified by nothing less than expectations of future profit. Unlike Google, which was profitable before the introduction of Gemini, AI startups need to establish profitability still. I think although initial expectations were for B2C models for these AI companies, most of the ones that survive will do so by pivoting to a B2B structure. I think it's fair to say that most businesses are more inclined to spend money chasing AI than individuals, and that'll lead to an increase in AI consulting type firms.
> in 5-10 years, most entry-level laptops will be able to manage a local 30B sized model
I suspect most of the excitement and value will be on edge devices. Models sized 1.7B to 30B have improved incredibly in capability in just the last few months and are unrecognizably better than a year ago. With improved science, new efficiency hacks, and new ideas, I can’t even imagine what a 30B model with effective tooling available could do in a personal device in two years time.
Very interested in this! I'm mainly a ChatGPT user; for me, o3 was the first sign of true "intelligence" (not 'sentience' or anything like that, just actual, genuine usefulness). Are these models at that level yet? Or are they o1? Still GPT4 level?
Not nearly o3 level. Much better than GPT4, though! For instance Qwen 3 30b-a3b 2507 Reasoning gets 46 vs GPT 4's 21 and o3's 60-something on Artificial Analysis's benchmark aggregation score. Small local models ~30b params and below tend to benchmark far better than they actually work, too.
> One could say that the introduction of the personal computer became a "race to the bottom." But it was only the start of the dot-com bubble era, a bubble that brought about a lot of beneficial market expansion.
I think the comparison is only half valid since personal computers were really just a continuation of the innovation that was general purpose computing.
I don't think LLMs have quite as much mileage to offer, so to continue growing, "AI" will need at least a couple step changes in architecture and compute.
I don't think anyone knows for sure how much mileage/scalability LLMs have. Given what we do know, I suspect if you can afford to spend more compute on even longer training runs, you can still get much better results compared to SOTA, even for "simple" domains like text/language.
I think we're pretty much out of "spend more compute on even longer training runs" atp.
> But I think the more exciting fact is that we can probably expect LLMs to get a LOT cheaper in the next few years as the current investments in hardware begin to pay off
Citation needed!