[ryanackley]

AI maximalism is making a lot of assumptions that I think are not a given

* The curve of AI improvement will continue at the current pace

* AI companies will have the capital continue to expand infrastructure

* there will be some kind of functioning economy if all knowledge workers are replaced

There are strong headwinds to all three of these.

Hey it may come to pass but it’s very speculative at this point. I see a lot of tech people simply overlaying the progress curve of previous tech booms which is reductive.

[onlyrealcuzzo]

> * The curve of AI improvement will continue at the current pace

Frontier AI is already good enough to be very useful for engineering. It's too costly for many places where it could be useful today.

The cost for the same quality of output is going to drop at least 10x over the next 18-24 months.

And likely again in the following 18-24 months.

At the same time, the cost per watt is going to down ~25%, and at the same time speed will increase (also valuable since time is money).

[coffeefirst]

> The cost for the same quality of output is going to drop at least 10x over the next 18-24 months.

How do you know that?

In 2026 the prices have been spiking. It now costs orders of magnitude more than it did in November.

[Ukv]

Price of the current frontier may vary, but price for a given level of capability tends to drop pretty fast.

April of last year you'd get 1431 ELO[0] from o3-2025-04-16 for $8.00 per million output tokens. April of this year you can get 1436 ELO from deepseek-v4-flash for $0.2 per million output tokens.

[0]: https://huggingface.co/spaces/lmarena-ai/arena-leaderboard

[saxenaabhi]

Sure, but i don't think it's reasonable to hold given level of capability constant in a landscape where a give consumer of AI also has competitive pressures.

I can't use last year's SOTA model when my competitors can use the current SOTA model.

This is also baked in the eye watering valuations of model companies.

[margalabargala]

> I can't use last year's SOTA model when my competitors can use the current SOTA model.

Lots of people can. Tools don't need to be top of the line to be useful. Snap-on may exist, but they don't put Harbor Freight out of business.

Advanced IDEs exist but complex projects were still built in vim.

The more capable the budget models get, the lower the marginal gains from using the frontier models, even if the frontier models always stay 6 months ahead.

[onlyrealcuzzo]

> I can't use last year's SOTA model when my competitors can use the current SOTA model.

You can use open source models of equivalent or better capabilities for ~90% less cost...

If you kick and scream hard enough, you can always find a data point to make sure you're correct.

No one is saying that the Opus model last year costs 90% less now than it does this year.

That's not how it works.

There are better, more efficient models with equivalent capabilities that are 90% cheaper (see DeepSeek v4 Pro).

[rzmmm]

The ranking is not comparable across time like that.

[Ukv]

I'm using the current ELO of the models, and both are still running in the arena.

[Denzel]

Aren’t DeepSeek models deliberately priced lower than the cost to deliver? They’re subsidized which means the true cost is more than $0.2/Mtok.

[Ukv]

DeepSeek models are open-source so there are a bunch of third-party providers offering similar prices. Factoring in that DeepSeek have to train the model (whereas third parties can make a small profit over just the inference costs) I'd assume that on net they're spending investor money, but I wouldn't think that's any less true of OpenAI.

[Denzel]

Yes, DeepSeek is open-weight, but these third-party providers offering similar prices are subsidized with VC money as well. And you can find a range of prices for deepseek-v4-flash going up to and over $1/Mtok.

Even that $1/Mtok provided by Together AI is heavily subsidized by more than $1B in VC money.

This makes it unclear how the true cost curve is progressing. It’s not possible to confidently comment one way or another on the rate that cost is coming down when the entire industry is so heavily subsidized.

[onlyrealcuzzo]

> How do you know that?

Historic trends, every 18 months, performance for the same level of quality has gone down 90%.

See: https://www.reddit.com/r/LocalLLaMA/comments/1gpr2p4/llms_co...

And Chart 13 here: https://www.rdworldonline.com/ais-great-compression-20-chart...

And here: https://epoch.ai/data-insights/llm-inference-price-trends

The technology already exists now on the algorithmic front for the next 10x drop between everyone adopting DeepSeek's MLA, MoE (mostly already done), Medusa (a better version of Google's speculative decoding), Kimi's Attn Residuals, and Mimo's Sliding Window Attn, and (possibly) Microsoft's 1.58b (this may be a nothing burger).

Historically, algorithmic gains are only ~30% of the pie, but there's enough out there to get to 10x, with just what's available already. The other ~70% of the pie is better training data (often synthetic) and distilling frontier knowledge. There's no sign we are tapped out on that front.

> In 2026 the prices have been spiking.

That's not for the SAME level of output...

[Der_Einzige]

MoE isn’t the magical improvement you think it is. Logprobs of MoE models are always worse in quality than the dense equivalent and they struggler harder at very long context quality than equivalent dense models. This is why Chinese companies like qwen are releasing dense and MoE versions of their models at near equivalent sizes. I always use/prefer the dense one.

Speculative decoding usually only improves decode and sometimes actually harm prefill and for agentic coding prefill matters more.

You’re right about the rest but I need to set the record straight on these details.

[senordevnyc]

It now costs orders of magnitude more than it did in November.

Really? Care to do the math for me? Just curious about exactly how many orders of magnitude it's gone up.

[DonsDiscountGas]

AI/LLMs have been dramatically improving for 7+ years. There's now a lot more funding to support continued improvement. You're correct this is an "assumption", but continued improvement at the same pace (or faster) for the next 3+ years is just extrapolating a trend. Believing we've hit the top today is based on nothing at all. Continued improvement is much more likely.

[cloche]

You can only tell which part of the S-Curve you're on in retrospect. It's not something you can tell while you're experiencing it. Both scenarios of AI maxing out or continuing to improve are both likely.

[somebodythere]

That is not true. You can tell you are on the latter part of the S-Curve you are on, if the rate of change of capabilities has decreased compared to before. That is not what we are seeing right now. The rate of change is increasing, or is at best, stable.

[vrighter]

you compare the derivative yesterday, you measure the derivative today. difference between them is the jerk. It tells you where you are on the S-curve

[saltcured]

It's only in retrospect that you know the level of filtering required to find this S curve trend among the noisy perturbations.

Just like people trying to time the market with "technical analysis". It is extremely easy to find whatever pattern you are looking for. A lot harder to accurately distinguish predictive power from fantasy.

[alfalfasprout]

This is probably one of the more level headed takes in the comment thread. There's been a concerted marketing push to frame AI maximalism as an inevitability. More or less a "it's going happen anyways so let's go all in".

It's hardly an inevitability though (nothing is... and analogues to the industrial revolution are iffy at best, we haven't ever had an attempted replacement for intelligence itself before).

Society is doing this at an unprecedented cost and it's clear a large portion of the population is uneasy with it. Whether society in the US, Europe, and Asia will continue to allow such investment at the expense of everything else remains to be seen.

[ryanackley]

The capital outlay is eye watering and it feels like an over extension. Neither of the two major pure AI providers are close to profitable (OpenAI, Anthropic) while having valuations close to a trillion dollars.

Their ROI is paradoxical. If they succeed in disrupting knowledge work. Who will be around to use or buy their product?

[hodgehog11]

Others have commented on the rate of AI improvement. It doesn't need to be current rate for it to be an even more serious problem in the very near future. That's irrespective of prior booms.

Regarding AI companies having capital to expand infrastructure; this is largely irrelevant. The cat is out of the bag, and you can already make serious gains by finetuning to local problems on a desktop machine. There is enough hardware out there to run these things en masse; it's more a question of power. Regardless, this stuff will always keep progressing, regardless of who is doing it.

Regarding the economy, it may be largely irrelevant if we, the people, don't do something very soon. The wheel keeps spinning as long as there are productive workers; it's just that those workers are being replaced by machines. The last year has increasingly demonstrated that you don't need normal people to buy your stuff to remain afloat. You can just keep selling amongst your rich friends while the masses starve, as long as _something_ is still producing what the wealthy want, and enough systems are in place to protect them.

[ryanackley]

You’re just hand waving. No actual facts in your argument.

Sure, if you have a desktop with 175GB of unified memory, you can run deepseek 4 locally. Not completely out of reach but pretty close for almost everyone. There are smaller models but we’re talking about state of the art stuff that can reliably be used to do serious work right?

Also, even in a mad max style dystopian future, the elite need the working class to enjoy their luxuries. There is a long supply chain of experts and workers to build a yacht for example. It’s kind of ridiculous to imagine robots and AI replacing that entire supply chain. It would be extremely fragile. While conceivable, it’s also an unrealistic extrapolation grounded in sci fi instead of reality

[hodgehog11]

> No actual facts in your argument

I don't see facts in any of these arguments, that's really the point. Doomerism isn't particularly productive, but I'm tired of the complacency and the suggestions that everything will sort itself out. There is a chance it won't.

> we’re talking about state of the art stuff that can reliably be used to do serious work right

The enormous data centers are needed to train new models and deliver to hundreds of thousands of people. For us plebs, yes, the biggest models are out of reach. But a medium-sized business can readily buy the hardware needed to run the state-of-the-art locally if they have the weights. Inference is not so bad. That will be even more true in the future. So it's crazy to me to suggest that AI would just go away everywhere because it is too expensive. The problem is that the current arms race is wasteful and cares not for profitability.

> elite need the working class to enjoy their luxuries

I think this is an extremely critical misconception, and it's sending the world into an increasingly bad place. They really don't, and the assumptions that underpin this statement breaks down when the elite own all the critical assets. If you need proof of this, look at how the working class is being increasingly priced out of almost all luxuries right now. That's the norm. Almost all of human history has been that way. The formula could get a lot worse if there is even the remotest chance that robots or AI can take the place of the workers that might desperately be fighting for the scraps of the wealthy.

> It would be extremely fragile

Actually I think the current state of affairs is fragile. Could you explain this?

> grounded in sci fi instead of reality

More history than sci fi, but a fair criticism. Still, I don't believe there are any "factual" refutations of my concerns, and that should be worrying.

[ryanackley]

>I think this is an extremely critical misconception, and it's sending the world into an increasingly bad place. They really don't, and the assumptions that underpin this statement breaks down when the elite own all the critical assets. If you need proof of this, look at how the working class is being increasingly priced out of almost all luxuries right now.

Huh? What luxuries? Us plebes can't fly business class? We can't buy that expensive handbag? A better argument would be they can't afford to buy a family home in a lot of markets but this has to do with generational wealth and a housing shortage in many parts of the USA.

> More history than sci fi, but a fair criticism. Still, I don't believe there are any "factual" refutations of my concerns, and that should be worrying.

It's economics. There is a tipping point where automation is self-undermining for capitalism. If nobody has a job, demand collapses. i.e. nobody buys the mountain of goods the robots and AI are producing.

If the economy collapses, many wealthy people would no longer be wealthy. Who is maintaining the robots that are doing everything? Other robots? Now we're getting into sci-fi territory.

Even during the industrial revolution, jobs moved from the farm to the factories. There was not a total replacement for human labor like you seem to be suggesting will happen.

[hodgehog11]

> Huh? What luxuries? Us plebes can't fly business class? We can't buy that expensive handbag?

Yes. Why do you need to ask this? This is the K-shaped economy. Demand is dropping (especially in luxury handbags) as the middle class gets hollowed out. Maybe you're well-off enough that you can still pretend neoclassical economics is still holding up. Must be nice.

> they can't afford to buy a family home in a lot of markets but this has to do with generational wealth and a housing shortage in many parts of the USA.

Yes it has to do with generational wealth, that's my point. A shortage is true in some cases, but not all. That's mostly fueled from massive demand from the wealthy. Buying the family home is the most obvious asset that is becoming out of reach. This past year, many other assets have gone the same way. I think that will continue.

> nobody buys the mountain of goods the robots and AI are producing

I don't get why you're appealing to modern economic theory when the whole point of this scenario is that the standard economic relationship entirely breaks down. You can dismiss it as sci-fi, but people are thinking about this scenario. The wealthy no longer need a mass consumer market in this scenario to stay wealthy. They could simply trade proprietary algorithms, real estate, raw resources, and automated services exclusively among themselves. The global economy shrinks down to a private, self-contained club. Everyone else is locked out of the market. This should sound familiar if you're paying attention to the current markets. There would only need to be a few to maintain the status quo, and they remain inside the market in exchange.

The exit from this bleak future is societal unrest, which needs to occur sooner rather than later in order to succeed. That's the source of instability. Later on, not so much.

> Even during the industrial revolution, jobs moved from the farm to the factories.

Yes, but living standards seriously deteriorated for a long time. That's not too far from the exterminism in the above scenario, just not as dire since factories still need far more human labor to run them. If the labor itself becomes redundant, that's very bad.

I'm not saying this will happen. But dismissing it as sci-fi doesn't seem wise when we're seeing the signs of that future already.

[ryanackley]

I was being sarcastic about business class and expensive handbags. You act like we all should be entitled to these completely unnecessary things.

What you’re suggesting is a return to some kind of neo feudalism which means the elites would need to field their own robot army to challenge governments because in that type of society the power of the sword is the only real power. With robot armies that build themselves and mine uranium for their own small nuclear reactors, we’re back in sci-fi territory.

[hyperpape]

> The curve of AI improvement will continue at the current pace

I guess this is trivially true if you say "maximalism" (hell, the maximalists think it will speed up as the AI becomes a super-AI-researcher), but as long as the rate of change is positive and not miniscule, it's hard to predict what 2035 looks like in software development.

These things are very hard to quantify, but making the progress that happened from Jan 2025-December 2025 repeat twice in 10 years would be enough for me to say I couldn't predict the day-to-day of a software engineer in 2035.