> when technological improvements that increase the efficiency of a resource's use lead to a rise, rather than a fall, in total consumption of that resource.
Las Vegas replaced the expensive incandescent lighting on the strip with cheaper to run LED equivalents. But the costs didn't come down because they were able to add more lights and larger displays.
I think the same will happen with tokens. As the cost of tokens comes down, these models will just consume more tokens.
> But the costs didn't come down because they were able to add more lights and larger displays.
People are a gas; they expand the fill the space they're in. If you give someone a big house, they'll fill it with crap. If you make food cheap, they'll eat too much, even when it harms their health.
Making things cheaper usually just makes making them more prevalent. It's why computers are not faster than 20 years ago. They're merely more capable -- developers quickly and aggressively fill up (and overflow) all that added capability until you're back in the same place you used to be.
Definitely not going to argue against the Jevons paradox in general, it is observed in various cases.
The lightbulb thing seems different though? Or at least it is a specific subset. Lights in Las Vegas are sort of an advertisement, right? In the sense that having the brightest or most interesting (or whatever) lights draw attention to your show, casino, hotel, whatever. It’s kind of a zero sum game in that the different shops are competing for the finite attention of a more-or-less set number of tourist. I think part of the Jevons paradox is that society generally finds more useful applications of the newly cheap thing. If the thing’s only purpose is to compete better in a competition with a set prize (all of the tourists’ money), that’s constrained in some way.
Intelligence is weird though. I guess we could eventually hit the point where, I dunno, maybe there’s some information theoretic bound where we process all of our signals as cleverly as possible and aren’t bound by intelligence anymore. Obviously we’re nowhere near that. It would be a very alien environment.
Douglas Adams captured this in "Dirk Gently's Holistic Detective Agency", a VCR watches TV for you and an Electric Monk believes things for you to save you the effort of doing it yourself.
I think speed is actually going to be a bigger factor than cost. Even projects where “money is no object” often hit a wall with LLM response times.
Sure you can speed things up with parallel work under subagents, but as with parallelizing traditional computational tasks, there are diminishing gains.
I keep hearing people saying just change the way you work to trust long-running agents and multi-task more, because they’re too slow to work with interactively for many use cases. I think that’s painful in a world where we expect humans to still heavily guide and interact with agents for their day-to-day work.
> they’re too slow to work with interactively for many use cases
This just demonstrates how much we already take for granted the LLMs that we have now. If you compare it to what we had before (hand the task off to a junior dev and wait for them to complete the work) then it doesn't seem slow at all.
Given the 750 tok/sec GPT 5.6 Sol Ultrafast (via Cerebras), the many-1000 tok/sec Chinese models, and the 15000 tok/sec Taalas HC1, I think we're well on the way towards seeing that solved too. Combine the two, and yeah, wild ride incoming.
What's especially bewildering to me is that translated back to raw bandwidth, even 15000 tok/sec is just like what, 75 KB/s? Extremely meager amounts of data, moving mountains.
It's already kinda funny seeing LLMs throw out effort estimates in wall time terms. It's always some "hours, days, weeks" tier thing, when in reality, it's gone and done in minutes.
I think a big one is robotics. A robot can today fold your laundry. It takes ~10mins per item. Seriously. It takes a long time to process the image find the corner move the claw to the corner of the shirt and attempt to straighten before folding.
Robots right now generally move at glacial speeds. You might have seen robots doing flips in semi controlled environments but watch how slowly they open doors etc. processing time is a major bottleneck.
You must not have been paying attention to development with robots, there are many videos of robots moving really fast in "non controlled environments"
Honestly, most of the videos I've seen of robots moving around quickly aren't actually doing anything useful. We've had really impressive tech demos for the past 15 years of robots dancing and jumping around. But I don't want a dancing robot, I want a robot to make me a BLT, wash my dishes, take out the trash, and fold my laundry.
The most recent video which actually impressed me was a demonstration from Gemini Robotics 2, where a robot was shown autonomously removing the bag from a trash can and folding the loops closed in real time.
I don't follow robotics advances closely so it's possible I'm just ignorant, do you know any autonomous robotics demonstrations of useful activities that you would suggest checking out?
Have LLMs improved at being able to process physics-based problems and environments? I remember that issue being discussed around generative gaming a while ago but I hadn't heard much about it recently.
100x seems like an underestimate. Even with no model improvements, we should see that sort of reduction. Looking at TSMC’s margins, Nvidia’s margins, and OAI/Anth (alleged) margins on inference, there is a room for a 100x reduction.
Right now all three of those are at abnormally high levels. Competition will come for all three.
There’s still 50-500x cost reduction in “this is only an engineering problem” low hanging fruit from specialized chips to run the models + improved distillation.
Entirely feasible that by 2031, Fable 5 (or greater) intelligence level models will run cool on smart phones, if not sooner.
You're betting on getting getting ridiculously powerful chips to run on batteries in a tiny housing without cooling, while we can't even get enough RAM? It would be a terrible waste of resources. Now we already have TFLOPs wasting in our pockets and backpacks, then we'll have PFLOPs idling, because there's so much time between prompts. Much more efficient to batch it on a server.
At some point we'll have enough RAM, surely. The incentives to produce more are huge and all the fabs are booked out.
Maybe it'll take 10 years or 20 years. <5 years is not long enough for manufacturing to catch up.
Not much of a comment on the phone stuff but I'd caution against suggesting technology will never be good enough to do X. Maybe it'll be horrendously wasteful but it might happen.
I’m loving small models. The gemma4 26/31b models have been deeply impressive on weird prose analysis tasks that I am working on. Nova-micro is really stupid but is extremely fast when it’s smart enough to do something. I’m trying to be disciplined about able to evaluate quality vs cost everywhere for real systems built on this stuff. I probably need to get off Bedrock because it’s missing a lot of other little models that might be good competitors.
A year ago I had an aha moment, when I realized that for my purposes, Gemini Flash was not only 9x cheaper, but 3x faster than Gemini Pro, while producing identical output. Who's the best model now!
For a lot of tasks, even small models have saturated them a while ago, and then going cheaper and faster is just pure gains.
For coding I also prefer to do it interactive/realtime, micro-prompting, surgical edits, which the small models can handle just fine.
And then at the top, the real question is consistency. Not "can they do it" but "reliably enough that you don't need to constantly double check everything." (In my experience, not quite there yet, although it's getting way better.)
I've been working with the chinese open models for 4 months. They are more than capable for a tiny fraction of the cost of the frontier ones. And yet they also continue to get significantly better and (Deepseek's recent price increase aside) cheaper. Its hard to fathom how the truly frontier stuff will be able to compete long-term.
And why won’t the frontier models continue to become better? The open models are getting better but so are the frontier models. The frontier models might remain in a constant race to remain ahead
I don't think it's a matter of "staying ahead"; the proprietary frontier models are better, but the trouble for them is that open weight models are good enough in increasingly many cases. This raises the floor on the frontier companies and cuts their total addressable market by commodifying the easier LLM tasks. This is really the central argument of tfa :)
They are running out of novel, clean training data and compute. There is probably a limit to how much improvement can be squeezed out of LLMs. Recent improvements have been more about orchestration and "reasoning" loops (i.e. iteratively feeding context back through the model).
> Reading everything becomes the default. At a cent per document, a model can read every paper in a field, every record in an archive, every email, or every message in a support queue as a matter of routine
This means a great de-risking is happening for the costs of deploying somewhat autonomous agents. This has profound implications on the timeline of deployable personal agents. Cost was a significant factor for many people during the OpenClaw frenzy, specifically when they let their agents run somewhat wild.
It will become much more palatable, or already has, to install whatever the next generation of token consuming autonomous systems will be.
Personally I still see LLMs as very advanced search engines which lack intelligence. To me it seems that the cost of getting data is reduced by LLMs, not the cost of intelligence. I mean: we tell the model what we want to achieve, and the model responds with the right data in de form of code in seconds.
That's why 'stackoverflow programmers' will have a hard time competing with LLMs but engineers are still needed for their intelligence.
I think this viewpoint fails to understand what "intelligence" is. The idea must be that intelligence is some special thing that only humans have. So when machines couldn't do jack s... we said "it's the Turing test". When machines blew through the Turing test we said "that was just prediction..not really intelligence, that's different".
It's not different. The delusion humans have is that intelligence is special and magical. It's not. It's just nature's prediction machine. A very fancy version to be sure. But not qualitatively different .
All statements that "oh but it'll never be able to do that" will prove false.
I notice that a lot of these terms are squarely humanist for certain people, so any kind of allegation that machines are exhibiting them as traits will be a complete showstopper. You'll be accused of either anthropomorphization [0] in the kinder cases, or straight up mental illness in the less so kind cases. Never will they stop to consider that maybe you're simply working with a post-humanist understanding of these words, as that basically doesn't make sense to them, and as they are usually quite vested to not even attempt to.
I remember in one of the Hugging Face incident threads here, simply acknowledging that the agents were operating autonomously was super controversial. Thousands of years old concept, still inherently human for some people. [1]
Author here. I made these plots because I had been searching for them for months and never found quite what I wanted: how the cheapest way to reach a fixed capability level has moved over time. Artificial Analysis publishes enough data to reconstruct it. If someone knows of a source that already tracks this, with historical prices, please share.
I wonder if it might drive the point even further if the graph scales were linear? Or maybe the progress has been so great that this would make the graphs unreadable?
> when technological improvements that increase the efficiency of a resource's use lead to a rise, rather than a fall, in total consumption of that resource.
[1] - https://en.wikipedia.org/wiki/Jevons_paradox
Las Vegas replaced the expensive incandescent lighting on the strip with cheaper to run LED equivalents. But the costs didn't come down because they were able to add more lights and larger displays.
I think the same will happen with tokens. As the cost of tokens comes down, these models will just consume more tokens.
People are a gas; they expand the fill the space they're in. If you give someone a big house, they'll fill it with crap. If you make food cheap, they'll eat too much, even when it harms their health.
Making things cheaper usually just makes making them more prevalent. It's why computers are not faster than 20 years ago. They're merely more capable -- developers quickly and aggressively fill up (and overflow) all that added capability until you're back in the same place you used to be.
The lightbulb thing seems different though? Or at least it is a specific subset. Lights in Las Vegas are sort of an advertisement, right? In the sense that having the brightest or most interesting (or whatever) lights draw attention to your show, casino, hotel, whatever. It’s kind of a zero sum game in that the different shops are competing for the finite attention of a more-or-less set number of tourist. I think part of the Jevons paradox is that society generally finds more useful applications of the newly cheap thing. If the thing’s only purpose is to compete better in a competition with a set prize (all of the tourists’ money), that’s constrained in some way.
Intelligence is weird though. I guess we could eventually hit the point where, I dunno, maybe there’s some information theoretic bound where we process all of our signals as cleverly as possible and aren’t bound by intelligence anymore. Obviously we’re nowhere near that. It would be a very alien environment.
That sounds a race to the bottom for AI companies profits.
I do not think that LEDs are high profit margin items.
But we never scaled intelligence like this. The industrieal revolution created for the people at that time quite a huge issue / it was disruptive.
What will hapen to us though?
Related:
- Parkinson's law: "Work expands to fill the available time." https://en.wikipedia.org/w/index.php?title=Parkinson%27s_Law
- Lewis–Mogridge position: "Traffic expands to meet the available road space." https://en.wikipedia.org/wiki/Lewis%E2%80%93Mogridge_positio...
And I pretty much just plain agree, this is exactly what will happen.
I love how in our day "reading everything" means "the computer reads it for me".
I expect soon the computer will be able to go on bicycle rides, and spend time with my wife.
Sure you can speed things up with parallel work under subagents, but as with parallelizing traditional computational tasks, there are diminishing gains.
I keep hearing people saying just change the way you work to trust long-running agents and multi-task more, because they’re too slow to work with interactively for many use cases. I think that’s painful in a world where we expect humans to still heavily guide and interact with agents for their day-to-day work.
This just demonstrates how much we already take for granted the LLMs that we have now. If you compare it to what we had before (hand the task off to a junior dev and wait for them to complete the work) then it doesn't seem slow at all.
What's especially bewildering to me is that translated back to raw bandwidth, even 15000 tok/sec is just like what, 75 KB/s? Extremely meager amounts of data, moving mountains.
It's already kinda funny seeing LLMs throw out effort estimates in wall time terms. It's always some "hours, days, weeks" tier thing, when in reality, it's gone and done in minutes.
Robots right now generally move at glacial speeds. You might have seen robots doing flips in semi controlled environments but watch how slowly they open doors etc. processing time is a major bottleneck.
Still slow compared to humans, but Chinese robots will be as successful as Chinese EVs, phones and solar panels.
The most recent video which actually impressed me was a demonstration from Gemini Robotics 2, where a robot was shown autonomously removing the bag from a trash can and folding the loops closed in real time.
I don't follow robotics advances closely so it's possible I'm just ignorant, do you know any autonomous robotics demonstrations of useful activities that you would suggest checking out?
Right now all three of those are at abnormally high levels. Competition will come for all three.
Entirely feasible that by 2031, Fable 5 (or greater) intelligence level models will run cool on smart phones, if not sooner.
Maybe it'll take 10 years or 20 years. <5 years is not long enough for manufacturing to catch up.
Not much of a comment on the phone stuff but I'd caution against suggesting technology will never be good enough to do X. Maybe it'll be horrendously wasteful but it might happen.
(Actually talking to it, it was about as coherent as you'd expect, i.e. 3/10)
The floor for "actually usable model" keeps dropping though. (Seems to be about 27B right now?)
For a lot of tasks, even small models have saturated them a while ago, and then going cheaper and faster is just pure gains.
For coding I also prefer to do it interactive/realtime, micro-prompting, surgical edits, which the small models can handle just fine.
And then at the top, the real question is consistency. Not "can they do it" but "reliably enough that you don't need to constantly double check everything." (In my experience, not quite there yet, although it's getting way better.)
Pretty much already happening
That's why 'stackoverflow programmers' will have a hard time competing with LLMs but engineers are still needed for their intelligence.
Well that's just my 2 cents.
It's not different. The delusion humans have is that intelligence is special and magical. It's not. It's just nature's prediction machine. A very fancy version to be sure. But not qualitatively different .
All statements that "oh but it'll never be able to do that" will prove false.
I remember in one of the Hugging Face incident threads here, simply acknowledging that the agents were operating autonomously was super controversial. Thousands of years old concept, still inherently human for some people. [1]
[0] as opposed to personification, which is what people are actually doing almost always: https://en.wikipedia.org/wiki/Personification
[1] https://en.wikipedia.org/wiki/Automaton
I wonder if it might drive the point even further if the graph scales were linear? Or maybe the progress has been so great that this would make the graphs unreadable?
https://xkcd.com/1162
Just like real life!
If we're going cynical, may as well go full throttle.