> 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.
I'm no convinced. Propaganda only works when it relies on something that people are dispositionally inclined towards. eg: hating foreigners, calls towards group unity, etc.
You can't use propaganda to trick people into becoming experts at calculus, for instance.
20 years ago is about when SSDs came to consumer PCs. Which is the last time I remember thinking that my computer became faster than the previous one was when it was new (if by computer we mean the hardware combined with a mainstream up-to-date software stack)
Your computer may not feel faster than a 20 year old PC, but CPUs have gotten way, way faster.
Compare the AMD Ryzen 7 9800X3D to the AMD A12-9800, you go from 4 cores to 8, and you double the power consumption. But it's not twice as fast. It's 10 times faster. Depending on the exact metric, it could be as little as twice as fast or as much as 1000 times faster, depending on the exact operations.
On top of that memory throughput between DDR4 and DDR6 is about 2.5 times faster as well.
Oh and this isn't 20 years apart, this is 7 years apart. 20 years will see almost an exponentially larger gap even still.
I kind of hoped my point would come across more clearly. Obviously my current computer is faster than one from 20 years ago. But, it doesn't really react more quickly. It just does more: higher resolution video, advanced graphics, etc. The moment we add capability we rush to use it all up. This is not _always_ bad when it comes to computing, but I wouldn't mind if people in general had more restraint.
Obesity is nearly everywhere. In the old days, "prosperous" might have been a euphemism for "fat," as only the rich could afford to be fat. Now even people who are "food insecure" are often quite fat. Some of this of course is due to food quality, but ultimately there are no starving fat people.
How does "people will uncontrollably eat if given the chance" turn into "there are no starving fat people".
You're right that obesity is not necessarily linked to overeating . How is this relevant to your original point? Isn't it a complete contradiction to what you were saying before?
The original post was hyperbolic to the point of nonsense and now it seems we're on a completely different topic.
>You're right that obesity is not necessarily linked to overeating
This can be true, but I think some people are confused. The vast majority of the time, it is linked to overeating. There are some edge cases where there is something else going on metabolically, but most people do over eat.
Jevon's Paradox says nothing about the size of the incremental demand though. Lighting is a great example. The US went from spending 15% of our electricity on lighting to ~4%. LEDs are maybe 5x as efficient, so that would be 3% without any change in usage. So we did see maybe a 30% increase in consumption as a result of lower prices, but that was nowhere near large enough to offset the efficiency gain in terms of overall consumption. Las Vegas may have, but that is a very small part of overall usage.
I anticipate we will see something similar with intelligence. There is probably headroom to consume 100x as much intelligence in R&D. But that isn't most of the economy. Will run of the mill service jobs increase their use of intelligence by enough to offset the effect of cheaper prices? I think that's the real question.
Intelligence is a lever, not a fluid. Power depends on how much force it can apply to specific points, not how much exists in general.
True high-intelligence outputs (Maxwell's equations, the Fourier Transform, quantum theory) are fundamentally transformative in ways that mid-high-competence (starting a generic B2B SaaS, making another CRUD app) aren't.
Which is why we've assumed we're already pretty far down the path to AI, but we really aren't. Solving random Erdős problems isn't the same as opening up a completely new kind of math/science with game changing practical applications.
I don't think you can get to that level with more compute and more tokens. I think it's going to take new higher level knowledge representations and new kinds of training to get there.
And the token count and compute may turn out to be lower than what we're using now.
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.
And I pretty much just plain agree, this is exactly what will happen.
I don't think there's anything wrong with it (in isolation) either, though I do already find myself pointing out that we're misusing LLMs at work sometimes (most notably, a recent mini project could have been a jinja template - and it did become one thanks to me pushing back on this). Abundance is one thing, waste and misuse is another.
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.
Sensors are a huge challenge for robotics. We have very precise force-feedback on our joints, pressure and heat (temperature gradient) sensors all over our body, and our hands have a sensor density that allows us to count needle heads and detect the exact grip strength needed by feeling the micro-slippage of objects in our hands. Robots don't have that.
You can do backflips with pretty much just visual sensors for your environment, a good IMU for your spatial orientation, and some feedback on the position of a small number of really beefy joints and the force exerted on them. Folding laundry and opening doors is much more difficult, and trying to compensate with mostly vision requires going slow enough that things have time to move over appreciable distances before you take the next adjustment
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.
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.
I think what a lot of people miss about jevon's paradox is the elasticity of demand of the underlying resource
textiles had jevons paradox, and many more textile workers were employed even when textile machines were being created, until we saturated the demand for cheap clothing in the world and then textile workers were kaput (same for farming, and horses)
software is currently undergoing jevons paradox, but it's very unknown how high the ceiling of demand for software is. web dev might be doomed, but software in general i think is probably limitless
Intelligence is also probably unbounded (atm software and intelligence are very closely tied together). its very possible token spend rides up the curve forever.
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.)
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'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 :)
The big question is if there are larger economic gains to be made from ever greater intelligence. It seems to me like that may be the case for only a select few hard problems, while the vast majority of tasks approach their economic ceiling asymptotically with intelligence.
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).
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.
Sooooo what does it matter if we do 99% of things a LLM can just solve as a 'advanced search engine'?
Btw. an advanced search engine is probably the worst comparision i have read so far.
A LLM is a latent space which is capable of a lot of things a search engine can't do. It can apply different type of patterns and flows onto data, it can combine these etc.
My 'advanced search engine' was just able to create a working PR for exactly what i wanted it to solve (fixing a bug) by analysing the bug, finding a valid solution then commiting the solution itself.
But you forget all development today is just searching for a template, copy pasting, changing some small things. And a smart-ish search engine can do all that.
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.
It's not about magic, it's that humans have an evolved embodied intelligence for surviving in the world as living organisms that reproduce, care for young and live in communities. That's quite a bit different from language models trained on human data.
Artificial intelligence is artificial. It can still be called intelligent, just not the same kind as biological since it's not remotely biological. It's human-like but also alien. To have something artificial be human you'd need something like replicants from Blade Runner which are synthetic biological robots.
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 for them. You'll either be strung along on an infinite goalpost moving exercise, or 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 for it to stay that way.
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 [1], still inherently human for a lot of people.
To be clear, I'm not trying to be judgemental with this, I more consider it to be a communications breakdown, and find that to be frustrating instead. I'm not really sure how to meaningfully help it either, cause no matter how one slices it, you will in the end ask these people do desecrate these terminologies in favor of some more twisted-seeming ones. Same the other way around, the humanist understanding of these terms is basically non-workable.
> 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.
It’s way too early to tell the true cost of intelligence. Inference is still heavily subsidized, and training is apparently being funded by a mountain of free money
The article is solid. But there is a nuance what he skipped — quality vs price. Sure, GPT-5.6 Luna for pennies can do the same thing what Claude 4.5 Sonnet did for a dollar a year ago. Except Sonnet back then actually carried the codebase, while Luna... eh, not so much.
And another thing, speed. You can make it cheaper as much as you want, but if a model thinks for half a minute you save cents but lose time.
The other day I saw a benchmark of coding models that fit in 8GB VRAM - for reference some version of Mistral was added, normally requiring 32GB, but moving along at 4-5tok/sec when partially offloaded to CPU.
Surprisingly, some of the small models would not only give worse results, but also took longer than Mistral, because they were thinking so much.
That is an important detail which I was previously overlooking.
Even if its not intelligence, a LLM found a bug due to one error message, fixed it, created a PR and it solved it.
If an LLM is only able to do all of this after training on it and never achieving AGI, we already at the point were it is cheaper to teach one LLM one problem than teaching humans to do so.
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.
It blows my mind how fast models are getting better and this is the first article I’ve seen that shows just that and leaves almost no room for disagreement. Well done.
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?
Jevons Paradox [1]
> 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.
> 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.
these things you are talking about are only true due to the unbelievably LARGE marketing ~propaganda~ spending
what you're saying reflects corporation tendencies
take away the propaganda, let people live, see how they react
I'm no convinced. Propaganda only works when it relies on something that people are dispositionally inclined towards. eg: hating foreigners, calls towards group unity, etc.
You can't use propaganda to trick people into becoming experts at calculus, for instance.
You're right, computers are definitely the same speed as they were 20 years ago.
20 years ago is about when SSDs came to consumer PCs. Which is the last time I remember thinking that my computer became faster than the previous one was when it was new (if by computer we mean the hardware combined with a mainstream up-to-date software stack)
You had SSDs in 2006? Was I missing out?
I don't think I had any in my builds until 2014ish
Your computer may not feel faster than a 20 year old PC, but CPUs have gotten way, way faster.
Compare the AMD Ryzen 7 9800X3D to the AMD A12-9800, you go from 4 cores to 8, and you double the power consumption. But it's not twice as fast. It's 10 times faster. Depending on the exact metric, it could be as little as twice as fast or as much as 1000 times faster, depending on the exact operations.
On top of that memory throughput between DDR4 and DDR6 is about 2.5 times faster as well.
Oh and this isn't 20 years apart, this is 7 years apart. 20 years will see almost an exponentially larger gap even still.
I kind of hoped my point would come across more clearly. Obviously my current computer is faster than one from 20 years ago. But, it doesn't really react more quickly. It just does more: higher resolution video, advanced graphics, etc. The moment we add capability we rush to use it all up. This is not _always_ bad when it comes to computing, but I wouldn't mind if people in general had more restraint.
many computers actually feel much slower than they did 20 years ago: https://danluu.com/input-lag/
(...or over 40 years ago)
Anyone not living paycheck to paycheck is just constantly gorging on fast food?
I don't mean to be glib, but for the most part, yes. https://www.nytimes.com/2025/09/11/well/child-obesity-un-rep...
Obesity is nearly everywhere. In the old days, "prosperous" might have been a euphemism for "fat," as only the rich could afford to be fat. Now even people who are "food insecure" are often quite fat. Some of this of course is due to food quality, but ultimately there are no starving fat people.
How does "people will uncontrollably eat if given the chance" turn into "there are no starving fat people".
You're right that obesity is not necessarily linked to overeating . How is this relevant to your original point? Isn't it a complete contradiction to what you were saying before?
The original post was hyperbolic to the point of nonsense and now it seems we're on a completely different topic.
>You're right that obesity is not necessarily linked to overeating
This can be true, but I think some people are confused. The vast majority of the time, it is linked to overeating. There are some edge cases where there is something else going on metabolically, but most people do over eat.
Jevon's Paradox says nothing about the size of the incremental demand though. Lighting is a great example. The US went from spending 15% of our electricity on lighting to ~4%. LEDs are maybe 5x as efficient, so that would be 3% without any change in usage. So we did see maybe a 30% increase in consumption as a result of lower prices, but that was nowhere near large enough to offset the efficiency gain in terms of overall consumption. Las Vegas may have, but that is a very small part of overall usage.
I anticipate we will see something similar with intelligence. There is probably headroom to consume 100x as much intelligence in R&D. But that isn't most of the economy. Will run of the mill service jobs increase their use of intelligence by enough to offset the effect of cheaper prices? I think that's the real question.
Intelligence is a lever, not a fluid. Power depends on how much force it can apply to specific points, not how much exists in general.
True high-intelligence outputs (Maxwell's equations, the Fourier Transform, quantum theory) are fundamentally transformative in ways that mid-high-competence (starting a generic B2B SaaS, making another CRUD app) aren't.
Which is why we've assumed we're already pretty far down the path to AI, but we really aren't. Solving random Erdős problems isn't the same as opening up a completely new kind of math/science with game changing practical applications.
I don't think you can get to that level with more compute and more tokens. I think it's going to take new higher level knowledge representations and new kinds of training to get there.
And the token count and compute may turn out to be lower than what we're using now.
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.
> I think the same will happen with tokens. As the cost of tokens comes down, these models will just consume more tokens.
That sounds a race to the bottom for AI companies profits.
I do not think that LEDs are high profit margin items.
Tokens for sure because we still waiting for agents talking to agents as a default. Like your agent team and very long or 24/7 running agents.
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?
Turns out Grey Goo was just thermal paste
Sounds like a variation on the induced demand principle: https://en.wikipedia.org/wiki/Induced_demand
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 don't think there's anything wrong with it (in isolation) either, though I do already find myself pointing out that we're misusing LLMs at work sometimes (most notably, a recent mini project could have been a jinja template - and it did become one thanks to me pushing back on this). Abundance is one thing, waste and misuse is another.
Also Jevon's Paradox: the more efficient a system becomes, the more it gets used, making it use more resources rather than less.
> Reading everything becomes the default. At a cent per document, a model can read every paper
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.
Hasn't the computer already been spending time with your wife?
Asymmetric multiprocessing with your mom
it's called Instagram
You're absolutely right!
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.
The next generation of bike computers will probably have links to an LLM cloud subscription service for real-time coaching and route planning.
Futurama did it first!
"Robot, experience this tragic irony for me!"
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.
One of my projects has an estimate of 3,000+ hours. It seems accurate. It has spent months working on it.
Can I ask what you are working on?
Around the clock?
Yes, with a few exceptions. It's much easier now that the five hour limit has been removed from codex.
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.
Sensors are a huge challenge for robotics. We have very precise force-feedback on our joints, pressure and heat (temperature gradient) sensors all over our body, and our hands have a sensor density that allows us to count needle heads and detect the exact grip strength needed by feeling the micro-slippage of objects in our hands. Robots don't have that.
You can do backflips with pretty much just visual sensors for your environment, a good IMU for your spatial orientation, and some feedback on the position of a small number of really beefy joints and the force exerted on them. Folding laundry and opening doors is much more difficult, and trying to compensate with mostly vision requires going slow enough that things have time to move over appreciable distances before you take the next adjustment
Xiaomi robots do this in double digit seconds.
Still slow compared to humans, but Chinese robots will be as successful as Chinese EVs, phones and solar panels.
~1min per shirt in a controlled environment. Better but still far away.
Which is to say, they'll be banned in the US.
You must not have been paying attention to development with robots, there are many videos of robots moving really fast in "non controlled environments"
Most of these videos are still controlled in some manner. We are not yet at the age of throwing a robot in a home and doing n different dynamic tasks.
Yes, I think I saw one in a video titled "Robocop"
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?
Wait til the robots get on Cerebras, it'll set your pants on fire.
Only if the fire manages to escape your wallet!
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.
Transformers in general are huge but if you say LLM you’re specifically saying the language model transformer. Robots use vision transformers
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.
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.
I think what a lot of people miss about jevon's paradox is the elasticity of demand of the underlying resource
textiles had jevons paradox, and many more textile workers were employed even when textile machines were being created, until we saturated the demand for cheap clothing in the world and then textile workers were kaput (same for farming, and horses)
software is currently undergoing jevons paradox, but it's very unknown how high the ceiling of demand for software is. web dev might be doomed, but software in general i think is probably limitless
Intelligence is also probably unbounded (atm software and intelligence are very closely tied together). its very possible token spend rides up the curve forever.
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.)
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 saw a 1B model yesterday that was fine tuned on Fable output. I found that hilarious, but it did actually make all the scores go up.
(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?)
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 :)
The big question is if there are larger economic gains to be made from ever greater intelligence. It seems to me like that may be the case for only a select few hard problems, while the vast majority of tasks approach their economic ceiling asymptotically with intelligence.
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).
Diminishing returns on both intelligence and training, mostly
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.
Well that's just my 2 cents.
Sooooo what does it matter if we do 99% of things a LLM can just solve as a 'advanced search engine'?
Btw. an advanced search engine is probably the worst comparision i have read so far.
A LLM is a latent space which is capable of a lot of things a search engine can't do. It can apply different type of patterns and flows onto data, it can combine these etc.
My 'advanced search engine' was just able to create a working PR for exactly what i wanted it to solve (fixing a bug) by analysing the bug, finding a valid solution then commiting the solution itself.
You're correct on the technology assessment.
But you forget all development today is just searching for a template, copy pasting, changing some small things. And a smart-ish search engine can do all that.
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.
It's not about magic, it's that humans have an evolved embodied intelligence for surviving in the world as living organisms that reproduce, care for young and live in communities. That's quite a bit different from language models trained on human data.
Artificial intelligence is artificial. It can still be called intelligent, just not the same kind as biological since it's not remotely biological. It's human-like but also alien. To have something artificial be human you'd need something like replicants from Blade Runner which are synthetic biological robots.
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 for them. You'll either be strung along on an infinite goalpost moving exercise, or 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 for it to stay that way.
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 [1], still inherently human for a lot of people.
To be clear, I'm not trying to be judgemental with this, I more consider it to be a communications breakdown, and find that to be frustrating instead. I'm not really sure how to meaningfully help it either, cause no matter how one slices it, you will in the end ask these people do desecrate these terminologies in favor of some more twisted-seeming ones. Same the other way around, the humanist understanding of these terms is basically non-workable.
[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
> 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
Pretty much already happening
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.
It’s way too early to tell the true cost of intelligence. Inference is still heavily subsidized, and training is apparently being funded by a mountain of free money
How do we know inference is heavily subsidized? Are all these third party Chinese model providers subsidizing the true cost?
The article is solid. But there is a nuance what he skipped — quality vs price. Sure, GPT-5.6 Luna for pennies can do the same thing what Claude 4.5 Sonnet did for a dollar a year ago. Except Sonnet back then actually carried the codebase, while Luna... eh, not so much. And another thing, speed. You can make it cheaper as much as you want, but if a model thinks for half a minute you save cents but lose time.
The other day I saw a benchmark of coding models that fit in 8GB VRAM - for reference some version of Mistral was added, normally requiring 32GB, but moving along at 4-5tok/sec when partially offloaded to CPU.
Surprisingly, some of the small models would not only give worse results, but also took longer than Mistral, because they were thinking so much.
That is an important detail which I was previously overlooking.
this guy gets it
Idiocy becomes rampant!
Whats your problem?
Even if its not intelligence, a LLM found a bug due to one error message, fixed it, created a PR and it solved it.
If an LLM is only able to do all of this after training on it and never achieving AGI, we already at the point were it is cheaper to teach one LLM one problem than teaching humans to do so.
AGI ≈ Ralph × Infinite persistence
Just like real life!
becomes rampant? Where have you been the past decades? Hell, centuries and millenia?
If we're going cynical, may as well go full throttle.
all the optimistic pundits ignore the blatant second order consequence of this: the entire economy is pegged on this NOT being cheap!
all the US economy is tied to video cards being used in lieu of gold. cost dropping 100x means the economy bottom falls out.
nothing because this shit is not "intelligence"
ppl are doing all sorts of gymnastics to tell claude to slow its roll with verbosity.
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.
It blows my mind how fast models are getting better and this is the first article I’ve seen that shows just that and leaves almost no room for disagreement. Well done.
Thanks for the effort you put into this.
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