How would we do that? There's no historic precedent for that. Its fundamentally an information theory thing: what's the max amount of intelligence you can get out of 1 KB/MB/GB? There has to be a limit and I'm not convinced that it's far off.
Once I start a conversation about Postgres query plans, maybe 95% of the knowledge will be almost certainly not needed, and for an inference provider there will be many more concurrent queries with reasonably large overlap. Maybe future architectures will be able to take advantage of this so not all parameters are needed in memory for almost all instances.
If we are talking about all knowledge, then I agree that the compression ratio is very impressive already.
Mixture of Experts is already used by pretty much all modern LLMs to address exactly this phenomenon.
Hopefully, future models can be trained to be even more aware of external knowledge, accessible through web search / RAG / whatever it will be then, and might not need to internalize much knowledge at all.
All the terms are squishy, but being sloppy about it, I think there's intelligence that needs to be in the model, and knowledge that could live in a database. Right now, models are memorizing a lot of stuff they don't need to. We know how to index lots of information on (comparatively slow) SSDs or across a network.
See Karpathy's "Cognitive Core" idea. (I don't have a good link)
So far, frontier capability keeps getting shrunk to fit consumer level hardware. The lag time being over a year, though, precludes this from being called SOTA by the time it arrives.
There must be a limit, I agree, but there have been no signs of approaching it yet. The most recent cohort of small models have shown the biggest leap in capability so far.
Historical precedent: Quickselect, or any other discovery of a surprising algorithmic speed-up.
LLMs haven’t been on the scene for very long. There’s still lots of room for efficiency discoveries. Plus, I keep hearing quantum is going to be a big deal in the next few years.
Super interesting, when Quantum computing would enable really large models, or much larger contexts. But QRAM is even more behind than pure fault-tolerant gates.
2K, 4K, and 8K+ TVs have been around forever, and although 4K has become the norm, it's widely accepted that there isn't much benefit for most people above 1080p
Depends on distance and size, there is a lovely chart. My projector at 10 feet away is night and day difference at 4k. Mathematically 8k should look identical, so I haven’t spent the money to try it to confirm.
The 1080p thing is true in the age of 40” TVs but not so much now, think Costco has a 95” TV if not 100”.
It depends on TV size. If you've ever experienced Netflix on a 75" TV, 1080p content looks mediocre on it. So really it's better expressed as "there isn't much benefit for most TV sizes."
I wont quite argue, that 1080p is always enough on a 75", but your example is more a Netflix problem; their compression results in hideous quality, and it's not really representative of what 1080p could be.
FWIW I still play PC games on 1080p (also has the nice side effect that I don't need to buy an overpriced highend GPU, nor use upscaling hacks like DLSS which trash image quality).
Your example also counters your point: A lot of people do not care about 4k, and even less do care about 8k or above. We reached a point where most people are completely happy with the quality they get.
> Your example also counters your point: A lot of people do not care about 4k, and even less do care about 8k or above. We reached a point where most people are completely happy with the quality they get.
And when (not if) we get to the 4k (or 8k) equivalent of LLMs, they'll just be baked into hardware and we'll have near instant responses while running locally.
There is no future where OpenAI, Anthropic, etc survive with their current business model; at some point we will hit a point where training a new model is done only every 5 years or so, and in that scenario we aren't going to be running models of pricey server GPUs with latencies measured in seconds and full responses measured in minutes.
We'll be running locally with sub-millisecond latencies and responses measured in milliseconds. There is no way for any big company to compete with the current business plan of selling inference or subscriptions.
Don't forget that it's only an assumption that scaling more results in better models. There may be a ceiling to that. If that is hit and the best performing possible model can run in little VRAM, your argument here doesn't hold anymore.
In a way it is actually the same thing that makes us accept AI as working in the first place. It only needs to be good enough for human perception. The same is probably true for compute.
I'd argue we've already hit the ceiling. Can you truly tell the different between SOTA models from 9 months ago and those from today? There are some improvements but they're mostly marginal.
Plus there is a chance the actual scaling that matters is beyond our reach. Think instead of TB models, PB or ZB models. We don't even have that kind of information. Humanity in its entire history hasn't generated 1ZB of information.
> Standard Def TV was plenty for 50 years. But when more was on offer, everyone went for it, and now you can't even buy a 480p TV.
And yet when we hit 4k, that's were people just stopped buying higher res. 8K is still useful, but only when the screen is so large that it doesn't fit in the room :-/
> How would we do that? There's no historic precedent for that. Its fundamentally an information theory thing
Maybe? Human science history is absolutely littered with examples of things that were “constrained” by a fundamental law … until they weren’t, because we’d misunderstood or misapplied the law.
I think this is the future - at least it will be for on-device models. Apple, for instance, will "bake silicon" once a year for their current model, and use that chip in all their devices.
> How would we do that? There's no historic precedent for that. Its fundamentally an information theory thing: what's the max amount of intelligence you can get out of 1 KB/MB/GB? There has to be a limit and I'm not convinced that it's far off.
Sure, it might be very near using the current approach, but... it might also be might be very far off because we are using the wrong approach.
I mean, look at the max amount of intelligence you can get out of a human brain powered by two bananas...
Though even if there are big further breakthroughs to be made here, they won’t necessarily be made soon, and if they aren’t made in the next ~10 years then they won’t really affect the outcome of chip investments made now or in the immediate future.
I think the chipmakers are pretty securely set up here, the usual cyclical issues aside. Between the stubborn resistance of the data to being compressed without significant loss, and to a lesser extent the ability to do more with more memory (we’re surely at diminishing returns for LLMs already, but diminishing doesn’t mean zero) it seems like a fairly safe guess (I am not an expert on anything) that for the next decade at least you’ll still want at least 128-256GiB of video RAM to have a comfortably “SOTA-enough” experience. Then the other side of the rectangle comes from putting that VRAM in every consumer and office-worker device. Selling everything at a big premium to a few efficiency-obsessed SaaS providers is what you do when you’re temporarily supply-constrained. Sidelining the SaaS middlemen and selling to gloriously inefficient consumers who will leave their laptops closed for most of the day is the dream. This is probably not unrelated to the fact that the country which is currently making a big strategic push into memory manufacturing is also championing open models.
The fundamental issue is that for these systems bigger is essentially always better (if affordable). So if we can squash something like Kimi K3 down to run on a 'normal' system, that just incentivizes devs to increase the model size until once again we're at the limit of what can be run.
SoTA AI will move well past 1TB+ memory requirements in 1 year or so.
Also, you will be able to play with much more competent models locally. They will still feel like children compared to the adults living in the SoTA region.
Are you thinking of fundamental architectural changes (like the Transformer)? Or incremental (like MoE or GQA)? Are there specific neolabs or techs you are following that lead you to this prediction?
Indeed it seems quite possible we are one architecture breakthrough away from existing chip stock driving us all the way to ASI.
Nobody has done AFAIK the information theory to prove it’s not possible.
Yeah, you can directly print the SOTA AI model on the chip. I also believe this can be done on older process nodes. The most important factor would be speed. How fast can you go from new model to new chip?
Hyperscalers are screwed, data center mania couldn’t even be completed during this massive spending spree, all while the people seemingly sitting on the sidelines are working on getting these things baked into the OS and chipsets in a way consumers wouldn’t notice
EU chip industry: best I can do is supply chips that are in 99% of industrial equipment and vehicles. Also supply the only machines in the world that enable leading edge lithography processes.
No comment on the second part, but the first part is wrong. EU doesn't have a moat in process technology anymore (since 2010s). They have a bunch of legacy fabs, like 350, 180 and 130 nm but Taiwan, US, and China (and others like Israel, or Singapore) have these too. EU is severely lacking behind chip design, production and packaging capabilities for industrial, automotive, and space kind of niche applications since the heavy divestment from semiconductors after 2008 financial crisis and 2015 Euro crisis.
On top of this, once European giants like NXP (formerly Philips semiconductors), Infineon (formerly Siemens semiconductors), and ST are actively reducing their footprint in Europe and moving to India or SE Asia.
In short: No, EU has no future in those markets as designer or producer of chips.
This is actually at the core of our (EU) problem. Instead of thinking, "here is this new technology, how can we build new markets/products out of it", our industry leaders always think, "here is this new technology, how can we apply it to our existing industries".
You can see it a lot in AI, right now. AI4industry is not where the gains are right now. We should be competing on training foundational models. Once the foundational models from US/China are strong enough, they will do AI4industry and we wouldn't have any edge on them anyway because their tech is more vertically integrated and they have better insight in how AI works under the hood.
There's no moat to AI models so it doesn't really matter who develops it first. If AGI happens none of it matters since it won't be controlled anyway. If there's an exponential wall then everyone will be at it within 12 months of the first person hitting it and then it's better not to have spent the cash pioneering the path.
The long-term idea that the we have to compete claim is based on is like AGI and it's gonna really care what Elon tells it to do. This is a dumb idea regardless of how it fails.
AI models are seeming the ultimate in fungible commodities at this point, why spend on R&D?
EU: Perhaps we should also spend a few billion on attracting american glue manufacturers so they open a factory and a half here so we can catch up in this chip making business.
CXMT uses old Qimonda patents (even though Qimonda's tech is now outdated), and the former CEO of Qimonda is currently building a fab for an FeRAM-based alternative to DRAM in Magdeburg. It's a gamble as FeRAM still suffers from a more limited lifetime than conventional DRAM at this moment but I think there are still somewhat interesting developments happening in Europe.
To be fair, all high end chips are made in China with EU tooling. US happens to be the place where the managers are and the trade is facilitated. There has been some good progress to push Chinese make cips in USA with EU machines however IIRC significant part of the chain is still in China and chips made in USA are shipped to China for further processing.
AFAIK the high end chips are made in Taiwan, that's not the same. Also there was an export control, so ASML couldn't sell the latest equipment to China.
> To be fair, all high end chips are made in China with EU tooling.
Did you mean Taiwan? Because depending on who you ask that is, or is not china.
The Chinese have been cut off from EU tooling for a while now. And their attempts to build out their own lithography setups have been a failure. Their 7nm process is, by all measures, not delivering the yields that one would need to make it really successful.
Shovels have always been the better out of the two investments.
There were rumours about China grabbing hold of one of ASML's machines, but with no additional information on this it's safe to assume they're still hammering out 7nm chips using DUV at piss poor yields.
I have nothing to say about this company's long term potential, but I'm curious to know whether this will become an enormous "tik stok" like SK Hynix became in South Korea. I've watched a short documentary about the investment mania in SK and it's really scary.
Morgan Stanley's target price for SPCX in their report is $300, with $8 being the space component, and the $292 being everything else, like what actually gets put into space. The Morgan Stanley report doesn't say SPCX is worth $8/share. Just wanted to be clear and not have a conversation based on vague implications.
If 30% of adults in your country suddenly lose, say, half their wealth from a stock market crash, it's going to affect everyone in the country - not just the investors.
To use an example, the Great Depression wasn't sunshine and roses even for the people who kept all their money under the mattress. Their job was just as likely to go out of business.
I lived through the crash of 1989, the Dot Com crash in 2000, the GFC of 2008, and whatever the crash and WSB speculation in 2020 were called. None of them had a tiny blast radius like you suggest.
Why does the US stock market keep having record years while the population doesn't seem to benefit, whereas the opposite is happening in China? Isn't that a failure of US capitalism?
The population in china is falling off a cliff....low birth rates, collapsing property prices (the main vehicle for personal investment), slowing ecconomy.
This memory stock popping on IPO has nothing to do with how well/bad capitalism provides in the USA. This is entirely about Chinese investors speculating on this company's ability to grow and undercut some Samsung, SK Hynix, Micron. Nobody knows on IPO day if a company will grow into the expectations of the investors.
And China's population mostly invested in Evergrande real estate, not stocks. Lots of workers in the US are at least invested in the stock market via their 401k/IRA, but the same isn't true of Chinese workers.
Also worth reiterating that most US stocks aren't doing great. The top 10 by market cap keep dragging the indexes into growth, but success is heavily biased towards the monster caps.
If the golden stick of memory sitting on Trumps desk has its intended effect - and the US allow Apple to buy memory from CXMT then another 500% surge might happen.
The more interesting question is: why does the rest of the world not increase their production prowess? I am aware of most of the reasons, but I am getting tired of the "China is better than the rest of the world". It is more that China did a lot of things that the others either failed to do or no longer wanted to do. Case in point is Germany: from "our cars are best" to "hey, chinese cars are now better" (or, cheaper, for a comparative quality). This constant complaining is so strange. Would that energy not be better spent to try to become better?
It is because in the west we suffer from self inflicted obstacles that make it impossible to get out of our own way. The Chinese gov must think they are playing on easy mode.
Because the capitalist west doesn't have strategic view, beside maybe on the defense sector.
Everything else must produce profits next quarter not in the next 3 years.
It's this. Wall Street won't allow american companies to make long term decisions and long term planning that cost money in the short term, because they want all that money immediately now. Like, right now now. If an executive promised they could deliver big profits in 5 years by spending the next two investing their profits, Wall Street would fire them and put in someone who will happily spend all the profit (plus debt) to do stock buybacks and dividend disbursals.
It's even worse, the entire ideology is that industrial policy can't, by definition, be better than what the market would decide without interference. So every kind of actual strategic decision making (i.e. investing a few hundred billions in the industries of importance for the next decade(s), or at least in the infrastructure supporting it) would actually make matters worse. That is essentially what neoliberalism is all about.
This is core reason why the German industry became so irrelevant in the last 2 decades: The rail infrastructure became pseudo privatized and now is in shambles (IIRC only about 60% of trains arrive on time). The PV industry was axed by the conservatives about 15 years ago and now is completely dominated by China. Even now, as the other EU countries demand strategic investments - and again the German conservatives say this ought to happen without any new debt. "We can only spend what we have", to quote Merz. Imagine China saying this about chip production, or even Sam Altman about compute hardware...
> Would that energy not be better spent to try to become better?
The problem is, shareholders have bled Western companies dry in the chase for returns on investments. There is no money left in the companies to retool them to be competitive with China.
Meanwhile, the CCP just pours dollars into their companies.
Unfortunately, there is no punishment in Western societies if you run a company into the ground as long as you pay your debts on the way out.
The rest of the world is chasing models of investment in procurement of goods and services that get stuck on neoliberal capitalist antipatterns where every action demands a short-term profit motive. These are now very large-scale, highly-complex-supplychain manufacturing projects, and you cannot remain competitive if you are just Putting Faith In The Market to solve your problems. There's too much planning and sustained, structured goal-orientation required for investors or for the sort of contract-bid, public-private actions Western politicals are used to engaging with.
It takes 2 years in EU to get approval to build a shed. Imagine how long it takes to get approvals for a chip factory which uses a lot of water and very toxic chemicals.
UK has been trying to get approvals for a new runway at Heathrow airport for over 20 years !!! now.
The EU isn't monolithic, you could probably do this fairly easily in someplace like Poland where there isn't a bureaucratic hell and a large supply of skilled workers and engineers. But of course, Poland and co. are the "lessers" in Europe where you outsource for cheaper labor, a prestigious thing like a cutting-edge chip fab would need to be built somewhere like Germany or the Netherlands, and so it never shall exist.
Mostly self-inflicted wounds in form of climate pledges, climate taxes and such. Because of the goal to achieve zero emissions goal and climate neutrality, we abandoned cheap energy sources such as coal and pushed our heavy industries out of Europe to China. We still need those products, so we import them from China while wondering why European industrial segment isn't growing. Even in the clean energy sector, most solar panels come from China.
Most of the price increase did occur in Q4 2025 and most prices are fairly stable since then. Maybe this is different for LPDDR, GDDR, or HBM, but I would like to see evidence for that.
Of course CXMT, Micron, and family will still be profitable, but maybe not 'surge 470% from IPO' profitable.
If we are talking about all knowledge, then I agree that the compression ratio is very impressive already.
Hopefully, future models can be trained to be even more aware of external knowledge, accessible through web search / RAG / whatever it will be then, and might not need to internalize much knowledge at all.
See Karpathy's "Cognitive Core" idea. (I don't have a good link)
There must be a limit, I agree, but there have been no signs of approaching it yet. The most recent cohort of small models have shown the biggest leap in capability so far.
Dogs are very smart!
LLMs haven’t been on the scene for very long. There’s still lots of room for efficiency discoveries. Plus, I keep hearing quantum is going to be a big deal in the next few years.
Standard Def TV was plenty for 50 years. But when more was on offer, everyone went for it, and now you can't even buy a 480p TV.
2K, 4K, and 8K+ TVs have been around forever, and although 4K has become the norm, it's widely accepted that there isn't much benefit for most people above 1080p
The 1080p thing is true in the age of 40” TVs but not so much now, think Costco has a 95” TV if not 100”.
And when (not if) we get to the 4k (or 8k) equivalent of LLMs, they'll just be baked into hardware and we'll have near instant responses while running locally.
There is no future where OpenAI, Anthropic, etc survive with their current business model; at some point we will hit a point where training a new model is done only every 5 years or so, and in that scenario we aren't going to be running models of pricey server GPUs with latencies measured in seconds and full responses measured in minutes.
We'll be running locally with sub-millisecond latencies and responses measured in milliseconds. There is no way for any big company to compete with the current business plan of selling inference or subscriptions.
In a way it is actually the same thing that makes us accept AI as working in the first place. It only needs to be good enough for human perception. The same is probably true for compute.
Plus there is a chance the actual scaling that matters is beyond our reach. Think instead of TB models, PB or ZB models. We don't even have that kind of information. Humanity in its entire history hasn't generated 1ZB of information.
And yet when we hit 4k, that's were people just stopped buying higher res. 8K is still useful, but only when the screen is so large that it doesn't fit in the room :-/
https://github.com/JustVugg/colibri
Maybe? Human science history is absolutely littered with examples of things that were “constrained” by a fundamental law … until they weren’t, because we’d misunderstood or misapplied the law.
Some groups are baking models into silicone, Deepmind has an example, it gets 18,000 tokens/sec on Llama 3.1, not sure about parameter size
Sure, it might be very near using the current approach, but... it might also be might be very far off because we are using the wrong approach.
I mean, look at the max amount of intelligence you can get out of a human brain powered by two bananas...
It's like people saying mobile chips are going to be better than PC chips... until they realize PC can be made of mobile chips too if that comes true.
Also, you will be able to play with much more competent models locally. They will still feel like children compared to the adults living in the SoTA region.
Their posterior. Or worse, their interest in going to the moon (HODL style, not Artemis style).
Indeed it seems quite possible we are one architecture breakthrough away from existing chip stock driving us all the way to ASI.
Nobody has done AFAIK the information theory to prove it’s not possible.
Hyperscalers are screwed, data center mania couldn’t even be completed during this massive spending spree, all while the people seemingly sitting on the sidelines are working on getting these things baked into the OS and chipsets in a way consumers wouldn’t notice
On top of this, once European giants like NXP (formerly Philips semiconductors), Infineon (formerly Siemens semiconductors), and ST are actively reducing their footprint in Europe and moving to India or SE Asia.
In short: No, EU has no future in those markets as designer or producer of chips.
You can see it a lot in AI, right now. AI4industry is not where the gains are right now. We should be competing on training foundational models. Once the foundational models from US/China are strong enough, they will do AI4industry and we wouldn't have any edge on them anyway because their tech is more vertically integrated and they have better insight in how AI works under the hood.
The long-term idea that the we have to compete claim is based on is like AGI and it's gonna really care what Elon tells it to do. This is a dumb idea regardless of how it fails.
AI models are seeming the ultimate in fungible commodities at this point, why spend on R&D?
Difference between Europe and the others is that it doesn't brag nearly as much.
I don’t think the goal there was to sit around praising past accomplishments. You always have to keep building.
Did you mean Taiwan? Because depending on who you ask that is, or is not china.
The Chinese have been cut off from EU tooling for a while now. And their attempts to build out their own lithography setups have been a failure. Their 7nm process is, by all measures, not delivering the yields that one would need to make it really successful.
There were rumours about China grabbing hold of one of ASML's machines, but with no additional information on this it's safe to assume they're still hammering out 7nm chips using DUV at piss poor yields.
Aren't they the only ones able to supply the chip-making machines?
i mean, it shouldn't be scary, because their failure is their own, and i doubt there would be collateral damage outside of those high risk investors.
It's scary for _them_, but it's an interesting case study for an outsider.
To use an example, the Great Depression wasn't sunshine and roses even for the people who kept all their money under the mattress. Their job was just as likely to go out of business.
I lived through the crash of 1989, the Dot Com crash in 2000, the GFC of 2008, and whatever the crash and WSB speculation in 2020 were called. None of them had a tiny blast radius like you suggest.
This memory stock popping on IPO has nothing to do with how well/bad capitalism provides in the USA. This is entirely about Chinese investors speculating on this company's ability to grow and undercut some Samsung, SK Hynix, Micron. Nobody knows on IPO day if a company will grow into the expectations of the investors.
And China's population mostly invested in Evergrande real estate, not stocks. Lots of workers in the US are at least invested in the stock market via their 401k/IRA, but the same isn't true of Chinese workers.
Also worth reiterating that most US stocks aren't doing great. The top 10 by market cap keep dragging the indexes into growth, but success is heavily biased towards the monster caps.
So others did spend energy on becoming better and now you see a different picture when you look at the world.
But what about SpaceX, Blue Origin... there are also counter examples that show it is possible for investors to wait out the big profits.
This is core reason why the German industry became so irrelevant in the last 2 decades: The rail infrastructure became pseudo privatized and now is in shambles (IIRC only about 60% of trains arrive on time). The PV industry was axed by the conservatives about 15 years ago and now is completely dominated by China. Even now, as the other EU countries demand strategic investments - and again the German conservatives say this ought to happen without any new debt. "We can only spend what we have", to quote Merz. Imagine China saying this about chip production, or even Sam Altman about compute hardware...
The problem is, shareholders have bled Western companies dry in the chase for returns on investments. There is no money left in the companies to retool them to be competitive with China.
Meanwhile, the CCP just pours dollars into their companies.
Unfortunately, there is no punishment in Western societies if you run a company into the ground as long as you pay your debts on the way out.
UK has been trying to get approvals for a new runway at Heathrow airport for over 20 years !!! now.
Is this accurate? It looks like they are no longer rising, at least for DDR5 memory:
https://de.pcpartpicker.com/trends/price/memory/
Most of the price increase did occur in Q4 2025 and most prices are fairly stable since then. Maybe this is different for LPDDR, GDDR, or HBM, but I would like to see evidence for that.
how much more do they need to ship to satisfy you?
In China they already make integrated circuits with a finer pitch than at TSMC or Intel before the transition to EUV.
This is achieved by multiple exposures, which increase the fabrication cost, but with the current prices they still have very healthy profits.