HarHarVeryFunny

What makes you think they don't have a domestic market?

The Chinese domestic EV market is the largest in the world.

Chinese factories are full of Chinese robotics.

Companies like Huawei are struggling to meet the domestic demand for AI accelerators, while growing volume very fast year over year.

The Chinese domestic AI market is massive, but we just don't hear much about it outside of the models that are also sold here. Can you even name the models that ByteDance (China's Meta) makes?

As far as exports go, China is a low cost producer for many things. If the US doesn't want to buy high quality EV's costing 1/2 the price of a Tesla, then there are plenty of other countries happy to do so. The US is only 5% of the global population - there are lots of other consumers out there.

It depends on what you are trying to make, but there are workarounds for most things, and it becomes more of a cost and volume issue than a showstopper.

If your yield is low and you still want the volume, then run more wafers, but then you need access to more DUV machines and more wafers. If your defect density is too high, then design smaller chips with a higher chance of avoiding defects, which is what Hauwei are doing - split processor into multiple chiplets (NVidia do this too, to raise yields and reduce cost).

HBM3E usually has 5-6000 vias, but doing this with multi-pattern DUV without defects is tough, so CXMT currently drop that to 3000, at a cost of some loss of thermal and voltage stability.

HBM:processor production ratios aren't what Huawei would like, so they mitigate this at system level by putting an optical memory bus on the GPU chip and sharing memory across the system.

Not everything is a huge LLM - smaller models like recommendation systems don't have so many parameters and need so much memory, so why waste HBM on them? ByteDance use their SeedChip acelerator for this, currently made by TSMC (using an older sanctions-approved 28mm process), which instead etches dense RRAM on die beside the processor.

There is also a time dynamic to this, with China still using pre-sanctions equipment and chips as their domestic alternatives ramp up to replace them. One interesting part of this is HBM ... HBM is very demanding to make, and everyone had problems with it, with initially only SK Hynix being successful. Samsung and Micron took a year or so to catch up, and during this time Samsung had made a ton of HBM that didn't meet NVidia's specifications, so ended up, pre-sanctions, selling it to China, where it has acted as a stockpile to carry them over as CXMT's domestic capacity ramps up.

China seems to be doing fine. Of course they would like sanctions lifted, moreso for HBM than anything else, but all that sanctions have really achieved is accelerating their semiconductor independence. They are still building 1T+ SOTA models, standing up 100K clusters of domestic AI accelerators, etc, etc.

On a related note, I was reading yesterday that apparently the real bottleneck for Chinese production of AI accelerators is HBM production, not processors or ASML equipment.

The lack of ASML EUV machines certainly hurts, and pushing DUV so hard results in abysmal yields of good chips, but you can compensate by running more wafers or making smaller chips, and the net result is that Huawei's Ascend production volume is limited by CXMT's HBM capacity not processor dies.

The problem is that HBM manufacture requires many steps (die thinning, via drilling, plating, alignment) where the equipment used by everyone else (Samsung, SK Hynix, Micron) is also blocked by sanctions, so the Chinese are having to develop all of this themselves too, which they have, but yields are currently low, even when using shorter HBM stacks.

Why would you want/need to?

The advantage of HBM over regular non-stacked DRAM is memory bandwidth, which also requires a super-wide memory bus - 2048 bits wide for HBM4. Compare that to the 128 bit wide bus of a modern CPU.

So to take advantage of it on the desktop, or anywhere else, you need that 2048 bit wide bus, and a processor capable of consuming 2-3 TB of data per second!

These are not normal requirements, other than for a GPU.

There may be a few scientists rejecting use of AI for that reason, but I would expect them to be a tiny minority, just as I would expect the number of developers who accept that AI is a powerful coding tool, but refuse to use it just because they want to feel superior, is very small!

A much more practical reason we're not yet seeing a lot of headline scientific mathematical breakthroughs is just that it is ungodly expensive! e.g. The Navier-Stokes result cost around $20M at API prices, and academics just don't have that kind of money to spend. You'll see more mathematical and scientific results from practitioners when either the cost of compute needed for these sort of brute force results is more in line with the size of academic grants, and/or the AI companies donate more compute to the scientific community.

There are different reactions from different mathematicians of course - Terrance Tao vs Cedric Villani, and no doubt a lot of shock at the speed of advance, but it seems the reasoned complaint why they don't want the AI companies themselves working on these problems is because the outcome is not the same - you get a result that in of itself may have been suspected or useless (Navier Stokes), but no write up of any new math or insights that were developed along the way, which is the real reason mathematics and people like Erdos pushed these famous problems in the first place - because they were expected to yield interesting mathematics, just as years of work on FLT had done. Imagine if instead of Wiles's work, and all that had gone before him, all we had was a $20M compute bill, hundreds of pages of impenetrable math, and a billion lines of Lean proving it was true?!

Sure - the AI companies want some quick trophy kills to feature in their IPO prospectus, but they are not going to themselves be cracking the genuinely tough problems.

There is a difference between what's easy/hard for a human vs computer, and LLMs haven't changed that. You might expect a computer to be good at tasks requiring prodigious memory and compute, and it turns out that some of these long-standing math problems are of that nature - not requiring new breakthroughs but rather just massive exploration of what is already known and what they were trained on.

There will no doubt be more math results like this, but presumably also ones that are "hard for a human, easy for a computer", requiring massive search (e.g. find an example/counter-example cf Navier-Stokes & Jacobian conjecture) rather than creativity.

It seems that physics -> chemistry -> biology is a layering of abstractions and/or emergent phenomena, and any explanations of mechanics at these higher levels is more along the lines of "how was this built from these lego blocks" rather than being some more fundamental discovery.

You keep on making assertions - perhaps you could back it up with some explanation of why you regard this as a more fundamental discovery (or however you would like to characterize it)?

What aspect of that is an unsolved problem, and one that feels more fundamental rather than engineering / in need of explanation ?

What seems ridiculous to me is putting together any such lists, even in math, and expecting/hoping the AI companies to solve them. The people working for the AI companies are not scientists or experts in anything outside of building LLMs. The best chance of making progress on genuinely tough (not just computationally challenging) tasks is to put advanced tools in the hands of actual scientists.

What would be your idea of a biology problem that isn't an "engineering" one?

It seems that understanding biology could be characterized as trying to figure out "how did nature engineer this".

For that matter isn't all of science this way?

The goal of the suggested "origins of life" problem, as written, isn't to guess at history - it's to create some similar plausible self-emergent system in the lab.

Seems a bit of an ambitious goal for a language model though! There are presumably dozens of steps before you get to an RNA-based full-blown cell, and it's only been very recently that humans have managed to make an artificial self-replicating "cell" (container) of any type in the lab.

That's not quite right - a fish does not have a cortex.

The pallium of a fish and mammal are structured differently. A fish's (& bird's) pallium has a clustered (aka "nuclear") organization, while a mammal's has a mostly layered organization which is what is referred to as the cortex.

There is also more to a brain than just brainstem + pallium. In a vertebrate it's brainstem (aka hindbrain) + midbrain + forebrain, with the pallium being part of the forebrain.

I've just being doing some googling to try to understand this research better.

It seems that it's been known for a while that more primitive animals have/had functionally distinct front (sensory) and rear (motor + autonomic) neural systems, with different genes (Otx, Gbx) associated with each. In 2003 it was shown that an extant species, acorn worms, have this front/sensory, rear/motor neural division, assumed to date back to the pre-chordates.

In vertebrates we also see this association of Otx gene expression in the fore/midbrain and Gbx in the hindbrain, but it had been assumed that these "two halves" of the brain had fused long ago... What this new research shows is that this fusion only happens during development of the embryo, which seems to be a fairly common phenomena "Ontogeny recapitulates phylogeny" whereby embryonic development follows the same steps as evolution (e.g. human embryo also has a tail, chicken embryo has teeth).

Of course the front/sensory and rear/motor neural systems aren't totally isolated, since the sensory signals serve to steer the motor programs running in the back, and our brain still follows this primitive division of functionality.

I really don't get any sense of what these events are from the look-alike posters. They are all just a wall of color and excessive detail.

For a poster to catch your eye and be informational it needs to be way simpler - should be able to take it in at a glance, not have to pore over it like a "where's waldo" exercise.

I was wondering that too - presumably they are all from the same image generator, but which one?

They are definitely all ghastly - way too busy for me to want to give a 2nd glance at them. Of the block of examples shown, the bottom right one on off-white background is a bit less visually objectionable than the others, but still gives me that AI slop vibe.

True - OpenAI did at least say they used their models to help design their Jalapeno chip, but AFAIK zero details on how they are using their models in their software development process other than to automate some part(s) of training.

Ziphu seem much more matter of fact about it. To me it' a shame that they've decided to the use this "RSI" name, but at least they are being fairly specific about what they mean by it, while the western companies seem to want to invite you to think it's more than just dogfooding and automation.

There is some truth (but not accuracy) to it, but not for this reason.

This article is about brainstem (aka hindbrain) vs the rest (midbrain + forebrain).

Evolutionally older animal lineages like fish, reptiles, birds all have midbrain + forebrain as well as the brainstem, so reptilian brain != brainstem.

Where there is some truth to our older "reptilian" brain is that the part of the forebrain called the pallium, that exists even in fish (& reptiles), has much expanded/evolved in mammals giving us our cortex and hippocampus, so you could kinda say that we have a reptilian brain + cortex. Note though that mammals didn't evolve from reptiles - they evolved in parallel from a common ancestor, so maybe better to say we have a fish brain + cortex (& hippocampus).

While the cortex is what gives mammals their intelligence, in birds the pallium evolved in a different way to give them theirs.

Honestly will not be surprised to see future Anthropic press release that their AI is so powerful it made something in their bio lab that escaped and killed people, and that the government really needs to step in and and stop them from doing it again.

Given Anthopic's self-admitted inability to control their own AI, it seems one of the last things they should be doing is putting the POS anywhere near a bio lab.

If an AI company wants to help with science, then give free AI to scientists who actually know what they are doing.

The reality is often going to be much worse than TSMC's N2 -> A14 "mixed design" 20% density increase marketing spin.

Apparently the N2 -> A14 density increase for logic is 23% (Tom's Hardware), and about 0% for both SRAM and analog.

Both CPUs and GPUs devote considerable space to SRAM (registers & cache), and are 50% or less logic, so for them the density increase will be more like 10% rather than TSMC's 20% marketing number.

The cost-per-transistor used to fall as density increased (Moore's Law), but has now basically flattened off.

As each generation of lithography tech (DUV -> EUV -> High-NA EUV) gets pushed to the limit, it becomes necessary to use multiple etching steps per layer (higher cost, less wafers-per-minute) where previously one had been enough, then the next generation (e.g. EUV vs DUV) resets that to one step, then that becomes two ...

The trouble is that each new generation of tech costs considerably more than the one it replaces, both in terms of machine cost and operational cost, so the overall trend seems to have become fairly flat.

A14 is more in keeping with N2, and it is anyways just a name, not a measurement.

What's wild is that the mirrors used in the EUV machines are smooth to a level of deviation even smaller that the feature size of the gates being etched.

These mirrors are up to a couple of feet in diameter, with an average smoothness deviation of less than the width of an atom, maximum bump size of 1nm or less. They like to point out that if scaled up to the size of Germany (Zeiss makes them), the biggest surface defect would still only be 0.1mm!

FYI, Nikon and Canon don't rely on ASML-owned Cymer anymore - they use a domestic Japanese supplier instead, Gigaphoton.

Isn't Cymer actually owned by ASML - push come to shove couldn't they just leave the US ?

FYI, the US is only 10-15% of ASML's sales. The bulk of their machines go to China, Taiwan and South Korea.

What I meant by "emotional state" (AFAIK normal scientific usage) is something with a concrete physical aspect to it - an altered state of mind/body caused by the release of neurotransmitters and/or hormones.

In a conscious animal there is also going to be a subjective experience of that as well, a quale of what it feels like to be in that state if you will, but that is certainly not what I was referring to, as I would have hoped was obvious - I was talking about prediction.

In any case, when the conversation becomes about the conversation, then surely it is time to stop.

Could go either way...

1) Looks like a duck, quacks like a duck - it's conscious!

2) Looks like a robot, built like a robot - it's not conscious!

I've always assumed that for the majority of the people it'll be 2).

So now you're trying to pivot to consciousness and qualia ?

There are other people in this thread who want to talk about that stuff, so try them instead.

I've got to wonder how long the US will still be able get much of the world to go along with requests like this. Europe is already showing some spine and refused to join Trump's war with Iran, and Italy and Spain have also recently shown Trump the middle finger.

It'd be ironic if Europe ended up banning ASML (a Dutch company) from shipping to US instead!

They already have working machines, and are slated to ship the first 5 early production ones right about now (before the end of 2026), and are projecting 20 for next year.

Small numbers perhaps, but this is happening right now.

What will the numbers be in 5-10 years time? Who knows, but ASML took about 5 years to go from 20/yr to 100+/yr.

Of course politically the world may well be quite different in 5 years time, as may be the AI market.

I said an LLM will learn anything that helps it to predict, then gave examples of playing chess by prediction and predictive emotions, both of which you seem to now accept, so you are now accepting that my "following paragraphs" did in fact follow. Go figure!

You want to argue that predictive emotions are just as real as animal emotions, but that doesn't stop them from being predictive (and that AI that smiles as it kills you still seems concerning).

¯\_(ツ)_/¯