Not to downplay it, but it's still only half as hot as would be required of a commercial reactor. Also this reactor had no mechanisms to recover energy or neutrons to breed tritium. Still impressive and encouraging.

Right but that's what ITER is for. This type of research is to validate control systems which can be transferred to that project (i.e. prove you can do it, then prove its not machine-dependent).

good one haha, the properly scary part of the other nuclear arms race is fusion too even!

The “20 years away” meme is stupid. There really are technologies that are possible but incredibly hard and require decades of sustained effort.

Cracking natural language comprehension with digital computers is an example from our field and it’s here.

>Cracking natural language comprehension with digital computers is an example from our field and it’s here.

Exactly, there are experts in the field less than a decade a way who said 50+ years easily. And there we are.

I would debate the fact that LLMs have "cracked natural language comprehension"...

Not that it's not impressive, but LLMs do not "comprehend", for a start.

I see what you’re getting at but it does feel like goalposts are being moved, no? By and large we can ask a computer today a question and it will almost certainly spit back a sensible (!= correct) answer. We can ask what the words mean and ask it to translate it to other languages, and we can have a conversation.

> I see what you’re getting at but it does feel like goalposts are being moved, no?

I don't think so. I am not setting the "goalpost" here, it was expressed as "LLMs have cracked natural language comprehension".

I just don't think they have. There are tons of statements I would agree with regarding what LLMs have achieved, but this one is not part of them :-).

Fair enough! Out of curiosity, in your words, what have LLMs achieved?

What is the metric to measure comprehension?

How are we disentangling comprehension of natural language itself from comprehension of the subject matter being discussed via said language sample?

I think that by most reasonable metrics LLMs can reasonably be said to comprehend natural language itself. However they clearly are deficient in logic and reasoning, as well as comprehension of many of the concepts that the natural language is used to express.

Even more impressive would be when humans can actually comprehend LLMs!

what's your definition of "comprehend"?

> The “20 years away” meme is stupid.

No, it's not. It's just a legit illustration of somethings state of development on fundamental levels. It simply means "we have no f**ing clue how we can do this, but future..". This is different from something we have already solved, and you just need to throw money on it to scale it to whichever level you need it.

> Cracking natural language comprehension with digital computers is an example from our field and it’s here.

That's the point, everything in research is always x0 years away, until the breakthrough happens and it's finished.

> This is different from something we have already solved, and you just need to throw money on it to scale it to whichever level you need it.

We can already do fusion, and by every metric it is scaling. Triple product is increasing, etc.

Fusion does not become a viable source of energy until it scales beyond a certain point, but there is no "leap" between here and there that we know about, just better and better containment. We are descending a gradient, not looking for one.

If we had never managed to get fusion outside, say, hydrogen bombs, then I'd agree that we have no idea how to do it, but we have -- using many methods. Tokamaks seem to be the best one for scaling it so far, but there's other possibilities that I wish we would research more.

That's not really how it works. We do know SOME ways to reach fusion, but they are not on the level we need. We haven't mastered fusion yet, we still need to research the foundations. We don't even know if scaling is all we need, and we don't know how to scale it to the level we seek.

Technology usually has a range of performance, what it can do and what not. And our technology for fusion-process is not in the range for a commercial reactor, so we still need a breakthrough in our understanding.

Hmm... in that case the analogy with AI is even better. This sounds like neural networks before things like deep learning and the transformer architecture -- before we figured out how to scale them. Turns out this did require some innovations. It wasn't just a matter of making a bigger model.

If reaching a sufficiently high triple product were enough, I'd agree with you.

But at some point the problem becomes not one of plasma physics, but of engineering and economics. Regardless of the plasma physics, the walls of the reactor can withstand only so much power/area and only so much cumulative neutron irradiation. Issues like this seem mundane and therefore easy, but they're perfectly capable of rendering a technology into a nonstarter.

I'm a layman, and so can't comment too specificaly. I found this Construction Physics article interesting, which was posted here some months back: https://www.construction-physics.com/p/will-we-ever-get-fusi...

I'm glad the bear case includes the "will never be economically practical" which is my core criticism of fusion, even with "high funding".

I also didn't see anything about vessel irradiation, which also never seems to be discussed. I get it probably isn't as big a problem as solid fuel rod fission in terms of waste creation, and tritium breeding may help, but it still will be kind of the same problem with LFTRs: a reactor design will fundamentally need an ongoing reconstruction/replacement strategy due to the vessel irradiation and transmutation from high energy neutrons.

Feel free to correct me if this isn't as big a problem as I think it is.