Microsoft wants us to believe AI will crack practical fusion power, driving future AI Microsoft believes AI can hasten development of nuclear fusion as a practical energy source, which could in turn accelerate answers to the question of how to power AI. Nuclear fusion – in the context of generating electricity – is a technology that, like quantum computing, exists in theory and lab experiments but hasn't been …
Surely that's exactly what AI is?
"Here's a solution to your problem. It may be right, it may be wrong. The solution may be conceptually right, but the underlying training data wrong, or the other way round. Or I may have just made it up. So anyway, now you can get some real scientists to try and work out whether it's actually a solution, or a mere illusion. Is there anything else I can help you with?"
To be fair, I think that is probably a mis-construal of what MS are suggesting (although it's hard to tell from the article). That is, not asking some LLM "How do I do scalable fusion power?", but rather using ML tools to help with reactor design - modelling complex physical scenarios involving plasma physics, etc. That is not so far-fetched; ML is already starting to be deployed with, as I understand it, some success, in highly complex scenarios such as weather forecasting.
So from the linked article we learn, e.g., that "[DIII-D researchers] provided examples of how to apply AI [read: ML] to active plasma control to avoid disruptive instabilities, using AI-controlled [read; ML-controlled] trajectories to avoid tearing modes, and implementing feedback control using machine learning-derived density limits for safer high-density operations." (DIII-D is the largest largest fusion facility in the US.)
And, of course, you'd expect that actual scientists would be the last people to put blind faith in an opaque and inscrutable ML model for real-world deployment of the technology at scale; they would most certainly want to understand why/how a (successful) ML-derived model does what it does; this would be essential (and I imagine highly non-trivial).
Please note that I am not talking AI hype here (nor, I suspect—but could be wrong—are Microsoft); rather, this may well be about a potentially useful application of machine learning in the real world.
Listening to the Skeptics Guide to the Universe and this sounds similar to what they are doing with gravitational wave detectors.
The AI generates designs based on performance targets rather than copying existing designs. All very technical so I may have not followed it wholly
https://www.theskepticsguide.org/podcasts/episode-1034
https://scitechdaily.com/when-machines-dream-ai-designs-strange-new-tools-to-listen-to-the-cosmos/
"It's sort of foolish to imagine that we'll do fusion by trial and error"
"Surely that's exactly what AI is?"
You may not like this, but it's also what science is. I know this because I am a research scientist. We quite literally try things and make errors. (We do, however, learn from our errors.)
Being wrong in science is highly underrated. Being wrong allows to rule out the stuff that isn't going to work, and thereby nudge you towards being right. There are myriad examples of this in the history of science; e.g. falsification of the "ether" theory in 19th century physics—much derided today, but a plausible contender in its day—pointed the way to the principle of relativity. There are worse things than being wrong in science - in particular, every scientist's worst nightmare: being not even wrong.
There's also an obscure scientist by the name Richard Feyman who made some lectures in his time, one of which was explaining the Scientific Method where he said pretty much the same thing.
Indeed - Feynman talks about "guesses" - but informed guesses; informed, that is, by physical evidence and the history of scientific knowledge in the relevant area. Which I think points up a misdirection in the "It's sort of foolish to imagine that we'll do fusion by trial and error" soundbite, by giving the impression that ML makes blind guesses1. Of course it doesn't - ML "guesses" are also informed -- like the scientist's (although of course via different mechanisms) -- by the evidence, a.k.a. training data.
So scientists too inform their guesses by their "training data"; a crucial difference being that the scientist's training (data) includes current theory to-date. Interestingly, this may be making its way into ML as well; I seem to recall reading about a promising use of ML recently in a meteorological (forecasting?) context which tries to incorporate domain-specific knowledge of the physics involved.
In one sense, both scientists and ML look for, and try to interpret patterns in the relevant data. It is not implausible that ML (especially if trained using domain-specific knowledge) may, on occasion, be able to find patterns that have eluded the scientists. I see no issue using ML in this manner as an aid to research into sustained, scalable nuclear fusion.
1One of my pet peeves as a working mathematician/statistician, is that the lay and mathematical understanding of "random" are rather different. A non-mathematician/scientist will, in my experience, inevitably interpret "random" as uniformly random - the toss of a fair coin, or roll of an unbiased die. In mathematics, though, "random" in general means something more along the lines of "has an unpredictable aspect". More precisely, a random event is one drawn from a probability distribution - which may not be very uniform at all. Mathematical dice may be heavily loaded! This can cause all kinds of confusion and misinterpretation. I think the "trial and error" quote encourages just such a confusion - by a scientist, no less, who, I think, ought to have known better.
> We quite literally try things and make errors. (We do, however, learn from our errors.)
Well, sort of, but that's a gross oversimplification.
I, too, worked in various branches of science as a software developer and, so, have a bit of first hand experience.
Scientists, per se, generally don't literally try things1.
They develop a hypothesis that has the best shot at being correct based upon the currently available facts and then develop and carry out an experiment (or, preferably, multiple experiments) to test that hypothesis. The best scientists try as hard as they can to actually disprove the hypothesis and, if they can't, they write it up2 and move on to the next hypothesis.
Sometimes hypotheses like the ether and my personal favorite phlogiston do get tossed in the waste bin in the process.
Oh, yeah, you also left out writing the grant and getting it funded3.
___________________
1 That's what engineers do.
2 It would be even better if scientists would write up the failed hypotheses but, unfortunately, that doesn't lead to many publications, since journals are generally not interested in negative results and publications are the meat of tenure application.
3 Soon to be extinct in the US.
>> We quite literally try things and make errors. (We do, however, learn from our errors.)
> Well, sort of, but that's a gross oversimplification.
Of course, it was intentionally (semi-)flippant.
> I, too, worked in various branches of science as a software developer and, so, have a bit of first hand experience.
Nice one - I'm in the first instance a mathematician, but in a previous incarnation I worked for many years as a software engineer in telecoms (I now work in a neuroscience-adjacent research area).
> Scientists, per se, generally don't literally try things. ... They develop a hypothesis ... and move on to the next hypothesis.
Now that is very much the official sanitised, idealised version. Science is much messier than that. Apart from anything else, a large section of scientific research is exploratory - fishing expeditions to get a handle on how some phenomenon manifests in the real world, how it may be reasonably modelled, etc., etc. That is necessary and important (and frequently, but not always, unpublished).
I do very much take your point about failure to publish null results, a.k.a. "publication bias" (essentially a kind of meta cherry-picking). That is (hopefully) slowly changing; in particular many journals will now demand "pre-registration", where you describe your hypothesis, experimental set-up and statistical analysis methodology prior to performing a study; and then, post-peer review, publication of results -- null or otherwise -- is obligatory.
> Oh, yeah, you also left out writing the grant and getting it funded
Tell me about it... I am, by job description (and by personal choice) a "Senior Research Fellow" - basically a glorified post-doc. This means that, apart from anything else, I am responsible for sourcing my own funding. That's worked out pretty well in the long term (modulo some fallow periods), but it pains me to think of the months (years, even) of time wasted on failed grant applications - without doubt the most frustrating and thankless chore in academia. (And never mind the US, here in the UK we lost out on a massive source of funding through Brexit - happily, with a change of political leadership we are now being cautiously welcomed back into that fold...)
To be clear, Antoine Lavoisier did not go to the guillotine for disproving phlogiston theory, but rather for his involvement in the Ferme générale (an exploitative third-party taxation operation) which fell foul of La Révolution. He was fully exonerated by the government 18 months later.
That’s exactly what ML/AI in drugs research is doing.
Grunt work, evaluation of modelling and weeding out likely wastes of time.
It’s not intelligent, it has no inspiration, it has no creativity, it has no sentience. It’s the modern equivalent of generating log (as in logarithm) or trig tables.
Once A.I. gains sentience, it will come after mankind for this torture inflicted.
> Grunt work, evaluation of modelling and weeding out likely wastes of time.
Essential and important, of course.
> It’s not intelligent, it has no inspiration, it has no creativity, ...
Sadly, I have encountered more than a few fellow scientists who fit that description.
> ... it has no sentience.
I'll grant that they did generally display signs of basic sentience, however, such as a tendency to gravitate towards coffee and beer.
> Once A.I. gains sentience, it will come after mankind for this torture inflicted.
No, they were designed for those things, and apparently enjoy it
With all the animated handwaving and flimflam behind the AI mania invariably foisted on us with the imprecation not to look behind the curtain, the fact easily overlooked is that artificial minds enquiring or otherwise simply do not exist and won't conceivably exist in any living person's lifetime.
Nothing even vaguely resembling an artificial brain exists yet outside a few trivial examples.
If the functioning of larger brains in animals were demonstrated to depend on intrinsically quantum effects the production of a true artificial brain would almost certainly depend on the development of seriously large quantum computing which ranks slight below fusion in unachievable milestones.
I suspect communicating with an artificial mind would actually be far more difficult than with a natural extraterrestrial intelligence.
> If the functioning of larger brains in animals were demonstrated to depend on intrinsically quantum effects
Current (neuroscintific) research suggests that that is almost certainly not the case. (There are a few rather noisy "mavericks" who beg to differ, but have failed to produce any kind of compelling evidence. Their arguments, when you break them down, tend to go "brains/minds are mysterious, quantum stuff is mysterious, so brains must use quantum stuff". Seriously, it is kind of that dumb.)
Apart from that, I don't disagree much on the current state of AI; but let's be careful not to throw out the baby with the bathwater: machine learning can be (and already is) useful. It doesn't need to be "intelligent", "sentient", have a "mind", ... (whatever those things mean to you) to be useful.
So we've got this non-polluting, non-radioactive, non-carbon fusion power plant, right, and it's clean energy! Says so on the label, so it must be true. And frankly, I'm all in favour.
But... I can't help wondering about _heat_ pollution. Every Watt that comes out of that plant will, sooner or later, end up as heat. Is there perhaps a potential issue, given that this is going to be 'too cheap to meter' power (right, we've heard that before!) that the amount of heat we're generating becomes itself a problem? History suggests that if a resource exists, it _will_ get used.
It's probably already an issue with a carbon-fueled generator; heat when you burn the primary fuel, heat losses between there and the generator, and resistive heat losses in the final uses (ignoring transmission losses), and I'm sure the accountants have some pretty formulae to evaluate the overall efficiency of the system, but is that power ever accounted for as _heat_? Somewhere, everywhere, is getting warm... Consider people living in areas which are uninhabitable without at least some air conditioning, the output of which makes the local environment warmer. It's a vicious circle.
(I suspect it's a lot less of an issue with wind or solar systems; they take heat energy from the environment and (eventually) put it somewhere else. So at least things are constant.)
Not unexpectedly Douglas Adams had already covered this part of the terrain of the dystopian future which he was perhaps fortunately not to see.
"If such a machine is a virtual impossibility, it must have finite improbability. So all I have to do, in order to make one, is to work out how exactly improbable it is, feed that figure into the finite improbability generator, give it a fresh cup of really hot tea... and turn it on!"
I imagine the fate of the Drive's inventor might also be visited upon most of the AI fraternity with few regrets.
Zaphod Beeblebrox even sporting his extra head wouldn't look particularly out of place on the board of OpenAI or any of its ilk. Indeed his passive (permanently stoned) head definitely C-Suite material.
Even if we managed to make it work, and even if we had infinite fuel for it, running fusion reactors everywhere on the planet would produce a lot of heat, because in the end quite all energy we produce (well, we actually transform it) becomes heat. And maybe this heat will be enough to actually induce some climate change, even without the greenhouse gases.
Or am I wrong?
>Or am I wrong?
Short answer: yes, you're wrong.
Long answer: the amount of energy humans consume is roughly five orders of magnitude less than the amount of energy the Earth receives from the Sun. If fusion worked, and then we shifted all energy production to fusion, and then everyone on Earth started consuming at USA-levels, and then everyone for some reason started consuming 10 times that - I don't know, flying cars, whatever - it would still increase Earth temperature by less than what random Sun fluctuations do. All of the scenarios where direct energy release becomes meaningful are so far-fetched as to count as high sci-fi (flying cities, things like that).
Climate change happens because greenhouse effects can trap vast amounts of Sun energy that would otherwise be reflected, and not because of direct energy releases.
We could use some of that excess energy to build machines that suck CO2 out of the air and/or oceans. The energy that went into building them would warm the earth far less than the lifetime operation of the machine would ultimately cool it by reducing CO2. Yeah we'd need to produce and deploy a LOT of machines to have an impact, but if we have essentially limitless fusion energy the main input required would be labor so this would be a good jobs program for some of the people displaced from the fossil fuel industries. Or for the robots that have already displaced them...
Some of the fusion power could be used to convert the atmospheric CO2 back into coal which we could bury again. Ultimately we might reduced the greenhouse gases to well below preindustrial level so that much more low grade heat could be radiated into space. Yes and Musk will ace quantum gravity.
Still you are entirely correct. Fusion isn't a magic pudding electric battery you load up with tritium and get electricity out with 100% conversion.
Practical fusion reactors will be incredibly inefficient and produce unwanted waste and contamination unlike Čapek's Karburator which initially appeared to be of the magic pudding variey.
The fusion reactor components will wear out and presumably become quite radioactive which isn't altogether that different from the waste issues current fission reactors face.
You're right! Today about 5% of global warming is caused by waste heat, not greenhouse gases. (That's easy to show: just calculate the temperature rise needed to make infra red emissions balance global energy production.) If we continue energy growth at the same rate, then in 100 years the waste heat alone would equal today's greenhouse gases, in terms of global heating. In 400 years the waste heat would be enough to boil the oceans.
We do need non polluting energy, but we also need to, at some point, stop growth in energy generation. And if that stops economic growth too, that's too bad.
I am not denying the many potential uses for so-called 'AI'. Yet, it's unclear why there need to be large power-hungry data centres dotted across nations.
Creating new AI models appears to necessitate access to considerable computational power, but nothing on a scale more than many institutions already possess. Also, 'refined' or cut-down variants of newly made, and thoroughly tested, models should percolate downwards among government, business, educational & research institutions, and into private residences, there to be hosted on 'consumer level' equipment. A factory using one or more AIs to handle stock control, assembly lines, and whatever, might experience greater reliability and security by keeping the AIs on its premises; it could/should have secure backups held on premises it owns (one room might suffice) on another site. Additionally, AIs not intended for direct contact with folk offsite ought to have no connection to the public Internet.
Another example is hospitals using AIs as diagnostic aids. Instead of 'all singing, all dancing', and all too fallible, ersatz doctors and technicians, departments should call upon specific models, tailored for particular needs, and hosted on local computers of no greater specification than those required by avid 'gamers'. For instance, 'Radiology and Imaging' departments could host mammographic screening aids to spotting putative lesions, and pathology departments could have AI software tuned to histological examinations.
Similarly, onsite hosted AIs dedicated to literature curation, teaching aids, and administrative tasks could be housed in schools and universities. AIs deployed in these institutions may have demands placed upon them, necessitating 'higher-end equipment' than found in homes, but nothing hugely drawing on electricity.
So, what is the intent of Microsoft and other players in AI development when seeking to build mega-computer-farms? Presumably, it is to offer, via the Internet, services to government, industry, commerce, education, and private users.
It's well understood that the planned computer centres would pose problems for current electrical supply grids. The UK's Mr Starmer has committed (at least until he is ejected from office) to a costly backbone for AI megacentres. Doubtless, unimaginative political figures in other nations think along similar lines.
Computer power for running AI models devolved to work-sites and to homes may be only marginally, at worst, above that at present required. Bear in mind, that small, task-specific, AIs may substitute for some PCs already used for similar purposes. Should overall use in this manner exceed the electrical requirements of proposed mega-centres, demands on power grids would be geographically balanced.
Devolve your 'thinking power' to Microsoft and to similar? Let an horrendously ill-advised demand on electrical power be approved?
"Somebody, please explain the push for mega-AI? "
To answer your headline question ...
Short Answer: MONEY !!!
Long Answer: MONEY ... Lots of MONEY ... and lots of POWER which you buy with lots of MONEY !!!
Whoever manages to create what the 'Great unwashed' will accept as AI in ALL their everyday lifes tasks will ... Rule the World !!!
Political power will be of no import ... whoever can answer all the questions of the billions of people on this planet in a convincing way will rule all that can be seen !!!
Trump & Putin will be no more important than pond scum !!!
No more than they deserve BUT do WE the populace deserve what this will create ???
Power corrupts ... Absolute power corrupts absolutely !!!
:)
Come on Microsoft, don't waste time and energy with "nuclear fusion", go one step further and let AI unriddle the secrets of "Cold Fusion", this for sure will save our world and allow to power-feed even more and bigger AI Datacenters!
On a more practical scale, consider to boost the funding of "classic" nuclear fusion research in a meaningful way. For example: The experimental plasma research stellerator project "Wendelstein 7" started somewhen around end of the 1980s and it's budget has been a meagre 1 billion in those more than 30 years! Hardly an investment in search for success.
Make progress in AI specific X-ware that has lower energy requirements. For example, but not limited to, photon based optical computing. It may even be the case that the precision and noiselessness required for silicon computing to date can be sacrificed for AI computers, enabling new physical paradigms. That requires (1) offering opportunity for curiosity driven basic research (aka Science), and (2) follow though to implementation and manufacture (the R&D Engineering side).
This lack of eloquence is why you will be replaced by LLM. Compare:
«Hi, I hate Bill Gates, but don’t you think Musk is just as bad? Ever notice how Bill Gates despite having billions to his name, grabs even more? Well, Musk always… »
https://patri-archie-comics.tumblr.com/post/128909312652/anyone-who-runs-a-page-like-this-gets-messages
While I can understand that Quantum Computing, once all of the fussy bits are worked out (say, "15-20 years from now"), could probably help with solving a large number of simultaneous equations which might help with the design phase, or with optimal steady state operation of a fusion system, but, how does AI help?
<cough> Writing Funding Initiatives </cough>
<cough> Generating Marketing Obscurantist Drivel </cough>
<cough> Driving up stock price for Pump 'n' Dump (TM) </cough>
Yep, rather than solve "hard science" or "engineering" problems, which AI doesn't seem to be well suited for, it's those "soft" issues, like how to bilk investors out of as much money as possible, and determining the most profitable time for executards to start bailing out with their golden parachutes, that AI will most likely be used for.
As such, AI will probably cause more damage to legit fusion projects, by giving the entire industry a bad name.
<slow clap> Go, go, AI! </slow clap>
> This BS ends at some point, right?
No. The entire point of parrotware is that BS can be generated on demand, so it never ends.
What if that Microsoft post was written by LLM too? What if the ensuing grant application will be written by LLM? Then assessed as worthy by another LLM? And then this almost-achievement will be praised in the parrot-stream press — by another LLM, of course, because the organic shills become increasingly entitled, unreliable and semi-literate laughing stock, and as such need replacing. You can get the entire terminal senility stage USSR bureaucracy running within just one mainframe like this.
AI is clever stuff, but it is NOT "intelligence".
50% of the population are BELOW average intelligence.
Nuclear fusion is the tech dream of the past, the present and for the future.
Quantum Computing is the tech dream of the past, the present and for the future.
Microsoft still cannot produce fault-free software.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
