MSFT needs so much power to train that it's considering small nuclear reactors | Hacker News Reader

MSFT needs so much power to train that it's considering small nuclear reactors

futurism.com

56 pointssigna1158 comments

The article on futurism.com references an article on theverge.com https://www.theverge.com/2023/9/26/23889956/microsoft-next-g... which in references a job posting on microsoft.com https://jobs.careers.microsoft.com/global/en/job/1627555/Pri... This story very much reminds me of "The Submarine": http://www.paulgraham.com/submarine.html I.e. Microsoft is not pursuing SMRs, but some marketing agency have successfully planted the job posting and is successfully making media amplify the buzz. Or if the story is true it is still just a job posting, and you can't draw that big conclusions from it.

I was thinking that someone at Microsoft wants their competition to waste a lot of money investigating nuclear for their data centres.

I've been watching a lot of Mad Men recently though.

I don't understand. How would a marketing agency plant a job posting in the Microsoft careers site?

I think the message might be "We are big on AI"

> Nevertheless, the company signed a power purchase agreement with Helion, a fusion startup founded by OpenAI CEO Sam Altman earlier this year, with the hopes of buying electricity from it as soon as 2028.

Except that's not actually true, Sam Altman had nothing to do with that company's founding though he did play a role in leading Series E funding in late 2021. I was a bit spooked by this because it sounded like a clear-cut case of corruption as there is no way Sam Altman's start-up would succeed where many others have failed... but now it's a bit more murky and less clear the level of funny business involved in that deal.

Perhaps Sam Altman found that startup in the same way Musk found Tesla, by making an investment and being recognized as a co-founder being one of the terms.

I'm almost as anti-Musk as it gets, and this parallel is still disingenuous. The duration of time necessary for a company to reach its Series E, and the act of investing at that duration of time, is NOWHERE near equal - or comparable - to Musk's investment in Tesla just 7 months into its lifetime.

Nuclear power plants to offset power use from new tech: Good.

Nuclear power plants being run by Microsoft Corporation: Concerning.

Why?

Past history of quality control issues

>In todays news, Microsoft reactor management tools have been found to be signed using stolen keys after one reactor was forced into an emergency shutdown after hackers reduced the coolant flow rate

I can see it now...

"It looks like you're trying to avert a nuclear meltdown. Would you like some help with that?"

[YES] NO

[ YES ] [ MAYBE LATER ]

If you think MS would likely be worse at security than whatever other shop could be running your local reactor, I think you are either not being serious or you are getting confused by visibility and what the difficult things about IT are.

Trying to secure computers for 6 billion untrained randos, who will do unimaginably dumb things to the system you provide, which you have to keep super accessible while doing so, might not look very exciting, but is probably at least several magnitudes more difficult than securing a highly controlled, controllable and (I assume) strictly audited power plant.

Because Microsoft has a difficult relationship with information security and developing internal corporate procedures that are more conducive to growth and safety than to pleasing interpersonal and interdepartment disputes.

To be fair, because of my job I hear about a lot of issues at the national level, and ICS (like power plants, water treatment facilities, etc) have terrible track record of security (think: admin panels with default password just left open on the internet). I think even Microsoft will better than that, since they actually know what they're doing (compared to a bunch of engineers managing the plant).

Oh, yeah. ICS/SCADA is definitely the worst industry I've heard of as far as security is concerned.

I don't think Microsoft has anything to contribute to that, because no, I don't think Microsoft possesses the technical competence to deal with these systems. They might have more /generalized/ infosec knowledge, but snark aside, I doubt they have the resources they need to really dig into SCADA security simply for the fact that they've never really needed to. I think it'd take them years.

It'd take them years to build a power plant, though...

> I don't think Microsoft has anything to contribute to that, [...] simply for the fact that they've never really needed to.

The unfortunately still very popular OPC protocol suite (except for the newer OPC-UA) runs over DCOM (making it very annoying to interface with if you're not running Windows). The hardening changes Microsoft did to DCOM, starting with Windows XP SP2 and still ongoing, directly affect the security of these systems (and also make them even more annoying to interface with if you're not running Windows).

I'd think private enterprise would be the last people able to be trusted with nuclear power. Not to mention there seems to be no support from dc for building them to power, you know, useful things like lights and heat.

How many Bitcoin networks' worth of energy are required to train the models that generate these silly JPEGs?

How many silly Bitcoin networks' worth of energy are required to train the models that generate these silly JPEGs?

FTFY

Exactly this. People who don't value Bitcoin or what it stands for say it's a waste of resources. Same can be said about AI image generation and LLMs

> People who don't value Bitcoin or what it stands for say it's a waste of resources.

What? Even if you do value Bitcoin, you're going to make a pretty uphill argument that it's an efficient/effective use of resources to mine.

Less resources than traditional money printing/distribution/banking.

What about traditional credit/electronic banking?

Probably still yes if you include all the employees' professional carbon footprint.

Since banking is a solved problem, it sounds like getting rid of BTC would save some CO2 output.

As a completely random guess I'd say... 1/10_000_0000 of a BTC network?

would love if ai forced the market to adopt nuclear

This would be so cool, hopefully this helps get nuclear off the ground.

I'm 110% for renewables as a path forward, but the promise of nuclear power is still tempting. Small Modular Reactors seem to be the only way to make it work in an cost-effective manner.

There's enough players to give one hope this might come to pass: https://c3newsmag.com/five-of-the-worlds-leading-small-modul...

That said, there needs to be brutally transparent examination of operational safety and possible contingencies (e.g., terrorism, earthquakes, tidal waves, heat waves, etc).

Are you implying that nuclear is not renewable?

Tidal wave energy usage would cause us to lock with the moon in 1000 years

https://cs.stanford.edu/people/zjl/tide.html

Wow. Good thing it hasn't caught on and shows no promise of doing so.

Why doI suddenly hear the Majora's Mask music?

Rebuttal on 1000 years claim:

https://news.ycombinator.com/item?id=37383283

"Energy consumption has decoupled from population growth rates and economic growth." and no citations...

Every bit of human progress we make is through more energy usage. Are we reducing the rate we build infrastructure? Are we building and manufacturing fewer things? Have we reduced transportation of people, goods, and services? No, it always goes up.

How are we going to dig ourselves out from under the weight of oil? Building renewables. They require massive amounts of energy.

It's all energy. The more energy you can harness the more comfortable your life will be.

https://www.eia.gov/todayinenergy/detail.php?id=44095

https://corporate.exxonmobil.com/what-we-do/energy-supply/gl...

That said, there needs to be brutally transparent examination of operational safety and possible contingencies (e.g., terrorism, earthquakes, tidal waves, heat waves, etc).

Exactly how much examination and experience do we need to have with nuclear before we start using it more?

Moving to SMRs are "new" and it would imply a lot more reactors out there as well in possibly more "casual" environments (e.g., sticking them in existing coal plants to leverage the existing infra).

Then getting third parties to vet designs/operational plans to look for weaknesses...

My point is simply that as promising as it is (very), we need a belt and suspenders approach to safety.

I’ve always wondered why, if we can safely propel hundreds of people in an aluminum tube, 35k feet in the air, at hundreds of miles per hour ground speed, thousands of time day in and day out - why is safely generating power from nuclear so often dismissed as unachievable? It makes me wonder if we as a culture have lost our courage to tame insanely dangerous things.

It is achievable, but we've yet to see it come in at its promised cost. It didn't help that the industry made some easily avoidable mistakes.

I think the key problems are the size (big projects always exceed budgets and timelines), as well as the bespoke nature of each power plant. That's why SMRs are attractive -- they address both of those issues.

> if we can safely propel hundreds of people in an aluminum tube, 35k feet in the air, at hundreds of miles per hour ground speed, thousands of time day in and day out - why is safely generating power from nuclear so often dismissed as unachievable?

The level of safety is different. There have been plenty of total loss accidents in aviation; a similar number of accidents of that severity would be unacceptable in nuclear power. Nuclear power plants need (and, AFAIK, have) a much higher safety factor than airplanes.

Always wondered why having the military running submarine units for commercial power has not happened. They have the experience of running safely in more challenging conditions.

Because submarine style high enrichment pressurized water reactors are both a nuclear proliferation concern due to their fuel type (i.e. much closer to weapons grade uranium than normal reactors use) and INCREDIBLY expensive. The reactors built for nuclear submarines are performant, not economical.

Thanks! I knew someone would have the details. Are there alternatives you think would work in a distributed model like this that address these concerns ?

There's a push these days to move towards HALEU fuel in the US; it's basically downblended submarine fuel stock [0]. It'll take about a decade to really start the downblending and distribution process at commercial scale in earnest [1], and there's some projects for instrumentation, transport containment, and so on in the works. Once commercial scale feedstock production is proven out, we'll start seeing shovels breaking ground on new reactor sites.

[0] https://www.nrc.gov/materials/new-fuels/haleu.html [1] https://www.orano.group/usa/en/our-portfolio-expertise/advan...

A HN comment chain claiming the opposite of what you're arguing here, that they're much cheaper: https://news.ycombinator.com/item?id=26348839

A1Bs are not submarine reactors but I guess that's not important. If it were as simple as just taking a carrier reactor and plopping it in the middle of a corn field everyone would be doing it already. For their cost calculation that user is that user is making lots of mistakes. You cannot compare the A1Bs thermal power output and to the electrical power output of power stations while also ignoring all of the cost of all of the additional systems which have to be built into the power stations themselves to manage them. These systems also have to be built into aircraft carriers but they seem to be glossing over those costs. Its also necessary to look at lifetime costs. The peak power output of naval reactors is very high relative to their size because they need to make ships go fast sometimes. Key word sometimes. If you run a naval reactor at its listed max power output its expensive highly enriched fuel is not going to last nearly as long as its supposed to. This gives people incorrect ideas about lifetime cost since you are comparing to power stations meant to put out a lot of power near constantly. If you take into account all the costs naval reactors are not at all competitive with normal civilian nuclear reactors and that should not be surprising to anyone. Civilian nuclear engineers are not stupid. Nuclear power plants don't just neglect naval reactors for no reason.

Also, as previously mentioned, naval reactors have nuclear proliferation concern. The reason naval reactors get to go 30 years between refuels is because they use use uranium enriched to >90% rather than the 3-5% used in commercial reactors. Not only is this very expensive, it adds a long list of concerns to the already long list of concerns surrounding nuclear energy.

> A1Bs are not submarine reactors but I guess that's not important.

I'll admit I was assuming that these naval are broadly similar, at least in cost per MW.

> If it were as simple as [...] everyone would be doing it already.

I think the main implicit claim here (and it's usually more explicit) is that using naval reactors is the only viable end-run around the morass that is civilian nuclear regulation.

> You cannot compare the A1Bs thermal power output and to the electrical power output [...]

True, but "thermal power station"[1] part of any nuclear plant is a well-solved off-the-shelf problem, as opposed to all the bespoke nuclear parts of it.

> The peak power output of naval reactors is very high relative to their size because they need to make ships go fast sometimes. Key word sometimes.

Do you have a source for this? I'm not doubting that you're right in practice, i.e. that naval nuclear reactors have variable power output, and spend a low part of their overall lifetime at a high percentage of the potential power output).

But by extension that means that they've got some average power output at which they could run perpetually over their shortened lifetime. What's that output?

> Civilian nuclear engineers are not stupid. Nuclear power plants don't just neglect naval reactors for no reason.

I don't think they're stupid, but they're clearly working within the narrow constraints given to them.

It's literally true that if our entire planet's electricity supply were generated with the RBMK-1000 design used in Chernobyl our mortality per MW would be vastly lower than it is today (look at current mortality statistics from coal etc.). So our current nuclear safety culture has clearly gotten out of hand.

> naval reactors have nuclear proliferation concern.

I don't think anyone's suggesting giving naval nuclear reactors to states that don't have nuclear weapons already, or at least those that are nuclear capable (e.g. Germany, Japan, etc. if they want highly enriched Uranium nothing's stopping them now).

If they're limited to those states I think these proliferation concerns can be dismissed.

1. https://en.wikipedia.org/wiki/Thermal_power_station

> Do you have a source for this [power output profile of naval reactors]

I was a submarine nuke reactor operator. Can confirm, we spent most of the time at essentially idle. There are effectively three things that will cause you to operate at high power on a submarine:

* Boat needs to get somewhere _right now._ I've spent days at All Ahead Flank, which is mildly unnerving at first, because every parameter you can think of is just absolutely maxed out. Imagine Le Mans, but for multiple days, and when it's done you don't get to rebuild everything.

* Coners (non-nukes) are practicing maneuvers.

* Return To Port. While your Squadron more or less dictates this, I have definitely seen dates flex a little if your boat mysteriously makes far better time than expected.

Carriers AFAIK have a somewhat higher average output, because they use their steam for flight ops as well as propulsion and electricity, and the topsiders have their own schedules.

> ...they've got some average power output at which they could run perpetually over their shortened lifetime. What's that output?

Actual power output for current naval reactors is classified. You can find guesses in various places, some of which are close, and some of which is hilariously wrong.

The question isn't what the current usage pattern of these reactors is, but whether they'd be amenable to different usage patterns.

Just like you might only use your car for 2 hours every day, but using its engine for 18 hours per day is probably fine.

Also, as far as I know, no one has ever dismantled a navy reactor. Every such reactor that has ever been built is in service or sitting in some yard. Any scheme that would make use of navy sized reactors would create a stream of reactors no one knows how to safely dispose off after use.

Compare this to civilian reactors for which dismantling and repurposing of the location is part of the cost calculation nowadays.

Recent discussion on hiring for Nuclear Manager

https://news.ycombinator.com/item?id=37627697

Yeah this feels like crypto again.

In case people missed this particular (micro)trend:

Bitcoin mining company buys Pennsylvania power plant to meet electricity needs [1]

A Bitcoin Boom Fueled by Cheap Power, Empty Plants and Few Rules [2]

Bitcoin miners revived a dying coal plant – then CO2 emissions soared [3]

[1] https://www.techspot.com/news/91430-bitcoin-mining-company-b...

[2] https://www.nytimes.com/2021/12/05/nyregion/bitcoin-mining-u...

[3] https://www.theguardian.com/technology/2022/feb/18/bitcoin-m...

Cyberpunk...

The article gave me Snow Crash vibes.

Obvious question: why don't Google/meta/etc have this problem?

What “problem” specifically? What’s the power usage/model performance for Google and Meta, who are both behind GPT-4? And, is that linear or some exponentially growing thing, with model competence?

Nuclear is an incredibly environmentally friendly way to train. It’s a problem if they keep using gas/coal or sucking up renewables that could be used for something better.

Reply on news.ycombinator.com