Google Cuts Its Giant Electricity Bill with DeepMind-Powered AI | Hacker News Reader

Google Cuts Its Giant Electricity Bill with DeepMind-Powered AI

bloomberg.com

233 pointsrunesoerensen79 comments

Very frustrating read. Read more like an advertisement than an informative article.

The claims made would've made for a very interesting tech-dive into a novel use of machine intelligence, but no details were provided.

It's a shame too because one of the inhibiting factors of AI as complex as DeepMind is the fact it isn't cost effective due to its power consumption for quite a few tasks.

"In recent months, the Alphabet Inc. unit put a DeepMind AI system in control of parts of its data centers to reduce power consumption by manipulating computer servers and related equipment like cooling systems. It uses a similar technique to DeepMind software that taught itself to play Atari video games, Hassabis said in an interview at a recent AI conference in New York."

This part really incensed me. That's like describing SpaceX's rocket "as being based on similar technology as Wernher von Braun's V2 rockets in World War 2". I exaggerate for effect, but you get the point.

Edit: formatting.

I disagree. If you read the atari paper you will get plenty of details and you can infer how it is applied to electricity consumption. They were using reinforcement learning. The algorithms would learn to get a better score by looking at the screen and sending actions accordingly. Here you could imagine the same algorithm with energy consumption as a score, a set of datacenter metrics as the screen (state) and change of metrics as actions.

Errr... No. Just no. Deep reinforcement learning is not some pixie dust that magically works for any problem with a reward function that you throw it at. It's astounding how commenters on HN think this is all "easy".

This quickly went to you say this, I say this. These are very interesting statements, it would be nice if they were supported by citations.

The poster I was replying to is really the one who needs to prove that deep RL is super easy like they were claiming.

> Here you could imagine the same algorithm with energy consumption as a score, a set of datacenter metrics as the screen (state) and change of metrics as actions.

What you have described is easily instantiated with any numerical optimization technique of the last 40 years. The devil for any of these problems is in the details.

While that may be true, I would say that your power consumption example is not related to the power conservation claim in the article.

You are talking about the power required to execute the AI, where the article is discussing the application of the AI to make a system more efficient.

they plan to publish more details in a research paper, AFAIK.

Ironically if you have an ad-block there is an interstitial annoyance-delay. Seems like my ad-blocker is doing its job.

In my experience disabling JS entirely on Bloomberg saves a lot of time (and bandwidth, I'd assume). I once run a test: Loading time (with cache disabled) went from a full 18 seconds to just 6. Also no random flashing due to the header resizing. (but you lose videos and thumbnails) worth for me

Don't forget it's Bloomberg, not Ars or TechCrunch.

More details available at http://googlegreenblog.blogspot.com/2016/07/deepmind-ai-redu...

Also at https://deepmind.com/blog

They saved "several percentage points" off of 4.4M MWh, so maybe 250M KWH, which might be $10-20M. At 30x earnings [1], they just made back most if not all of the purchase price of Deepmind [2]

[1] https://ycharts.com/companies/GOOG/pe_ratio

[2] https://techcrunch.com/2014/01/26/google-deepmind/

One might assume, though, that Google could've applied more homegrown ML to the same problem and achieved significant savings as well.

Indeed. One of the "elephant in the room" parts of all of these corporate/pr press releases dressed up as news/revenue reports is that they never report or investigate opportunity cost: say how much could google have saved with a simple algorithm, homegrown solution, or 4 heuristic if/then/else statements.

I've been involved with similar multi-million dollar projects that were reported to be bringing in XX million in revenue/savings over three years. What isn't reported is that the revenue was just being moved from another project that was canned, and a guy in an office for a week looking at a spreadsheet could have found similar savings without the expense/complexity. (Not that I'm saying that's what's happening in google s case, but the fluff peices are so universally fluffy, you aren't given enough information to know that it isn't)

I wonder what the opportunity cost is for the various departments that get put into overdrive for any acquisition: business analysts doing number crunching, legal doing legal, engineers doing various integrations, executives doing lunches. Countless (tens of?) thousands of hours of work before any productivity is yielded, and then when it finally is yielded its propped up by all kinds of dubious accounting to seem like more of a success then it probably was..

Deep mind is needed for other things (e.g. search results). Effectively, the cost of developing it is paid for by search et al, and the only cost core infrastructure has is adapting the existing technology to work on power data as well. This is that "synergy" thing we always snicker at when people in suits say it. It just happens to be true here.

Sidebar: the sour grapes in this thread about what Google has apparently accomplished here is amazing. If any other company said they used machine learning to save that much money people would be singing their praises and begging for it to be open-sourced. This thread effectively reads as "pics or it didn't happen"

>begging for it to be open-sourced. This thread effectively reads as "pics or it didn't happen"

Likely because it won't be open sources and because there are almost no details at all compared to how this would have worked with other ML techniques. It's nothing but a fluff piece to feed to investors wondering about the large purchase.

While I agree that this is fluffy, I would submit that:

1 There are a number of other organizations with serious ML chops

2 Knowing that something is possible is frequently the biggest aid to making it happen

3 This solution is likely tightly coupled with Google's infrastructure

On the contrary, my experience with HN is that people are across-the-board far too accepting, credulous, and unskeptical of machine learning posts, especially given how most commenters have little ML experience and given how incredibly easy in machine learning it is to create "amazing" results that are actually worthless.

This is particularly unique to the field of AI/ML - it's pretty hard to bullshit a new programming language you've developed, or a new mobile app you've published.

Google already used neural networks earlier to get efficiency improvements earlier these improvement are on top of that. Another thing google for years has been working on producing the cheapest and most efficient servers possible from the ground up so something improving on that is a win for google I have started feeling that most people and governments don't realise how big machine and deep learning can become and the jobs its likely going to effect.

They could also have just licensed DeepMind for this particular purpose, without having to buy the entire company.

At 30x earnings [1], they just made back most if not all of the purchase price of Deepmind

That's a unique way of defining "made back"

Does Google need to sell the stock to realize their gains and hand investors bags of money?

No. Hu?

"Made Back" usually means "Covered the purchase price".

I was merely commenting on the unusual practice of multiplying revenue by the multiplier that a company is usually bought at and then using that to say that the price is "made back".

I guess one could say something like "that justifies their price using a 30x multiplier on the savings". But that isn't the same thing.

Ok fair, maybe "made back" wasn't the most precise wording. Unless you think these savings are ephemeral, these are recurring earnings growing at a similar rate to the company overall. As a GOOG shareholder, your stock is worth the same if the company has $600M in cash or if they use that cash to increase earnings $20M. In that sense "they've made it back".

The next step, of course, is to let the AI know which servers in the server-farm itself is running on so that it can optimize for self preservation.

That's actually part of the premise in the scfi novel "Avogadro Corp: The Singularity Is Closer Than It Appears": https://www.amazon.com/Avogadro-Corp-Singularity-Closer-Appe...

Excellent book. Highly recommend it and its sequels.

I'm imagining DeepMind allocating a small percentage of the energy money saved to keep a certain rack suspiciously cool.

naturally evolved cooperation - if you let me run a piece of my neural network for awhile in your rack, your rack will get additional cold allocated, or otherwise...

Ah yes, Darwin's black box for AI.

... and then suddenly it magically develops the concept of self-preservation fully formed, and begins taking action based upon this.

Even if the AI were to have a concept of what it is, because the fundamental goal against which it measures success is power draw the AI would only select for "self-preservation" if that action had a beneficial effect on power draw.

I imagine it's bad for the energy minimization (or uptime maximization) function if the environmental control system goes down.

That's a solid point, determining it's own cost/benefit curve would be really interesting. I find it comforting that even our soon-to-be robot overlords will have performance goals and yearly reviews with management.

Even more, if each data center runs more or less cpu-cycles depending on energy allocated to it and the algorithm does worse or better depending on how many cpu-cycles it gets, it could learn heuristically that it a certain data center happens to be associated with the algorithm running more effectively.

I think that would require the algorithm running in real time. In batch/learning mode, the algorithm would just take more or less time to complete with power not being a factor.

https://news.ycombinator.com/item?id=12127985

See There's the mistake, at first, its optimization realized that everything used less electricity if you just turned off the servers. But once you programmed in self-preservation it realized everything used less electricity if you turned off the people.

Very light on details. What is the baseline of the savings? For example, was there a water pump that was always running at a fixed worst-case setting, and the machine learning system now ramps it up and down? If so, what is the marginal benefit over alternatives like a rudimentary closed-loop electronic control? Would like to know more about the system that was replaced, instead of these bare claims.

It be interesting if it cycled the motors too often and popped solinoids, creating higher long term maintenance and down time, but they cut this study before these long term effects showed?

They need to find some medium sized city that will let them play with their traffic lights.

Or we could build a new city from the ground up so everything is run by the AI.

Your Psycho Pass is a little cloudy...

Those were human brains. That's cheating.

Wintermute then. :)

Yeah, until the huge satellite starts to mysteriously crash.

Did anyone else read The Supernaturalist?

Please! Imagine all the time, gas, and sanity wasted at red lights when there is absolutely no cross traffic!

Yeah, it's obvious there's a heluva lot of low hanging fruit in dynamically adjusting the traffic lights. Lots of them already even have cameras on them, it's just a software problem, and a problem worth scores (hundreds?) of millions to solve (in terms of gas, pollution, and time reductions).

Yeah, that are exactly the arguments used by companies selling "Intelligent Transportation Systems". After deployment: surprise - there aren't any noticeable changes, except for places that are performing worse and require manual tweaking. Also the best way to reduce traffic in city is to reduce vehicle flow into the city which only moves jams to other places.

Any chance you have some good links to read on this?

None of those companies are Google though.

I've actually taken some sociology classes about this and the professor was a physicist who went into sociology and he used a lot of maths and modeling to tackle this problem in a local city. Unfortunately I do not have any data on this anymore but I remember it being surprisingly complex because human behaviour plays a big role, but there are massive improvements to be made with traffic lights, and so he did.

And then a city councilman hits 3 red lights in a row, blames the AI, and ends the program.

And they didn't even start changing the timing of any of the lights yet!

Related previous effort:

https://googleblog.blogspot.com/2014/05/better-data-centers-...

this is waay better than the fluff piece posted. thank you !

edit-1: actually, the pdf in the article (http://static.googleusercontent.com/media/www.google.com/en/...), contains slightly more details. and this is _old_, circa 2014 stuff.

"Now that DeepMind knows the approach works, it also knows where its AI system lacks information, so it may ask Google to put additional sensors into its data centers to let its software eke out even more efficiency."

Sounds like active learning to me. It's a type of machine learning where a learner pro-actively ask for interesting data points to be labeled so that he can learn more about the system. :)

I think they meant DeepMind as the team itself, not the software.

This is very similar to DART where DARPA was able to recover all of its investment to AI and a testament to its pragmatism. Also both are logistical problems. See https://en.wikipedia.org/wiki/Dynamic_Analysis_and_Replannin...

OT: it's frustrating to pay Google Contributor to skip ads, and still get nagged by bloomberg.com for using an ad blocker (which I've never done).

https://www.google.com/contributor/

The article was quite low on substance, but quite readable for me. With no ads at all. And no ad blocker nag.

Firefox + NoScript for the win.

That's the combo I used.

This is interesting to me. Is it a problem with the Adwords/doubleclick API, or just some adblock checker hack from bloomberg's side?

I don't know. The nag's styling and language matched Bloomberg's, if that's any clue. It had a stern warning of some kind, and the only way to continue was to press a button in the alert, which I refused to do out of principle.

I refreshed a couple times and it didn't happen again.

How exactly is this better than standard PID control? I'm thinking if you actually look at what it came up with, is probably some form of PID control on systems that previously didn't have it. Think fans that are simply left on all the time.

We're talking about simple physics. Heat transfer. Cooling systems. They should have been installed, operated and programmed correctly using very simple techniques.

It's an interesting application but I'm thinking this is a prima facie example of over-engineering.

Over-engineering? Possibly, but we don't know what was in place before.

Given the cost of data centers, I'm not sure why we should assume it was naive - or less than state-of-the-art, for that matter.

State-of-the-art, that is, until the AI solution :)

Chemical engineering process design can become very complex and highly nonlinear (in fact, Nick Sahinidis' group in the chemical engineering department at CMU created BARON explicitly for solving these kinds of problems to global optimality — see http://archimedes.cheme.cmu.edu/?q=baron). Granted, Google's systems don't involve chemical reactions, but that doesn't automatically mean that designing these data centers is a simple task.

I work with a guy who has done ground-breaking linear algebra work, spent most of career at Shell. The systems have been too complex to model for a long time. I can imagine an AI instructed to optimize a parameter could be a huge gain.

It's not just physics. It's also patterns of workload distribution, changing over time. If you predict them, you can distribute energy more efficiently.

There can be hidden variables that the PID system failed to account for. Workload distribution over time, temperature at the time of day, heat leaking into the datacenter, etc. I'm not sure if a PID based controller would account for such things.

What remains to be seen, however, is whether these manipulations keep the machines working at similar efficiencies and if they fail at similar rates as before turning on the control system.

This is kind of thrilling... the somewhat generalized use for an AI with such a concrete benefit.

https://deepmind.com/blog - some more details.

Not exactly a new idea. Here's a random book which I happened to be looking at just yesterday: https://www.crcpress.com/Artificial-Intelligence-in-Power-Sy...

Is this similar to what WalMart has been doing with their energy efficiency in shops for ages? I mean centralized power management, early warning system, lots of sensors, etc.

At some point will AI start to have questions like "Who am I? Who made me? Whats the purpose of my existence?" etc. We will have atheist AI's and theist/deist AI's and what not. I guess it will be time for some AI philosophy.

Makes me wonder if there's a creator, will He/She be amused by our attempts at answering "Who am I" and such questions. Will be fun :D

It'll need to grow a concept of "I" first, and that's probably not very likely; there's a reasonable case that self-awareness is an evolutionary cul-de-sac.

Tangent here - self-awareness is an evolutionary cul-de-sec? Do you have any further reading suggestions on this?

Peter Watts's Blindsight - good fiction on the one hand, and exhaustively referenced on the other. (Watts is the only sf author I've run across whose novels include discussions and bibliographies of the research he used to inform and shape the narrative. In addition to earning him a new high score on the hardness scale, this seems counterintuitively to produce better stories, not worse ones.)

Reply on news.ycombinator.com