We need to talk about data centres.

For the 2nd or 3rd time this week I've seen someone comment on a new data centre build with a stat about how 80% of data is never accessed. Then they talk about the energy and cooling used in modern DCs.

The reality is that data storage is actually incredibly efficient, and uses fuck all power. A hard disk is less than 10w and stores multiple users data.

Storing data, our photos, our memories, our history. Is not the problem.

What is? 1/n

The thing driving the need for the bigger more power and water hungry data centres is AI. Sparkling autocarrot. Where as a machine in a rack full of hard disks might consume a couple of hundred watts. A machine loaded up with a typical load of 8 "AI accelerators" can be pulling in the region of 5kw. Over an order of magnitude more power than the energy needed to store the lifes photos of hundreds of people.

And why ? To what end?

2/n

I've worked in this industry for over a quarter of a century.

At no point have I found myself thinking "I wish I could just ask the computer to write this email for me" "I wish the computer could write my code for me". MS is adding co pilot function to lots of products. Not as opt in. But opt out. And it's a right hassle to turn it off? Why? So someone can ask it to right a longer email from a prompt that the recipient can then ask the AI to summarise for them ?

3/n

There are certainly some areas where machine learning (note, I'm not calling it AI) has it's uses. Medical research springs to mind. But a ubiquitous AI assistance rolled into all our products? Why? It's just using too much power, too many resources, and for what ? Sparkling autocarrot.

Encoding the worst of our society in a bit stream. Exacerbating inequality, prejudice, and hate.

In comp sci there's a term. GIGO. Garbage in. Garbage out.

4/n

These large language models are being fed on the combined mass of the world's online content. Your tweets, Facebook posts, forum posts, that blog you forgot you started. All of it is being fed into the black box of the LLM. The internet provides us unprecedented access to the world's information. But it is also an unprecedented collection of hate. We've seen this time and time again. From chat bots that start shouting nazi propaganda, to CV vetting systems that won't hire women.

5/n

Garbage in. Garbage out.

And what makes this even more terrifying is that when you look at a webpage, it's often hard to tell if it's been generated with sparkling autocarrot, or written by a human. If we can't tell, then what hope does the LLM? And so we're gonna end up with the next generation models being fed on the output of the previous. This is going to create feedback loops. Reinforcing the worst the model has to offer. Strengthening the hate. The prejudice.

6/n

And because we don't know what has been created how. There's no way to control what feeds the models. It's just gonna enshitify. And fast.

AI has all the hallmarks of a bubble. Like crypto before it, and half a dozen other bubbles before that, that all share their heritage right back to the south sea bubble (no, not tulip mania, but that's something for a different thread).

Except this bubble has gone more mainstream. It's consuming way more resources. Than any before it.

7/n

Water is going to become the next big inequality front. As the climate changes. Clean fresh water is going to become harder to come by. More expensive, and more unequal. That same water is being poured over panels I'm data centres to cool the servers. To cool the AI accelerators, generating content noone asked for. Enshitifying the knowledge base of humanity. Just so a few people can make some money.

8/n

Storing our data, our memories, our photos, on servers in data centres that are built in sensible places isn't inherently a bad thing. And we shouldn't allow ourselves to fall for the trope of 80% of it is never accessed. But building datacentres that use ten times the energy, and need even more water, in deserts, and water stressed areas, to drive sparkling autocarrot that noone asked for. That we should be more vocal about.

9/9.

@quixoticgeek Training GPT-3 took as much water as 460 hamburgers.

You raise a very good point about location of data centers. There's plenty of room along the Wisconsin and Michigan coastlines with access to more fresh water than could ever be consumed in a thousand years. If water were priced appropriately in drought-ridden areas, I imagine data centers would be happy to relocate. First step is to fix the broken politics that subsidizes silage/beef

@jamiemccarthy @quixoticgeek why is fresh water being used to cool data centers? The data center machines couldn’t care less if it is clean or not. That seems a big oversight.

(And before someone says salt water corrodes, why the f would you need to put the water right against the machines, it just needs to absorb thermal energy, not make out with it.)

@davidaugust @jamiemccarthy @quixoticgeek Some companies do use seawater when when it's the best option.

When I was at Google a number of years ago, we had facilities in Finland using seawater, and facilities in a number of other places that used grey water.

(grey water being sewage, runoff, etc. that's gone through water treatment)

For a long time most companies used colos run by companies that had zero incentive to make their facilities power-efficient (they were the "industry standard"). However, over the last few years, I think a lot of companies have followed Google's lead and stopped wasting massive amounts of energy and water on cooling.

(Microsoft has done some really interesting work with passive cooling by sinking clusters in sealed pressure-vessels under the sea)

@distributednerd @davidaugust @jamiemccarthy @quixoticgeek Excess heat is like any other pollution...there is no "away." I guess sinking it in the ocean depths has less impact than dumping it in a river, but it's still going to disrupt ecosystems on a local scale.