Rob Pike Goes Nuclear over GenAI

Assuming that the price of energy includes the costs of the external costs to society it's production cause

So, the fair price of making the planet unlivable in many regions is 18 cents per kilowatt hour (averaged across the US)?

Either the assumption is bad, or I'm way overvaluing a habitable planet.

What you are basically saying is that there is no such thing as too much energy in the laws of nature. Which is true. But for most people "too much energy" is a compresensible notion. I for example think LLM vendors are using too much energy because they are noticeably increasing household energy costs where their datacentres are, because increase in energy demand is heavily outpacing expansion of clean energy infrastructure.

I would argue that “too much” means the opportunity cost of the energy use for AI outweighs the benefits. One issue we’re currently facing is that both the costs and benefits are delayed. Another is that computing the cost of externalities is both complex and subject to political and other influence.

But the concept of “too much” is not at all nebulous. It’s the same reason I think certain people have too much money, or certain individuals and countries have too much power.

LLMs are making people less educated by discouraging them from reading content written by people with actual thoughts.

This is obviously incorrect if you spend any time trying to learn a new topic with an LLM. You get answers to your questions that are hard to exactly Google that are right more often than they are wrong. It's an amazing learning tool. Yes you have to be cautious about making sure you really understand, double checking what it says and making sure it's consistent with what you've learned so far, etc. But if you can do that there really has never been a better tool for learning.

Anyone could already learn to write software before AI.

Yes, and now people have opportunities to learn more things, faster?

These students have carefully studied the code and appreciated how it was written. AI training mindlessly gobbled it up like a tsunami gobbles up houses.

So? How does this affect whether or not students appreciate something? Who cares whether or not AI appreciates anything?

The printing press made people more educated by giving them access to literature. LLMs are making people less educated by discouraging them from reading content written by people with actual thoughts.

The printing press made both excellent writing, such that of Primo Levi, and also garbage such as People Magazine accessible to the public. It's your choice on what you choose to read! The same applies to how you choose to use LLMs and how you apply them to your life!

How is AI "raping the planet?" How is AI leading to spending trillions of dollars on "toxic, unrecyclable equipment"?

Is that supposed to be ironic? TPU and NPUs are toxic and usually out of date when they reach the way into data centers. And even for GPUs they end up in landfills and aren't reused. And since this is companies moving money back and forth this is the goal in its own right. Playing this as long as possible is exactly what fills investor pockets.

His reaction seems based in fear.

I'd say so too, fear of where the world is heading with the above.

I don't get the rest of your post in this context however doesn't make sense at all.

Rob argues that the AI is a "monster trained on data produced in part by [his] own hands" — okay. Thousands of students at thousands of universities every day are learning by reading code on GitHub. Should they be providing attribution to every piece of code they've ever read whenever they go and produce something original?

LLMs have been shown countless times that they are incapable of originality. So have text to image GenAIs. So many examples, such as a wine glass filled to the brink that you cannot even hack around.

You'll learn that they aren't actually learning, by simply playing around with any of them for a bit.

Or in other words: Using lossy compression on an image, piece of music (or text for that matter) also doesn't mean your computer learned the essence even when these tools use statistical methods to simplify. Even scanners do that stuff. Same if you from scratch reproduce a logo from a bitmap into a vector graphic it doesn't mean you can just skip attribution. It's just not how this works. And just using a giant data center to do the same large scale also doesn't give you that.

AI is here to stay, just like the printing press or the industrial revolution.

You realize that not even Sam Altman seems to be so sure about this. At least that's what internal emails and preparing/begging for a government bailout seem to suggest.

But then... you are right in a way. AI is going to stay. It's been a thing for half a century. That isn't going to magically disappear. That said there seem to be fewer and fewer printing presses though.

However comparing the current hype around a couple of companies moving money in circles and people having fun with chatbots. That's here to stay, just like blockchains and NFTs were here to stay, Oh and the Metaverse, and wasn't Second Life the new internet,

And yes, AI has big use, but it did have great use ten and twenty years ago. I remember pattern recognition being utilized to count bacteria in labs. With the hype a lot of institutions have an easy way to fish for funding when they do literally the same thing having a machine count stuff. Same thing as OCR.

Somewhere along the line OpenAI jumped in and realized that Chat Bots (which anyways were a bit of a hype, being slammed onto so many website) with ANNs make people excited. I mean actually the question that was interesting was whether you could jump huge volumes of text and images and it would do cool stuff and the hardware finally was there. But oversimplifying. Just like a couple of years ago people realized combining Hashcash and endless logs (had a couple of names, hash logs, etc.) made people really excited.

And it's cool and exciting stuff. And fun stuff to work at for a lot of reasons.

But the stuff that is hyped now is extremely on the side of net negative. The fact that it hits at a point of pre-existing economical and societal volatility surely plays a part. And sure like so many unsharp statistical algorithms from what would now be called AI coloring the black and white images that a lot of digital cameras make to image editing where a lot of things that would now be called AI have been there for decades there are uses. But again, nothing really having anything to do with the current net bad hype.

Spam is also here to stay, and fun fact detecting it also has been using Neuronal Networks as well as other statistical methods. So are biological and nuclear weapons, and again people having stuff to do with it apologized, even during the hype. Ransomeware too. Just like the industrial revolution that - fun fact - fueled a significant drop in people's life spans until they realized it's a good idea to wash hands every now and then and that overall live spans rise a lot when you cut child mortality (so live spans of 0-2 years) when you don't have doctors run around proudly birthing babies with reddened hands from the previous baby. I bring this up because things hailed like the industrial revolution also are two-sided. The whole idea of "ah it's fine to burn the world and poison everyone in the name of progress" even stopped before the topic of climate change was on the table.

That means even if somehow the current hype ended up positive it very much is not now. And it is extremely far away from even helping the economy. Again, not even Sam Altman seems to think that. The industrial revolution did not have that issue at all. The productivity gains aren't realized anywhere. The idea that one has to just wait a bit longer has been told for years now all while progress outside of "Let's just burn more resources" and "Let's run the LLM write input for LLMs" is nearly non-existant. There is a good chance that what we'll get instead is being stuck on the level of 23 both technologically (Python, etc.) and societal/ethically, because companies end up feeding the output into the input and people like talking to those LLMs optimized on avoiding anything remotely controversial or novel.

How hard is to get that semiconductors is quite hard to recycle and they are toxic?