Controlling AI's Growing Energy Needs

> I'm going to keep trying to find uses for these garbage generators

Is it really worth your time? Sometimes the simplest explanation is the right one. If you've already mastered your craft past a certain point and care about the quality of your work, this will only drag you down.

I just wrote a multilinear algebra research paper using natural language to generate all the latex. This used to take me several hour of tedious mind numbing boiler plate. I finished in in 10 minutes today.

People who aren't on board with llms as the biggest driver of growth since internet search were the type of people who still thought the yellow pages were safe from Google in 2004.

Our engineers feed a LLM engineering plans as a quality control. Not as a replacement of any of the regular process, simply as an additional step. It discovered what would’ve been a fairly serious error in some component design for one of our energy plants that nobody else had caught. Similarity it “fact checked” one of our subcontractors on something with cable sizes that I honestly don’t understand. Finding a better option than what the subcontractor was suggesting. I say “fact checked” because you shouldn’t think of it as that since everything it outputs need to be actual fact checked.

I mostly view LLMs as advanced auto-complete for what I do and haven’t been very impressed by them, but some of the examples our engineers show us are wild.

For all tech people, LLM are an absolute boon: it brings productivity by 10x (at least!) onto the next level, esp. for develops.

I developed an financial servic application with 1MB+ source of code (not including 3rd party libs) within 3 month - thats more than 1 million keystrokes that created a system that works perfectly fine and produces exactly the results as specified. (for comparison: my masterthesis had around 110.000 letters, and i wrote it over a couple of weeks, while the app is slightly more complicated: its a complex system where each part relies on each another, while the masterthesis is just a dumb document in which i could write in theory everything and nobody would care/check if the descript is just bs)

This reminded me of a conversation I recently had at work. Someone was turning over a project to me so he was explaining the context and history of the technical design decisions they've gone through. One stage of the project was basically linting through someone else's source code, evaluating possible code branches.

"So, we want to know all the possible states this could go through. We gave that task to OpenAI with the following prompt."

He shows me a sample prompt they would've sent to OpenAI. It started with the premise "You are a senior Ruby developer with ten years experience. Read through this source file and give me a list of all the code paths which involves global variables".

"Really?" I said, amused. I knew ahead of time that this project they were turning over is LLM-heavy. I have misgivings about LLMs but I try to keep an open mind so it was really impressive that it was used at this stage. "How well did it work?"

"It didn't. Instead we parsed it to a syntax tree and just performed static code analysis."

(Usual disclaimer that details of the anecdote were changed/obfuscated but the spirit of the interaction has been preserved.)

Eventually.

And if not we accelerate climate change because we built fast solutions for the energy problems instead of sustainable ones.

The number one predictor of human wealth is energy consumption. Since the stagnation of the west in the 1980s we have seen the same stagnation in energy consumption there.

On the vendor's side, there's incentive to reduce the costs per token, but also an incentive to sell more tokens.

On the consumer's side, it's more about what the alternative would be, and the alternative is getting a human to do it.

Right now the general quality of the LLMs is good enough to be a tool, but not good enough to fully replace humans using it — I've seen GPT-4 called an intern, I'd agree having seen student code not too long before ChatGPT came out, and you don't want to let students do a whole project unsupervised either. Even though o1 is better, I think it's still at the "interesting, but not good enough to leave unsupervised" level*.

For tasks where the AI is competent enough to be interesting for whatever task, even if it took 10 kW to run the compute, at $0.1/kWh that's often cheaper than offshoring to a poor country.

As the world's entire electricity supply is 2 TW, which is 250 watts per capita, I think it's not at all implausible that enough people decide that the output of various AI (not just LLMs, also vision, audio, robots etc.) are valuable enough to demand enough of that electricity to make the prices rise.

* and I'm not just saying that because when it gets that good none of us will be employable any more

You make it sound like banks tolerate fraud. A financial cleanup is coming to the carbon markets, because that’s the only way serious money will flow.

Because the AI that can barely match the cognitive power of a human brain running on a cup of coffee and some calories doesn't sleep, eat, get sick, go on maternity/paternity leave, go on strike, want to be paid, want their pay to increase with the cost of living, or have pesky human rights that limit the value you can extract from them.

One is miniaturization and the other one is some sort of science breakthroughs to understand how the brain works.

You're not calculating the energy required for a human being to do computation. It requires food, clothing, recreation, temperature control, healthcare, and education and all sorts of things that require oodles more energy than a simply calculation of daily calorie consumption of a human being.

If it helps the AI doesnt 'match' humans at all, it does a completely different thing.

The computing machines that the Manhattan project used were famously as fast as the average secretary doing the same calculations. Yet the computer age still happened. Ask yourself why.

Didn’t know there was a human brain running on a cup of coffee and some calories answering hundreds of millions of cognitive API requests across the world each minute and continuing to grow.

Tbh AI is already at the stage that its more effective and insightful on most subjects than a huge chunk of humans and their coffee and calorie sippin brains.

I don’t think this sort of sadist hot take on nascent revolutionary tech that’s becoming better exponentially each year, is productive or fits into HN.