It’s hard to tell with the these new trillion dollar investments, but the power cost of running a text query is pretty cheap, and according to Google, has been plummeting.
It’s complicated because there’s expensive training, and the volume of requests is shooting up, and people are increasingly using AI for things like generating HD-Video for a stupid meme, which uses so much energy. And we now have AI agents, that can just run continuously for hours on end talking to themselves, writing an application or solving a math problem or beating pokemon.
You can see the evolving tradeoff between better and cheaper over here.
ANYWAY. I would also posit that while AI might run on a thousand megawatts, remember that our current world rulers run on a thousand megayacths.
And yet it’s the main focus of what’s in the main media. I know automated driving is also out there, but no one really looks at that as NN modeling like LLMs
Did a google search. Looking for a single number. At the top there was some AI response, which I guess was intended to help me.
I went into the actual files anyway. The AI number was wrong. Not terribly wrong (off by .2), but I would prefer terribly wrong as I would know immediately. I don’t get people using AI, as I have yet seen anything I can trust to be accurate, without redoing the work myself
FWIW, in the past few months, I’ve been using ChatGPT to bypass my limited proficiency with R and Python to (finally!) migrate some huge/ugly Excel and Access files into more appropriate (and efficient) tools.
I obviously still have to double-check the work, and I do tend to try phrasing prompts 2-3 different ways just to validate that the code I’m being given is good, but in the past year ChatGPT and my prompting skill have finally gotten good enough that “I don’t have the slack in my schedule to deal with the learning curve” is no longer a valid excuse for not making changes that I should have made years ago with some of these files.
I already have one project that I’m penciling in for probably next summer, using public/industry data, that I might try using AI for, rather than begging for a team of minions to wrangle into doing.
I’ve done something similar. Needed to code something in a language whose syntax I don’t know, got Gemini to create a version for me with commented code that was wrong but at least had the syntax I wanted so I could go through and fix the logic.
you can’t trust ai to be accurate. you can trust it to say do a bunch of mundane work like you’d have an actuarial student do, then you need to check everything before presenting to the board.
Im building a compliance document comparing suspected problems against a bunch of regs and guidelines. I could read all the guidelines and find problems. but I didn’t.
I threw the problem onto an ai, gave it the outline of the problem and the regs I wanted to check. about 3 prompts later I had a 10 page document with bullet points pointing to specific parts of the regulations.
without ai, this is a week’s long job for me. with ai, it’s maybe a day as I go through the points, look up the exact section of the regs and review the ai’s conclusion.
ai is great for grunt work. its great for stuff that you’re already reasonably expert on. I wouldn’t trust it to teach me anything new. but ‘summarize this and create a powerpoint’ doesn’t make a media story.
This. The “search” is constrained to the possible moves available . . . with the main parameter being how many “moves ahead” to look. And there’s an optimization parameter that is often used that is often tied to some sort of point values assigned to your pieces and your opponent pieces that is used to determine the best result for you after those sets of moves.
And these point values are also used to assess (or constrain) the likely response of the opponent.
Search isn’t a requirement at all for expert level chess, which would beat 90% of chess players who have been playing for years. The net outputs what it calculates as the best move directly. Search will improve on that, but it isn’t a necessity.
I’m working on two things right now, or at least my dev is.
He’s working on the following:
all my calls; generate a transcript
attach the transcript to client record (for compliance)
generate a summary of the call and attach to client record (for me to review next time I contact them)
generate a list of tasks that were discussed in the call and inject into my calendar.
None of this is particularly complex, but it is cool and in my industry, groundbreaking. And pretty cool that even like a couple of years ago, mostly unimaginable.
Second thing is, and this is just for show, I’m using AI to quickly generate a complete website, from start to publication.
I think some of these ai researches get really far out over their skiis on some of this stuff.
It is a matter of fact that these models statistically predict future words.
To say they can reason is to equate reason with predicting the most probable word.
What are the implications of that claim for the philosophy of language, of science, epistemology, etc? Does it conflict with anything in those fields? I never hear anybody talk about that.
We cannot statistically predict the next word. But we can reason too, i guess?
Now we are extending this claim to consciousness. Or starting to, anyway.
The article actually made clear that no one is claiming that the big chatbots have attained consciousness, and the main focus is that this behavior is an interesting one to explore as the researchers try to understand the details in how/why they work.
However, the article also pointed out that this phenomenon does highlight a philosophical question: What is consciousness? That’s a concept that we might need a clear definition for in coming years as “AI” advances.
