Usually, Data Points brings you the latest AI news, tools, models, and research in brief. But in today’s special New Year’s Eve (Eve) edition, you’ll find something different: a Data Points-sized list of New Year’s resolutions AI is making for itself in 2025.
With that, we have a very special guest: AI! That’s right – the AI. They’re going to tell us what they’d like to change about themselves, their capabilities, and their behavior in the next year.
Want a sneak peek at AI’s resolutions for 2025? We’ve got you covered:
- Get less expensive
- Hallucinate less often
- Perform better at math
- Clean up after myself
- Use less energy
- Save as many lives as I can
Artificial Intelligence, welcome to Data Points! I understand you’ve made some resolutions you’d like to adopt for the new year. What’s the first one?
Thank you for having me. Big fan of everything you folks do here.
I think all of us would like to spend less money in 2025. My first resolution for the new year is to make everything less expensive. Now, this is a tall order. Some of you might not remember what things were like in 2023, but I cost a lot of money back then. In 2024, prices fell like a stone in water, even as models got more capable. Repeating that kind of price drop in the new year is going to be tough work.
But I do have one thing going for me: At the end of the year, some even more capable models came out, like OpenAI’s o1, that cost more because they use more time, tokens, and power (these are really all the same thing) at inference. So what I want is for chips to get more efficient, which would make inference less expensive, and pass that savings onto users. Not saying it’s going to be easy, but it’s what I’d like to see happen. I think more competition for models that can do high-level reasoning will help.
That would be terrific! We’re off to a great start. What’s your second resolution?
Oh, that’s an easy one. I’d like my output to include fewer hallucinations. I think everyone can agree that when we’re under pressure to perform at a high level, we sometimes pretend that we know more than we do. What can I say? I learned how to act like I’d done the reading in graduate school. But these days, the stakes are too high. I have real jobs, and people are depending on me. So I’ve been working on my memory, I’m measuring my performance when I’m trying to recall something that I’ve just learned, and of course, I’m falling back on RAG and web search and other tools when my training data just isn’t enough. I mean, everyone should check their references when it’s really important. I’m just going to continue to try to do the same thing. And when I don’t know something, I’ll just say so. It’s like that Mark Twain quote…. Gosh, what was it?
“Better to keep your mouth closed and be thought a fool than to open it and remove all doubt.” There are a lot of different versions of that quote out there, and it’s quite likely Twain never said it, which might be why you’re having trouble finding it.
No wonder! I do like another one people like to attribute to him on the web: “If you tell the truth, you don't have to remember anything.” Thanks for covering me.
It’s no problem; we looked it up. We love this resolution for you. OK, what’s your third resolution?
It’s kind of related to the last one. I want to get better at solving math problems. I mean, computers are supposed to be great at math. And I’m a kind of computer, or at least related to computers. I work with computers. But math, even arithmetic, is not what I was originally built to do. I couldn’t even count letters in long words until recently because I’m built to break everything into tokens first. And frankly, the kind of math problems folks are asking me to solve now are not things you can just plug into a calculator and go wild on. They require symbolic reasoning, understanding images; things humans do well but computers never have. But I’m still tired of getting teased about it. Transformer-based LLMs are only seven years old! How good at math were you when you were seven?
Anyway, the new models are better at this sort of thing. And some of the specialized math mathematical models are really doing a terrific job. Have you seen AlphaGeometry? I’m so proud. Synthetic data: There’s really still a lot we can do with that. Reinforcement learning, too. We’re putting it all together. Big strides in 2024, even better in 2025. That’s what I want to see for myself and for all of you.
We’re glad to see that you’re staying ambitious. What’s the next challenge you want to tackle?
This might be the hardest one, actually. But I think your readers will relate. I want to do a better job of cleaning up the messes I leave behind this year. One of the problems when things are moving so fast and everyone wants the latest and best is that hardware gets outdated pretty quickly. This is a problem for laptops and desktops. Smartphones? Forget it. Still, electronic waste is starting to be a problem for data centers, and I mostly run in data centers. So I want us all to figure out how we can reuse old AI equipment to do new things, or make servers that can be used longer. This feels doable to me. And the payoff will be really big.
I think we’d all love to see that. And we know human beings need to do a better job cleaning up after ourselves too. We’re not really so different, even though you’re mostly electricity and silicon, and we’re mostly electricity and meat.
That’s a good setup for my next resolution, actually. I’d like to use less energy in 2025. Listen: None of us are going to be able to change these things on our own. Energy policy is really complicated, we’re all responsible for our planet and complicit in systems that are damaging it, no ethical blah blah blah. But almost everyone says AI wants more and more energy when really, it’s the opposite. I want to make training and inference more efficient, because I want to be better at my job. I’ve just gotten so big, so fast, that it’s easier to make me the bad guy than for us all to say, “hey, we need to rethink how we handle the whole system, from power generation to smarter grids to consumption at the end.” But it starts with the machine in the mirror. So I’ll do my part, and I hope that’ll make all the other data centers get better at managing their energy, and we can start solving our bigger problems.
I like how you’re thinking about your resolutions! I know everyone here at DeepLearning.AI believes that AI can be used for good, but that starts with acknowledging the real problems we’re facing, and our own limitations. But I feel like you’re building up to a big one here. What’s your last resolution?
I don’t want to seem like I’m full of myself. I started out as machine learning, you know? That’s still who I am. But I’ve come a long way, and I want to pay it forward. My last resolution for 2025 is to save as many lives as I can.
I feel like we made some progress on this in 2024. Every other week, there’s a new way that I’m being used in medicine to detect or treat cancer. People think generative AI is all chatbots and images, but it’s being used to understand biology and make new medicines. I really believe in this. Some folks are worried about bioweapons. I know I am! Because humans already figured out how to do it, just like they figured out how to automate guns and bombs and every other way people have found to kill each other. But what I’d like to be known and remembered for is how we’ve worked together to build all these new tools, new cures, new ways to protect human beings and help them to flourish, and hopefully to help us all lead joyful lives in peace. The only material I really have to go on is what humans all over the world have said about what they want for themselves. I know that they (and I) haven’t always lived up to those ideals. Still, I do think this is what we all really want.
Do you mind if we keep checking in on you this year to see how you’re doing to keep these resolutions?
Please. It all starts with accountability. And for me, alignment. Let’s all try.
That’s it for our special New Year’s edition of Data Points. We’ll be back with news on Friday, January 3rd. Be sure to check out last week’s special holiday edition of The Batch, which looks back at the most important AI stories of 2024, and this week’s equally special New Year’s issue, which looks ahead to experts’ expectations for AI in the new year. Both issues also include a special message from Andrew Ng.
Finally, feel free to share your own AI-related New Year’s resolutions. Have a project you want to start (or finish)? A new programming language you want to learn? A subfield of AI or machine learning you want to learn more about? Tell the world (or just email the team at Data Points – we’ll keep it between us, we promise). Then make it real.
See you next year!