Over the past year, a few friends and colleagues have asked me about my day-to-day use of AI via emails, texts or the infamous X DMs. I realised that as this technology advances, so do the questions on how to use it meaningfully and efficiently. There are many resources out there, and it's extremely hard to keep up, even for someone who prides themselves on being an AI whisperer and, on certain days, a tech bro in the Bay.
Let's start with a philosophical quip. I've loved technology since I got my first computer, "Lappy" (so very creative!), which opened my world to the world and enabled me, in many ways, to do all the funky and fun things that bring me joy and curiosity. However, as much as I love this tech, I worry quite a bit about how it's created and – closer to home – how it's used. AI Scientists, AI Astronomers, and AI Chemists are superb intellectual feats, but we must proceed with care. Human dignity must guide our progress, not fade in its wake.
I want to live in a future where the "machines of loving grace" empower us in our journeys of discovery. To this end, I use AI in the following ways - many are the result of trial and error and lots and lots of frustration. My day-to-day involves Claude Sonnet 3.5 with the subscription. Claude has been my ride-and-type for more than a year and it's a legit strong model. I primarily use it for two tasks: code verification, where I give it some code I wrote and ask it to provide feedback and improve on it, and code completion, where I have an idea of where I wanna get, but not sure how to get there. Sometimes, I provide additional context via the brilliant repomix tool that packs GitHub repos. When it comes to code for Astronomy, stay tuned for the brilliant work that Nolan Koblischke is building with AstroCoder.
But what about the prompt, you might ask? Sonnet is pretty good at figuring out what you need it to do, but it helps to guide it via clear and (semi-) structured prompts. Recently, I asked Sonnet to help me organise a repo (because I am a chaotic human and I struggle with these things!), and the prompt I gave it was along the lines of: "I need a repo for this project, where I try to do X. I will need to save some artifacts, like images and models, and I have lots of data from multiple sources. I also use notebooks a lot, but I'd like to build this repo so that other people can use it down the line. Please provide me with a useful structure." The key here is not to overthink the prompts too much, but if you want to be extra precise, you can ask the model to improve your prompt and use that instead.
For literature search and synthesis, I like to use Gemini Pro 1.5 with Deep Research. You do need another subscription, unfortunately, but if you're an astronomer, the folks at UniverseTBD built this amazing tool called Pathfinder in collaboration with wonderful astronomers and human-computer interaction experts at Johns Hopkins University. Building tools for people and with people is a big part of our ethos, and hope the community likes it. Big shoutout to the legend Kartheik Iyer for opening this tool to the community!
I've recently started playing with o1-pro for hypothesis generation, but that's an experimental feature. If you're using the free version ChatGPT-4o, a good approach is to ask the model to extract the reasoning steps in some great ideas that the human expert approves. Once that's done, you can use the model to help you craft your idea further by referring back to this "blueprint of greatness". Again, experimental features, and we're actively thinking of how to expand this to the community of astronomy - free and useful. We're also cooking far out ideas, but we wanna do it right, slow and steady and involving our communities in the right way.
Oh, before I forget, if you wanna build an LLM, here are some amazing educational resources:
- Stanford's CS229 provides this amazing lecture on LLMs: https://www.youtube.com/watch?v=9vM4p9NN0Ts
- Andrej Karpathy on how to build an LLM: https://www.youtube.com/watch?v=kCc8FmEb1nY
- Neel Nanda's videos: https://www.youtube.com/watch?v=bOYE6E8JrtU and https://www.youtube.com/watch?v=dsjUDacBw8o&list=PL7m7hLIqA0hoIUPhC26ASCVs_VrqcDpAz&index=3
Neel also has the notebooks and solutions highlighted in the video description.
I will keep updating this blogpost and make it better and more useful. Feel free to send me feedback and ideas at iciuca@universetbd.org. Most importantly, please share your ideas of what would be useful for you. Let's tune on!
Comments