Love this writeup. I haven't had time to get my hands dirty with Code Interpreter yet, but I've got a few things I want to try.
I've seen folks mention that GPT-4 + Code Interpreter 1) is more advanced than GPT-4 (some are calling is GPT-4.5) and 2) has a bigger context window. Do you agree/disagree with that analysis?
I'd say both of those are true but not because of the underlying language model, as that's very much still "just" GPT-4. But:
1. The interactions are more advanced because the Code Interpreter can do far more sophisticated analysis using Python for many purposes in an intelligent way. It can also self-correct (as I indicate in my post) by catching and reviewing any Python errors on its own. As far as I can tell, this is the closest we've been to giving an LLM a serious toolkit to call on thus far.
2. This also comes from the Code Interpreter part rather than the context window for the chat itself - as far as I can tell, that part's unchanged. But because you can upload PDFs of up to 100MB (some articles have indicated it goes even beyond that), the context window is essentially equal to whatever document(s) you upload.
In one of the video showcases by "AI Explained" that I've linked, he feeds the Code Interpreter the entirety of Anna Karenina by Leo Tolstoy (340K words) and has it accurately find all instances of the word "England" and correctly analyse the tone in which it is mentioned.
I haven't personally tried doing anything major just yet. Just goofed around a bit. But I'll try to do some more thorough tests of its capabilities at a later stage. So far, it looks very promising indeed.
There are so many potential use cases that it's best to define them by what real-world purpose they serve. I didn't have any immediate productivity-related tasks, so I went for the fun and somewhat useless tests. But if you're serious about crunching data, the Code Interpreter is apparently a massively helpful assistant!
Ohhhhhh yeah. I am really looking forward to making sense of the data.
Some folks hoard clothing or magazines or whatever. I hoard data, and we have a lot of it. I feel that the data is valuable to us, and there are actionable conclusions I can't currently see.
I'm still hesitant for trust reasons to put a proprietary spreadsheet with our actual data in it into ChatGPT. In fairness, Google probably has access to a lot of the info, and it does seem inevitable that whatever Code Interpreter can do today, Sheets will do tomorrow, so maybe just waiting a bit longer is my best play.
Meantime, at least I can do as you're doing, and just have some fun! I did end up expanding an image earlier with ChatGPT, and that was a first. I give it a D right now, but I can see clearly how it's going to improve over time (and how I can learn its limitations, just as I have for ole GPT).
Love this writeup. I haven't had time to get my hands dirty with Code Interpreter yet, but I've got a few things I want to try.
I've seen folks mention that GPT-4 + Code Interpreter 1) is more advanced than GPT-4 (some are calling is GPT-4.5) and 2) has a bigger context window. Do you agree/disagree with that analysis?
I'd say both of those are true but not because of the underlying language model, as that's very much still "just" GPT-4. But:
1. The interactions are more advanced because the Code Interpreter can do far more sophisticated analysis using Python for many purposes in an intelligent way. It can also self-correct (as I indicate in my post) by catching and reviewing any Python errors on its own. As far as I can tell, this is the closest we've been to giving an LLM a serious toolkit to call on thus far.
2. This also comes from the Code Interpreter part rather than the context window for the chat itself - as far as I can tell, that part's unchanged. But because you can upload PDFs of up to 100MB (some articles have indicated it goes even beyond that), the context window is essentially equal to whatever document(s) you upload.
In one of the video showcases by "AI Explained" that I've linked, he feeds the Code Interpreter the entirety of Anna Karenina by Leo Tolstoy (340K words) and has it accurately find all instances of the word "England" and correctly analyse the tone in which it is mentioned.
I haven't personally tried doing anything major just yet. Just goofed around a bit. But I'll try to do some more thorough tests of its capabilities at a later stage. So far, it looks very promising indeed.
Great stuff, Daniel. I really need to get my hands dirty on this. I've played with a spreadsheet and PDF so far.
Any thoughts on a good next use case (from your list, or otherwise) I should try next?
There are so many potential use cases that it's best to define them by what real-world purpose they serve. I didn't have any immediate productivity-related tasks, so I went for the fun and somewhat useless tests. But if you're serious about crunching data, the Code Interpreter is apparently a massively helpful assistant!
Ohhhhhh yeah. I am really looking forward to making sense of the data.
Some folks hoard clothing or magazines or whatever. I hoard data, and we have a lot of it. I feel that the data is valuable to us, and there are actionable conclusions I can't currently see.
I'm still hesitant for trust reasons to put a proprietary spreadsheet with our actual data in it into ChatGPT. In fairness, Google probably has access to a lot of the info, and it does seem inevitable that whatever Code Interpreter can do today, Sheets will do tomorrow, so maybe just waiting a bit longer is my best play.
Meantime, at least I can do as you're doing, and just have some fun! I did end up expanding an image earlier with ChatGPT, and that was a first. I give it a D right now, but I can see clearly how it's going to improve over time (and how I can learn its limitations, just as I have for ole GPT).
Appreciate the dialogue!