I use GPT-4 for so much stuff. I was pondering the other day how much I’d be willing to pay should they decide to increase the price and realised it was almost order of magnitude more.
I love new technology and tried really hard to like crypto but never understood what I could actually use it for. Tried making some web3 apps etc. just seemed like it wasn’t really any better than a mysql db.
But I find new things to do with AI daily. I save myself hours and hours of time in work and personal life. If you can’t find something useful to use it for, I’d say try harder.
Also, it’s specifically GPT-4 that is amazing compared to anything else I’ve used.
If I had better privacy I would use this for even more tasks. I’d love to train something like this on reminding and prioritizing tasks for me based on my actual behaviour vs actions, for example. There’s enormous potential that previously wouldn’t have been possible with my skills as an individual (and evidently, within the capacity of entire companies. That product doesn’t exist in a form that I’d care to use yet).
Keep in mind that the valuation is based on speculation, too. It isn’t saying GPT-4 is worth 80B. It’s also saying there’s confidence that GPT-5 and successors will unveil meaningful possibilities with real value. I don’t think that’s unreasonable. I don’t think it maps to 80B, but that’s not my business. I just don’t think it’s crazy to see value here.
How so? I find it to be most useful in that exact area. Taking your existing code and turning it into another language, implementing something with a library you want to try but can't be arsed to read the docs for, making a first draft of an implementation you're not sure how to begin, or just researching what exists in general. Learning new things has never been this easy.
Having said that, I still like to dig around the internet and go on tangents. Having fast answers is great in some contexts, but I still value going on deep dives and collecting all kinds of unexpected (even if irrelevant) bits of knowledge. That’s actually preferably, but only possible when I’ve got a lot of time on my hands. Which isn’t very often.
I created the serialized state as JSON and planned to work backward from there. This would allow me to describe the capabilities of the machines and how their features would work in practice. I was able to feed that serialized machine into GPT and ask how to implement various features, and it was remarkably effective.
Most of it was trivial to implement, but there was the odd bit where GPT saw various connections between features which I was missing. It also made some mistakes of course, but generally did a good job of keeping the ball rolling.
I could have done this alone without a doubt. It was so smooth and easy with GPT though that I finished fairly quickly and really enjoyed it. I generally got what I wanted out of the project without going down rabbit holes or dumping more time than I could afford.
Otherwise it helped me figure out how to arbitrarily generate a jigsaw puzzle based on an image and a few parameters in three.js. That was a bit out of my comfort zone but all of its suggestions and solutions were sufficient to keep me moving and solving the problems. I’m not sure I would have finished otherwise. My goal is to make a multiplayer puzzle game with the intent of demonstrating how to manage moves, interactions, and play states over multiple clients and servers. That part is east for me, but creating a usable demo in three was not. Yet relatively attainable in a reasonable timeframe with GPT.
It could be that I learn slowly. I’m over a decade into my career and still find myself poring over documentation of things I suspect I should know by now. Like the math for creating a jigsaw puzzle. It’s basic stuff but I totally drew blanks at first. If that’s the case, perhaps it’s why GPT is so useful to me. It breaks down the odd barrier that might otherwise become a rabbit hole or a major time sink.
In fact given what we know now, it likely wasn't nearly as good as it looked. So was hard to say at that point if this whole language model thing will pan out or become yet another useless learning approach that's unfeasible to apply to anything.