2084: Chatting with the future: Redux

Tried out Chat GPT3 - it's absurd

Like a million other people, lately I’ve been checking out ChatGPT, OpenAI’s new text based conversation focused chatbot that they’ve released. And it’s absurd. Some of the responses are so good its magic. I’ve gotten it to write an essay comparing Pierre and Andrei from War And Peace, I’ve gotten it to partially do my project on Frobenius’s Integrability Theorem(in correct LaTeX no less), and so far I’ve tried it on a few mathematics homework problems, and the results were, if not always correct, at least close enough that they usually gave a hint to the solution. In a few months or so, they might fix that hiccup and then it’ll just straight up be better than most undergraduates - much less lazy, that’s for one.

What’s nicest about it is how smooth the whole conversation flows. ChatGPT has 8000 “tokens”, which are how many words it can process at a time, which means that it has one of the largest working memories of any large language model out there, and it shows in how much context it can remember - sometimes it’s spookily real. It’s mostly an evolution of GPT3, but it shows how much better the state of the art has gotten. It’s made me especially excited for GPT-4.

How it’s been modified is also rather interesting. Essentially they took GPT-3 and made it more “helpful” for people - able to more easily respond to requests with what people want. They did that by training another model using a bunch of human testers who evaluated what people liked and didn’t, and training GPT3 based on this models outputs on what it output. Here’s OpenAI’s explanation for those interested:

But what’s been the most interesting has been seeing what other people have been coming up with for ChatGPT. The best one I’ve seen so far is the virtual linux terminal, where someone got it to simulate a linux terminal and execute commands with real looking responses in it - it even managed to do an API request to its own website to have a simulated version of itself reply. Some other guy then got it to both write and execute a python program for tic tac toe, which is absurd - it’s hallucinating functioning code into existence. This is most interesting, because it indicates a possible future direction for GPT and programming in general - programming general models, instead of writing code. In the future, we might all be fiddling with the AI models to get it to output the data we want, rather than doing code. Seeing as most of programming is just data processing in various forms, a large heavy duty model could replace a lot of code with only one specialized model.

Since these neural networks models are computationally equivalent to our spiking neural networks that we call our brain, this raises an interesting odd point, in that we might come full circle to having “brains” take care of most of our work, instead of automating it away. Hyperintelligent, massive, all-knowing brains that never tire or take a break, but fundamentally brains all the same. Given the last post, where we looked at neural networks that can decode images and text, it might be possible in the future to stream information like that back - imagine knowing anything you could ever want, instantly at a thought.

And ChatGPT isn’t even that dumb.

It did better than most on the SATs, being in the 52nd percentile, and it got 83 on an IQ test. As the SAT is, to some extent, a test of intelligence, this means that ChatGPT is about as smart as 30% of humanity at least. Given the breadth of its knowledge, being as it is, knowledgeable about all of coding, all of Linux terminal management, obscure Russian authors, fluent in German, French and Japan and many more languages(given that it’s been trained on wikipedia and random webpages and reddit, that’s not surprising), solving crosswords, recipes for souffles, pizza and methamphetamine, knowing how to get a date and hotwire a car, give good advice on how to run a party, write letters to your kids telling them there is no Santa in a kind and empathetic way, there seems to be no limit to its capabilities. It can even help debug and explain code, and write it for you too.

Yes, it fibs sometimes, but that makes it more human, not less - who hasn’t ever told a few white lies, or made up some half remembered facts on the spot?

Now ChatGPT only works with text at the moment, but remember GATO? This might be only the start. I think we’ll see OpenAI come out with newer and newer models soon that can do a lot more than text - audio, and video generation can’t be too far away. If OpenAI ever releases the models behind the magic, if even to verified and vetted companies, expect to see the amount of applications explode - I don’t think that we’ll be able to explore the edges of the AI model’s internal mental universe anytime soon.

It makes me quite excited for the future. Programming is going to be so much more exciting now that we’re started to be free from the bounds of simple data processing. Expect to see ever more applications built using stringing together prompts - it might be a brave new future in 2084. I can’t wait.