ChatGPT Is at the Peak of the Hype Cycle
danschnau.com
danschnau.com
ChatGPT has it's shortcomings, but I've found myself returning to it over and over again. Then consider what this technology will look like in a few years time and the conclusion is that, imo, the hype is justified.
I'm becoming more and more disillusioned. I was sceptical at first, then thoroughly impressed by all the blog posts, but actually using it is extremely frustrating. It's really good at composing answers that look plausible, but it just has no idea what you want.
The only thing it is truly good at is composing unnecessarily verbose responses to spam emails.
I've found its knowledge to be flawed if you poke enough into it though that feels more like it doesn't "understand" the facts that are embedded into the words.
However, using it to do the classification natural language, extract certain pieces of information from the query, and then convert the structured data back to natural language it shows itself to be very capable and able to be "better sounding" than rules based summaries of that data.
I've also found it to be quite capable at rewriting poorly written questions (e.g. Stack overflow questions - yea, I know, banned and all) to remove extraneous parts ("I'm a noob", "Thanks in advance"), fix grammar, and change the tone to be more professional... and creating a good summary title.
I have definitely had it get stuff just flat out wrong, but it’s easy enough to pick it out and I degenerate to old practice of searching and filtering in those cases.
As has been posted about here many times in different ways, as they graft IR and reasoning systems onto chatgpt I expect these problems to disappear. I find it amazing how dismissive folks are of a tech that is barely a few months old. It’s like a whole generation of nerds have lost the capacity to dream.
It results in much more relevant and grounded answers(well... grounded to whatever you documents claim, for better or worse). If you want to try it, there's a demo at fragen.co.uk
> Nobody has figured out how to make money from AI/ML other than by selling you a pile of compute and storage for your AI/ML misadventures.
I work at a company where we process vast amounts of retailer advertising data, and one of the things we need to do is tag images and extract data from deals. Although humans are still involved in this, at this point they are mostly checking the work of our ML models. This dramatically increases the speed with which we are able to do this - and this speed is one of the things that makes us attractive to our clients.
Though, I wouldn't be surprised that if overall, people have spent more time on compute than have made from the models. There are companies that treat data science properly and are doing it well - but in the previous financial era people were inclined to throw money at things without clear criteria for success.
For the tweet, note the hyperbole and mocking of AWS more than the mocking of ML. The entire chain of tweets is mocking the cloud with hyperbole and a grain of truth.
For me, it has turned out to be a massive productivity enhancer, when used to augment an existing set of skills rather than replace it. I think of it as a "cognitive workhorse" that can do the crappy parts of generating lots of stuff that I can go back and edit later. I much prefer this method over having to do the initial generation myself. Its worked great for me for both writing technical(ish) docs and for generating code.
Much of the time though I feel like it only works as well as it does for me because I'm not asking it to do anything I couldn't do myself (albeit with more mental effort), which means I am able to immediately spot any issues and ask it to improve or expand on its answer.
The nice thing about generating code too is that you'll know pretty quick if it works or not. By the time its actually working, I've gone through and checked each line and made sure all the tests pass and test the correct things.
ChatGPT already does well when essentially regurgitating factually correct information, but the obvious weakness of an LLM is that when combining sources its just combining them linguistically so it produces fine sounding streams of often factually incorrect bullshit.
It's hard to imagine how GPT-4 will be much different unless there have been some fundamental extensions to the model architecture, or changes to the training regime. Presumably it'll do better in "search engine mode" where people are hoping it'll "understand" the subtleties of their questions in selecting what to generate, but hard to see how it'll avoid the "bullshit chained deduction" problem unless they've added a way for it to learn from it's own mistakes.
In terms of perception/disappointment, I guess it depends on individual expectations. If you want GPT-4 to be a better "search engine" than GPT-3, then you may be happy. If you expect GPT-4 to be "more intelligent" than GPT-3, then I'm guessing not so much. Let's see!
I've a hunch it may do well as the basis of an improved Copilot since this seems to be an easier subdomain.
So, perhaps the scale upgrade from GPT-3 to GPT-4 will improve it's world model enough that it does a better job of combining sources using a consistent bias/POV rather then happily combining true and false sources as it currently does.
Any program that only gets things right by being lucky enough to come up with the right "search key" for a correct answer isn't AI in my book. YMMV.