Comment by mintplant | Hacker News Reader

Comment by mintplant | Hacker News Reader

Hi HN! I've been playing around with ChatGPT a bunch since it came out. This experiment has a little bit of a backstory. Some friends and I were out at a pho restaurant; one of us put the whole bill on his card, so the rest of us needed to figure out how much to Venmo him. We were talking about how many bill-splitting apps there are, and I made a joke about doing it with ChatGPT. Then I actually tried it out.

I OCR'd the text with Google Lens, described who had what, and after a bit of prompt engineering (e.g., adding "Be sure to get your math correct" to make the AI's arithmetic check out, and convincing the AI to split shared items evenly), it totally worked: https://gist.github.com/spinda/967322dda1c04d9864f3efd45addc...

Then I started experimenting with describing a hypothetical check-splitting app to the AI, and asking it to feed me JSON commands to update the UI in response to messages from me telling it what the user was doing. The results were promising! And then the similarity to the Redux data loop jumped out, and I built this generic plugin to wire ChatGPT up to apps for real.

Cool project, but it reminds me of one of the golden rules of AI “if you can solve it with a simple rules based system, you should”

Having to tell chatGPT to make sure it gets it’s math right is not confidence inspiring.

I see your point, but I think that's kinda neat, personally. Fixing a bug by asking nicely for the bug not to happen. The pinnacle of declarativeness.

It makes me think of INTERCAL, that you need to prepend commands with "please" some of the time, but not so often that it chastises you for being overly polite.

Wild times to be alive and a software engineer.

Wonderful example @mintplant.

It also did not work perfectly for me, running GP's prompt. ChatGPT left off the shared "Egg Rolls" entry entirely :(

Can I ask why it's effective to tell it to get it's math correct?

Two sub-questions: 1) I thought LLMs worked by predicting the most likely next token. Why would telling it to check it's math actually make it's predictions more accurate?

2) Why would asking it to check it's math change the results it produces? Shouldn't it just do the same thing a second time and produce answers that are just as (in)accurate as it did the first time?

My intuition is: the AI is trying to predict a likely response to the input query given what it's seen in the dataset it's been trained on. That dataset will include instances where math has been done incorrectly, so the AI producing a response with incorrect math isn't inherently "wrong" from its perspective. However, queries containing reminders to be careful with the math are more likely to be followed by responses with correct numbers.

So it’s pretending to be an average human, and the average human gets it wrong a lot? I wonder if you can improve coding results by telling it to act as an expert programmer?

It's not "pretending". As the comment said, the dataset has something like "paths" and that prompt narrows it down to those that (hopefully) are related to "getting the math right".

I noticed as well that by default the output is effectively average. The right context bias does help with getting better quality output.

You can also have it check it's self after the fact. which is the advantage of this over GPT3.

Imagine you are navigating a tiny robot on mars, or let's say a giant land region with hills and valleys.

Saying it "split it evenly" and giving it numbers is something like saying "find me the hill next to this valley". This will lead you to the valley on the left, then to the tiny hill right behind it. The hill right behind it in this example is where all the weights about splitting stuff evenly are hanging around.

Giving it "split evenly and make sure the math is correct" is like saying "find me the the hill next to this valley, but make sure it's the tallest". So it will lead you to the same valley, but then to a different hill, because of adding "tallest" it will lead you to the intersection of where all the weights about splitting stuff evenly and correct math were found.

I.e. there is other stuff you can engineer about - saying stuff like "Write the result step by step and check for correctness." afterwards will navigate it to the same hill but then towards the "all math tests are here" hill which is where step by step calculations weights are and it will find it's new home there, giving you a more detailed output and more chances to be right since the prediction is easier - basically it "splits" the task into smaller, easier to predict chunks, like you split ugly code into small testable functions to understand it.

I told it to sound 100% more human when writing a blog post and it moves the score with the AI detectors.

Then I told it to sound even more human and it does even better.

Yes, it's weird this happens. It's all because this is a conditional model. So it can be influenced by anything in the prompt, including the persona it is assuming.

But it is not deterministic. For each token it picks at random, from a probability distribution generated by the neural network. You can make it deterministic if you set temperature=0. In general it is stochastic, if you sample again with the same prompt you get a different answer.

A calculator is a "simple rules based system" but I still do mental arithmetic.

> Having to tell chatGPT to make sure it gets it’s math right is not confidence inspiring.

We don't compute in our heads, we use calculators. chatGPT could have a calculator too. Why kill a fly with a sledge hammer?

Exactly, ChatGPT has a calculator to compute the math, the calculator with regard to one query may be correct, but to a query specifying that the calculator must be right the calculator to that query will most likely be correct.

It's just a regex. When the model outputs "[calculate]12+34=[/calculate]" you append the result "46" and continue with the model.

In some papers they even use named symbols. So you could have "[code]width=10; height=20[/code]. What is the area? ... The area is [code]width * height[/code].

Pretty sure you've invented something here. A coolest form of copilot. A new way of programming. This is really lispy, this is really cool. Love it 10 times.

Edit: you ask the AI to run the program in its head, its a new form of computer actually afaik.

Next up: I asked the AI to pretend to be the backend, but also front end! Apps spawn into existence to be used once and discarded.

Wow that's an impressive backstory. Probably one of the more useful things I've seen come out of chatGPT

This seems like it could be perfect for making prototypes. "Pretend that you are a backend for a text messaging app that supports group chats. The data model should look something like this: ..."

Why use chatGPT instead of using the GPT-3 api directly?

For the type of queries you are doing (sending whole context), the output is comparable (and just as wrong) between chatGPT and GPT-3.

ChatGPT is currently free, GPT-3 is not

Afaik there's no way to build upon your prompt to have it change the output in specific ways in GPT-3?

You feed it back in as the final tokens of the next prompt.

So, you needed a multiple step process of technology transfer to an ai or an app to figure out how to split the check?

I find this fascinating but not terribly impressive. Its fascinating that such a rudimentary skill prompted you to dive down this rabbit hole. The unimpressive part is that it's a rudimentary skill yet your over engineered solution still only required rudimentary skills.

It was more of a fun exercise, and an excuse to play with some cool tech! The part I was most impressed with was that ChatGPT could make sense of the mashed-up OCR text from the receipt—I can see that being super handy for speeding up expense reporting data entry. Everything else was extra.

It's not, like, a practical way to solve problems, but it's interesting in its own right.

Reply on news.ycombinator.com