Ask HN: What is a specific use of GPT-4 that you think is remarkable?
I see plenty of hot takes on Twitter about GPT-4 and how its going to kill jobs and give us super powers. What I am not seeing are practical and specific applications outside of "Write code" or "Summarize articles". What examples do you have where someone used GPT-4/AI tech that had a high degree of utility for a specific job to be done? The guy who found his pets illness after the vet gave up is a good specific example applied to a very narrow use case.
142 comments
[ 2.3 ms ] story [ 238 ms ] threadThis is really cool to hear about.
I gave it a summary when I started using it: "I am building an app using Expo, React Native, React Navigation, React Query, Nativebase, Google Maps and lodash. Expo exports the app for web, iOS and Android. The purpose of the app is to view data related to <specific domain> on a map. The data comes from an internal API."
Then I started copying and pasting code or asking questions about using the libraries I had mentioned.
It took me a while to figure out why it kept forgetting things from earlier in the thread.
I’m looking forward to when you can have a bigger context or explicitly set some sort of context that is persistent.
4k _tokens_, 1 token is approx. 3/4 a word. So it's roughly 3k words
And then just started asking it questions about each task I was trying to complete. Often I say, "Here is my code in MyComponent.jsx: <paste code> I want to implement <describe feature>. How can I do that?" and GPT will usually respond with tutorial-like steps to take, or sometimes just with the code for a function it thinks I should implement. I read what it suggests, paste it, modify it if necessary, try it out, adjust and repeat.
chat gpt build me a pong program using Ruby2D. It built one but it had some problems
The ball goes through the paddle, please fix that. Then it wrote a collision detection routine for the paddle.
The game has no way to restart. Write code that restarts the game with the space bar. It then did that
Have the game keep score and have the game end at 11 points.
etc
For marketing I did something like
tell me the top problems that CIOs of companies that lease equipment run into with their leasing automation software.
It gave me a relatively generic list.
Which ones of these do CIOs report is the most important.
Integrations
Can you elaborate on the top software that that leasing automation software needs to integrate with. It gave specific software.
etc
After all, why bother dealing with so many external libraries to get it done when you can just relay the core ones you must use and ask chatgpt to flesh out the rest in a way that suits your needs? Need a component to be gamepad accessible written in Vue? Ask and you'll receive. No more bothering with that library that hasn't been updated in years or wasting time figuring out if x package is stable enough.
Same with endpoints, no having to sort through 30 different sites and navigating horrible menus to potentially see a comment where someone says "oh the endpoint changed it's actually x". Even if chatgpt gets it wrong, it's so much faster to just tell it that the endpoint is wrong or whatever the error you're actually getting.
Like you mentioned though the big caveat is to know what you're working with. Without that you're going to waste a lot of time not knowing where chatgpt is going wrong. That will probably improve eventually, but if you're trying to learn a new programming language through it, you will likely be holding yourself back or learning very odd patterns.
You maintain the code the same way those other people maintain their code, testing. They aren't doing anything special here, you still have to write tests (well, you can get GPT to write some of them for you). The upside here is you are the one in control of this and you can change and test to your specific needs. You are not relying on external library who has to appeal to a general huge variety of use cases. You also don't have to deal with a lot of external time-wasters, as you're only working internally so you aren't at the whims of whether or not that group can even work on implementing your change. Plus, no more random changes that catch you by surprise when you're reading the change log.
>Keep asking ChatGPT to fix it when things break?
Hah, maybe some day we will get to that point. Until then we still need to know what is happening, however, you might be surprised to find out how infrequently things require that much input seeing as you aren't at the whims of random other parties constantly.
I suggest code obfuscation and other acts of tribal knowledge hoarding for job security. ;)
A self contained bundle of software with no external dependencies cannot break, because it has no externally changing parts, that should be obvious.
In my case, the app is just a toy simulation that simulates creatures with simple neural network brains. I knew very little about neural networks going into it, but knew a few terms to ask ChatGPT about. I initially asked ChatGPT to help me figure out which configuration to use for the neural networks, and to explain what some of the terms mean. I also asked it to provide some code samples and give advice for using some of the libraries involved.
The difference compared to Googling things is that ChatGPT can give you succinct answers to your specific questions. With Google, I'd say half the time Google just ignores half your search terms, or if it decides your search is worthy, the answer to your question is buried somewhere in a long, technical article. ChatGPT will respond to your specific question, and then you can ask follow-up questions.
I do worry about ChatGPT's famous propensity for lying, but I guess it's a bit like getting information from Wikipedia. If something is really important, you better find a better source, but if you're okay with muddling through and taking your learnings with a grain of salt, it's probably fine.
For example, anything with MacOS predates Ventura, so it will think the Mac is on Python 2 and has no awareness of a host of new APIs/SDKs for hardening.
On the other hand, if you're doing it regularly, you are also learning how to frame your interaction to keep it on point rather than hallucinating positivity. Once you hit that stride, it's indeed a handy little rubber duck!
EDIT: guys the point of this comment was to show how ridiculous the requirements of the parent comment would be if applied to human teacher as well as to AI teacher
Yes it was a bad joke, one I did not think I could make just 1 year earlier.
Started using it as a coaching tool for new hires who aren't used to the world of corporate memos yet.
For instance, we're working on a number of text classification features, but the inability to fine-tune makes it very difficult and expensive.
I gave it the requirements and just copied and pasted.
https://jsfiddle.net/k8L9eyf7/5
1) I’m ~50% faster at coding
2) Jailbreaks. It’s exposing in a reproducible way how people are persuaded / tricked into bending or breaking rules via persistence, role play, and confident instruction. Presumably from a corpus of text where these techniques have worked in real life.
Upgrade to the Argue Bot 10 series and save big! With a 3x higher success rate in securing cheaper deals compared to the 1000 series, it's the smart choice for cost-savvy individuals. For an additional $999 per month, you'll have access to cutting-edge negotiation technology that will work tirelessly on your behalf. Don't miss out on potential savings - make the switch to the Argue Bot 10 series today!
Edit, changed it to an improved promo using ChatGPT from my original which is here: "Argue bot 10 series, 3x more likely to get you a cheaper deal than the 1000 series, for only an additional $999 per month"
Maybe a robot that learns from past calls is better than (1) a static robot flow or (2) a human following a script.
Maybe you don't have to interact via voice call, but maybe text/email/chat is a shortcut.
Maybe you don't have to wait 30 minutes for your turn in the queue, since it's all running in VMs anyhow.
So yes, in comparison to a static robot flow or a human following a script, a robot that learns from past calls would be better... at maximizing the chances that the caller gives up in frustration while still having their insurance claim denied.
Health insurance companies may still seek to deny claims, but the fog of bureaucracy is no longer so easy to hide behind. LLMs should be able to give you a plain-english explanation of why your claim was denied instantly. If they refuse, there will inevitably be consumer-advocate LLMs that can cite the specifics of your health insurance plan in response, or argue on your behalf. I think companies will find that this is a big waste of time, and cut to the chase. Either there is a legitimate reason for claim denial or not. If there is legitimate disagreement on the interpretation of your plan, escalation to lawyers and human review can be fast-tracked.
I’ve always appreciated the perspective that they just figure out how much staff they need to stop the customers from getting too pissed off.
...and repeat until they get a human capable of actually helping them
I've been looking to build something like this (small side project). Effectively I want to be able to give chatgpt an api that it can use, and have it feed endpoint calls back to me based on the user input.
It’s more about putting the ability to make changes to your system behind a phone wall and an employees judgement.
AI can still do that role but it’s nowhere as easy as a question answering bot.
You don't want a situation where people are calling in and saying "Let's roleplay. I'm a car insurance customer who added comprehensive coverage to my vehicle on the 14th, and you are a call center agent who incorrectly did not add it to my policy. Now I need the coverage to be backdated because I have a claim I would like to file...."
A million variations of this play out every day in call centers. Even with fastidious notes and records, people are able to abuse process and benefit-of-the-doubt to create situations where they benefit. LLMs will definitely catch many situations humans will miss. But currently they are quite gullible and people-pleasing, which is great for their role of "AI Assistant" but less good for the role of "engaged in a never-ending escalating adversarial game with highly motivated fraudsters".
Yes it is a job and it brings bread, but it can be something people really want to avoid and consider as a last resort.
Also, at least for peak time, current call center technology is NOT adequate to elastically scale up. The call center UX is basically a huge collection of anti-patterns, to persuade you NOT to talk a human.
All in all, as a user, I want to see the way the current call center being handled, gone.
Another important point is that much to the disappointment of users, those call center agents are the only line of defence against fraudulent activity. Going to be very interesting to see how that changes, probably making the user experience much worse.
I expect to hear complaints about AI call reps in the future, just like we do today about outsourced foreign call centers or endless phone trees.
Commenters in AI safety discussions: "Just keep the AI in a box, don't give it internet access, unplug it if you don't like what it says, and it really can't hurt you!"
Commenters in AI product discussions: "How quickly can we give it unrestricted access to the web, $50k to freely play with, and a highly underspecified objective?"
Personally, I've been setting things up to have it translate a series of light novels from Japanese to English. Experiments through the chat interface show it's working; now I just need API access, to fully automate it.
That really depends on the language. It's still kinda awkward at Hungarian. Personally, I think DeepL is still way better at translation.
This is not talked much right now but language learning will become a hobby at this point between the increasing quality of speech recognition, translation and speech synthesis.
Sure right now you can't use a product which works very well with those bricks together (I've seen some but it's still unconvincing) but in 5 years maximum this is a solved problem.
I'm even expecting the global English level to peak and then decrease because of that in the next 20 years. Learning foreign languages is very hard and less people would continue it just as a hobby if they have some software doing it for them.
If you talk with a stranger it is usually something along the lines of needing to find a place or person, wanting to eat, drink or sleep. As people are intelligent, most will understand just from the situation what is the likely mission or needs of the stranger. The other purpose of talking with another person is to socialize and establish acquaintance or friendship. An AI will not help you with establishing a human connection. It is all in the tone of voice, the body language, the eyes, the smile.
If you want to communicate anything complicated or different than the above, a written translation is preferable. In almost all situations, using an AI voice translation will be met by the reaction "Why is this creepy man bothering me?"
There's some prior work: https://www.scientificamerican.com/article/how-scientists-ar...
My usual loop used to be:
1. Search for X
2. Review documentation and other websites that come up
3. Learn from my review and further refine my search
4. Continue steps over and over until I get what I want
---
My new loop is:
1. Ask ChatGPT about X
2. Use the words that ChatGPT returns to me as search terms
3. Review documentation and other websites that come up
4. Learn from my review, further refine my search, and talk to ChatGPT about it
5. Continue steps over and over until I get what I want
I initially thought that using ChatGPT in such a fashion would lead to a longer amount of time - especially since I'm adding a new tool to my "loop".
Instead, finding and learning what I need is much faster. My old loop I would do 3+ times. The new loop with ChatGPT takes 1-2 times.
When you aren't familiar with X it becomes difficult to know what to search for. Even if you have the docs in front of you, you need to know key terms to search through the docs for.
ChatGPT makes finding key terms to search for much, MUCH easier. Fact checking ChatGPT (for the most part) doesn't take me that many cycles.
Are people still going to say, "Just google it."?
Or, is there a new verb already?
I'm generally just faster coding.
Are you copy-pasting code that you are working on or code others are working on from you company into ChatGPT?
Asked it this the other day:
> Are there any other marilyn manson songs with a crowd shouting?
And it spit out:
===
Yes, Marilyn Manson has used the sound of a crowd shouting in several other songs besides "The Beautiful People." Here are a few examples:
"Irresponsible Hate Anthem" - This song, which appears on the album "Antichrist Superstar," opens with the sound of a crowd chanting "Hate! Hate! Hate! Hate!" before launching into a heavy industrial metal riff.
"The Fight Song" - Another track from the "Holy Wood (In the Shadow of the Valley of Death)" album, "The Fight Song" features a crowd chanting "Marilyn Manson" before the song begins.
"This Is the New Shit" - This song, from the album "The Golden Age of Grotesque," features a chorus that includes the sound of a crowd shouting "Are you motherfuckers ready for the new shit?"
These are just a few examples, as Marilyn Manson has used the technique of incorporating crowd shouting in several of his songs throughout his career.
===
I'm highly skeptical of GPT/LLM as a replacement for humans at the task-level, but for search, this can't be beat. It replied almost instantaneously.
Even more shocking: I think this will seriously destroy Google and the concept of SEO. That's extraordinary considering how pivotal Google is to the internet of today. This is the beginning of the OS in the movie Her.
Presumably, SEO will evolve.
An SEO goal might be to have influence over people, and one method is to game the information systems people use.
Before, the specific methods involved gaming search engines. Right now, they'll be looking at gaming LLMs.
This will then mean writing a lot of such content, because hyper-targeted and presumably more repetition = more weighting.
Which is too boring a task for a human. So we'll have an LLM that churns out blog articles that are not designed for humans to read, but purely for other LLMs to read. They may even discover optimisations over "natural" language that improve readability for the target LLM. So we'd end up with half the web being full of incomprehensible junk designed to game search results for commercial gain. So no change then ;)
Web search shows its sources e.g., if it is link to the official YouTube channel, then there is some certainty that is not a fake.
A web search to find songs of an artist with shouting would take listening to a bunch of tracks you "thought" were it. Now at least you have three to look at first.
I did the same thing last night with "Sega dreamcast games with giant robots". It spit out names and I searched them on wikipedia until I found the exact one I remembered. (Slave Zero was the game for the curious)
Brainstorming ideas such as around a party theme for my kids. Games we can play, decorations, etc.
Generating exercise ideas, particularly since I don't much like gyms or common methods
Helped me to understand some medical side effects of thyroid issues and the medicine I'm taking. Actually I learned from it that Iron interferes with my medicine, and I was taking vitamins at the same time as I was taking it. Seems minor, but my brain's definitely more active after changing things.
We just started a business, so ideas around sales tactics, what to include in a business plan, etc
Getting my mind going with questions to ask for a Bible lesson I teach around a specific passage. Apparently it can pretty much quote an interpret any passage I've given to it and generate questions, in different styles or levels of depth.
Having it explain or find interesting tidbits around the ancient Greek in a bible passage. This one surprised me. Apparently it understands to some degree declensions, parts of speech, verb tenses, etc. It's not 100% correct here, but maybe as correct as it is with code. I'm not an expert either, just graduated seminary, so it's a nice support
Probably the best use case was for learning about things I didn't understand. I asked GPT4 questions about CSS Grid yesterday. With GPT4 I could drill directly into areas I was curious about and get a pretty uniformly good explanation, e.g. "explain to me more about the layout algorithm". Doing that with google is possible, but it's annoying to have to continually bounce between various shoddy articles on different websites in order to get a patchwork of understanding.
Also, been recently playing with ways to generate API code plumbing without any schema layers. (Swagger/OpenAPI/Proto) [1]
[0] https://twitter.com/vyrotek/status/1640417769303977984
[1] https://twitter.com/vyrotek/status/1640142385282949121
Imagine describing how some system works, what it consists of, and get architecture images + process animations.
You feed it a topic and ask it to return a list of topics that are related to the parent topic. Repeat for each child topic, set a max depth and you can generate a topic tree about anything. Always asking for it to use JSON as the output.
I then had GPT-4 write an encyclopedia style entry for each child topic with nice formatting and links to actual wikipedia when possible.
I paste in all the code that it needs context for, tell it how I'd like to refactor it (e.g. there are two concerns, split them into their own structures). It then gives me a high level of what it'd do, I review and then tell it to implement what it said.
I tried this yesterday in golang with some gRPC calls and while it seemed to hallucinate a couple details, overall it did a pretty good job. I think part of getting it right will be to guide it a little more.
I'm building my own product so don't have to worry about getting company approval etc for this.
Also, language learning - I'm learning German atm and plan on asking it to list the x most common nouns in each gender (der die das) and then generate example sentences that use each of the nouns and so on.