Ask HN: What is a specific use of GPT-4 that you think is remarkable? | Hacker News Reader

Ask HN: What is a specific use of GPT-4 that you think is remarkable?

88 pointsjasondigitized133 comments

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.

View on HN

I've got a thread going with ChatGPT about an app I'm building. It knows all of the libraries I'm using, and has seen the main App component, the Router and many of the key components. I'm asking it about every task I need to do, feature I need to implement, bug I need to fix. Basically rubber ducking with it and talking through everything I'm doing as if it were my coworker. It's reducing my time to solve problems by about 80%. I'm able to do 4-6 times as much in a day as I could without ChatGPT when I was stuck figuring out everything tediously on my own. For context, I'm a full stack developer with about 15 years of experience, so I know what I'm doing. ChatGPT is just taking care of a lot of the trial and error and figuring out how to use different libraries or implement certain algorithms.

So I have yet to use ChatGPT (I know, I know), but are you able to keep long running threads with it? How does it know context about your app? Are you simply copying and pasting code?

This is really cool to hear about.

Yep, the threads are permanent. I've got approximately 200 separate threads going with it right now.

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.

Wow thats incredible. How are you managing the 200 threads - are you using the api for that? Or just using their chat interface on chat.openai.com and scrolling to the relevant one to continue the converstaion

You can run into issues with long running threads. It only has so much context (I think around 4K words). It doesn’t have all the context of all the thread. My understanding is that it is summarizing the thread and resubmitting it as part of the prompt each time.

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.

32k words with GPT-4, no? That's what they said in the demo video.

only with api and only if you're invited to one.

32k _tokens_. There are approx. 3/4 tokens per word, so that's around 24k words

Oh, that's interesting. I hadn't noticed it forgetting context, but I just asked "Do you remember the original prompt I gave you when starting this conversation?" and it said "Yes, the original prompt you gave me was: 'Hello, can you help me convert 0.05 BTC to USD?'". So, yeah, it's definitely forgetting some things.

> think around 4K words

4k _tokens_, 1 token is approx. 3/4 a word. So it's roughly 3k words

Can you describe the structure of your conversation? Did you describe what libraries you were using and send the full code of the main app component?

I started with a pretty simple prompt: "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."

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.

here is one I used for teaching girl scouts to code:

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

Would love to read a blog post showing this in detail.

I'm seeing similar, chatgpt replacing a lot of external libraries and digging through the weeds to find the right endpoint.

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.

If you're not using external libraries that are maintained by other people, how do you handle your code's maintainability in the long-term? Keep asking ChatGPT to fix it when things break?

I want to make something clear, which is that there isn't a total avoidance of external libraries. Only that you're not so quickly choosing them for every problem. It makes sense to use Vue (just an example), it doesn't make sense to use a library that's a wrapper on top of another wrapper to use within Vue. That sort of deal.

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.

When this is possible, most human coders will be out of work.

I suggest code obfuscation and other acts of tribal knowledge hoarding for job security. ;)

Usually things break because you used an external library.

A self contained bundle of software with no external dependencies cannot break, because it has no externally changing parts, that should be obvious.

ChatGPT replaces StackOverflow

Indeed, but it's sad because ChatGPT won't be indexed like SO, so in the future, we won't really get the same sorts of improvements as we have now.

Hopefully that far in the future, it will actually understand coding on a deep enough level that we won't need to train it anymore.

I have a similar-ish thread going for an app I'm building.

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.

The catch with this is when there's been a fundamental shift in best practice within the past year or two.

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!

I've been using it for long-term memory supplementation. Had a faint memory of driving around in a friend's truck back in high school and was trying to remember a Marilyn Manson song we always played.

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.

90% of searches are trying to find websites not teach concepts. If they want to learn something it's usually current.. who won the game.. fire locally. If you want to learn how to lift weights simple instructions are not going to replace videos of fit influencers you follow. I think google is safe

Similarly, I used it to answer the question, "Who was the Black female author whose works were saved from an estate bonfire?" Neither DDG nor Goog could answer for me (kept coming up Maya Angelou which is totally wrong--it was, of course, Zora Neale Hurston as I should have remembered in the first place.

I like it for tip of my tongue questions too

> I think this will seriously destroy Google and the concept of SEO.

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.

Presumably by writing content that is hyper-targeted at specific questions (e.g. "which is the best X for me to use?") that the LLM will then consume, hoping that it will be regurgitated as fact when the LLM answers that specific question. [I am not an expert on LLM training, so I might be wrong]

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 ;)

How certain are you that these are real songs and not just hallucinations? How would you check it?

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.

You check the output by listening to the songs.

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)

You are assuming that a web search is somehow fundamentally worse at finding _existing_ content. Ridiculous. Given that it can use LLMs (among other things) behind the scenes. The opposite can also be true (LLM can use web search), the difference is how likely that BS is generated.

Is this a trick question, or something? You look up the song name on YouTube/apple music/Spotify/a song index to listen and see if they ended up being the song you thought about.

over time it will be able to play the songs and the exact location in the songs. You can ask for references now and it hasnt worked well.

The fact that you can put in garbage unstructured data, ask it questions, and then "convert" it to JSON is kind of magical. It's like computers have a new analog/organic input system. [0]

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

Maybe we can finally code a 100% functional CSV importer which deal with all the broken stuff people invented haha

Impressive, but it seems to have hallucinated Lilly's location.

Also, was M from Male or the initial of another name?

Call center employees are pretty much as good as gone once the price comes down on GPT4. You can pretty easily give it a good prompt that allows it to answer very specific questions about your company, or interact with external APIs to schedule services, make changes based on the customer's requests, etc. Throw some Speech to Text and Text to Speech services in front and behind it, and voila you have an AI receptionist that blows literally every automated answering service out of the water today. IMO, 90% of call center employees will be out of work in the next 10 years.

I am dreading the inevitable future day when I have to argue about a denied health insurance claim on the phone with a fucking robot prompted to make it as difficult as possible for me to win.

Why would you do that when you can dispatch your own AI to argue for you?

It's Warren Ellises, all the way down.

We'll probably have tiered AI systems.

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"

The startup opportunity is to have an adversarial chatbot argue on your behalf. Feed it the terms and conditions of your policy, claim info, and then have it duke it out with the insurance company.

Slow down.

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.

"better" for whom? The key issue raised by the parent post is that from whatever sources the robot learns, we should expect it to be trained to use that data to make things better at achieving the goals of the robot owner which are different from - and often opposite to - the goals of the customer.

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.

Optimistically, it might be an improvement. Imagine a world where you can call in or chat. Hold time is zero in both cases. You never need to "transfer" to a new department, or supply the same information over and over. LLMs have access to all the case files and needed info on the backend instantaneously.

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.

Often, holding times are not only due to a shortage of call center employees available, but also to encourage the caller to give up. I doubt the hold time would be zero, unless the government legislates a maximum hold time. Even then, the AI could stall with endless patience and requesting clarifications and further information, and repeating details back to you to confirm, with minor errors requiring correction, which might be even worse than waiting on hold.

Do you have direct experience with a call center where this was happening? This sounds like a ridiculous conspiracy theory. In every call center I've ever had experience in, median/mean hold time was a KPI that was relentlessly optimized for. Hourly demand was modeled and staffing levels were set such that hold times were kept to the minimum possible with the resources available. Calls-not-answered was an inverse metric. That's not to say the optimal hold time is always zero. It's also bad to have too many staff sitting idle "just in case" demand suddenly peaks. It's a tricky problem. But LLMs obviously obey the laws of compute scaling, not human scaling, so I would expect it to scale to demand very quickly.

Hold times are optimized for customer retention and profits. This implies that you’re both correct. It’s just the perspective: if people give up on their query but stay a customer that’s fine. Most call centers don’t exactly have the degree of customer service of a hospitality institution like a luxury hotel.

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.

New security threat unlocked: DDOS for chat agents

Fight fire with fire, train an LLM to respond on your behalf.

I bet GPT-4 could more effectively maintain that 0.002 cents equals 0.002 dollars.

Don't you have that exactly today where everyone presses 0? People call because they want a human.. otherwise the website can provide those answers.

> Don't you have exactly what you have today where everyone presses 0?

...and repeat until they get a human capable of actually helping them

They only want to press 0 because the system sucks, GPT-4 is good enough to be your best friend.

> or interact with external APIs to schedule services, make changes based on the customer's requests, etc.

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.

Most call centers aren’t about answering questions though.

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.

Bingo. The biggest problem with implementing LLMs-as-call-center-agents, at least in contexts like insurance, is fraud. Even GPT-4 is just too easy to fool currently. Call center conversations are often adversarial, where the caller wants the agent to create a change to the system that is somehow fraudulent or to their benefit, and it's the agent's job to hold the line.

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...."

that wouldn't happen though because the program would see that the user didn't in fact change their policy.

Didn't they? Or was it a system error? Perhaps the LLM has no record of it because the user's sister's boyfriend called in with a different phone number and misstated the policy number. But the change was definitely supposed to go through. In fact it's an absurd failure on the insurance company's fault for not processing that change correctly. Let me speak to the manager! I'm calling the Department of Insurance and reporting you!

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".

And at the other end of the scale, you calibrate it too be too suspicious and you end up with "You have been a bad customer, I have been a good Bing" style arguments with lots of legitimate claims and a class action suit. You can absolutely have a bot answer general questions and especially recurring ones, but you're going to want it to escalate to humans pretty early on. Even human agents often escalate to human agents.

This is why well architected systems will require the AI agent to do things through other services via well-defined APIs. First, it allows the business to limit the space with which the AI can interact with the rest of the business. Secondly, it allows the business to use good ol' fashioned AI, logical rules, etc that can prevent an AI agent from doing things that it shouldn't. Of course, there will be exceptions to the rules, the same as there are today, but it should drastically cut down the number of humans involved in the process, and when a human is required to intervene on the business' behalf the AI can summarize the entire conversation + the reasoning on the business side for not acquiescing to the customer's demands.

I think call center is already a job that nobody wants.

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.

Remote call jobs are a solid job to have these days. A far cry from the horrors of a typical call center, plus if you're someone who struggles with a "normal" physical job, a remote call center position is near the top of your list.

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 think companies will try to eliminate call centers, and customers will treasure a good call center even more than they do today.

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.

I did a medical lab test the other day and I've been trying for several days to reach my original doctor who for some reason I still can't get. I don't know what to make of the abbreviations and scales on the document and I sure as hell wasn't going to visit another doctor and pay a fee to grok it without a need to or google each line to find out I have "cancer or some other random ailment".

By chance I was using cgpt for something else and I just copied and pasted the unformatted data from the pdf and voila. https://i.imgur.com/tCy1dWU.png, there is a more personal page that I didn't share that gave me a eli5 explanation for the exact problem I got checked for that I couldn't tell and it pretty much worked the same using the terms and even gave me recommendations as to what I could do next.

I am surprised at how good ChatGPT is at with abbreviation. I'm working with a lot of discharge summaries at the moment, and ChatGPT is a god-send in interpreting all the shorthand used.

It will take me forever otherwise to find out that the shortform CR2 means second complete remission.

I think GPT4 isn't amazing (that is, better than excellent humans) at any specific task, so I doubt that many things it does, in isolation, are remarkable. Where they shine is that GPT4 and ChatGPT are really good at doing lots of menial tasks really quickly. e.g. I can ask it to refactor my code, rewrite it in a different style, add a new function etc and it'll happily do that all day.

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.

Unlike 3.5, GPT-4 is competent at translation. Not just competent; it's better than most human translators. We might not want to leave it to its own devices yet, but there's a large number of novels that were never translated, or were only translated badly.

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.

> GPT-4 is competent at translation

That really depends on the language. It's still kinda awkward at Hungarian. Personally, I think DeepL is still way better at translation.

In what ways would you say its awkward? I'm really curious about its performance in languages that might not have shown up in its training data much. I will say, I've tried using it with a dialect, and it's good when it chooses to use the dialect, but slips into formal/prestige language extremely often.

Text translation is basically solved with DeepL, at least on the languages it supports, it's indistinguishable for my native language.

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.

No electronic solution can ever come close to having a real spoken conversation with another person. DeepL and other AI translators are great for text, though.

The bricks are already there, just not perfectly stitched together I give you that, we already have near perfect text translation for some languages, high quality earbuds, near perfect voice synthesis in some research papers.

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.

It can come close enough that if the choice is between an expensive live translator and a cheap (or free) voice-to-voice machine translation, then many people might prefer the latter.

You don't need either. We as humans are very good at communicating with each other even if we use different languages. What is important is not the words that are said, but the interaction between humans, which include much more than just words.

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?"

Translation is not solved with deepl at ll. Not with difficult language pairs (English/Chinese, English/Japanese, English/korean etc ). That's where LLMs shine. Far ahead.

I admit I don't speak Japanese or Korean, for the French / English pair, the current level is identical to a professional translator.

I agree deepl is excellent at English/French and other similar languages

We should really start pointing the giant LLMs at animal vocalizations and get cross-species translation.

There's some prior work: https://www.scientificamerican.com/article/how-scientists-ar...

I really hope we're able to use LLMs to understand animals in the near future. They are communicating, and we should be able to communicate back.

Do you have an example of this!?!! There's so many untranslated light novels I want to read but the ML translation has been historically terrible where it's not readable. Being able to translate light novels without a translator/scanlator would be huge.

i don't compare with GPT-4 and it's chinese but https://github.com/ogkalu2/Human-parity-on-machine-translati...

Mostly code here, but I've found a few uses, mostly around brainstorming that I use as a starting point to further thought:

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

I've been using it as a discussion machine and a way to cut down search.

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.

A lot of folks can relate to that I bet. To typing in something in Google, clicking on a few results, learning some new keywords you didn't know before, putting those back into Google, and rinsing and repeating until you get the results you are looking for. It's search, but supercharged!

Are people still going to say, "Just google it."?

Or, is there a new verb already?

One thing you can do for quick research -> summary is ask it to produce a detailed mind-map of the topic at hand, along with related/relevant long-tail keywords. The latter is effectively a SEO strategy in a nutshell, but having those keywords in front of you can really jog your memory.

I want to share a personal experience with GPT-4 and other AI tools that has helped me a lot, given my recently acquired physical disability.

My creative and critical thinking are still in great shape, but I've been struggling to execute my ideas because of my physical limitations. For example, I used to paint as a way to deal with traumatic experiences, but that's no longer possible for me.

I got depressed and frustrated, being no longer able to transfer traumatic memories out of my head.

I found a workaround using Stable Diffusion and ChatGPT. By doing this, I can recreate my paintings in my own style without having to physically paint them. It's now part of my weekly routine and has been a real support for my therapy and PTSD.

Another thing GPT-4 has helped me with is managing anxiety attacks. I'm not good at coding, it's simply a skill I'm not good at, but with some assistance from GPT-4, I made a simple browser game using HTML, CSS, and JS. This game is now my go-to distraction when I feel anxious and is easily accessible on my phone.

GPT-4 has also made everyday tasks, like writing letters or emails, a bit easier for me. I can shave off a few minutes here and there that way each day, which is a lot of time I can do stuff I enjoy if you add them up over a year or longer.

While these benefits may not seem groundbreaking to everyone, especially if you're healthy, they've made a significant difference in my life and for me personally, and I think that is remarkable.

I'm also not good at asking other people for help, especially for stuff I managed to do on my own for all my life.

So if I can ask AI Tools to do stuff for me instead of taking up a friend's lifetime, it helps me to keep a sense of independence and being able to take care of myself.

I cut and pasted my child’s school’s multi page complaints policy without reading it myself, then asked it to show me the stages I need to follow as a sequence diagram and it can up with this:

@startuml actor Parent participant "Class Teacher" as CT participant "Head of School" as H participant Clerk

== Level 1 ==

Parent -> CT: Informal Complaint CT -> H: Meeting Request H -> CT: Meeting Scheduled CT -> Parent: Meeting Scheduled Parent -> CT: Attend Meeting CT -> H: Meeting Summary H -> Clerk: Meeting Summary Received

== Level 2 ==

Parent -> Clerk: Formal Complaint Clerk -> H: Formal Complaint Received H -> Clerk: Response Preparation Clerk -> Parent: Formal Complaint Response Parent -> Clerk: Request for Level 2 Meeting Clerk -> H: Request for Level 2 Meeting Received H -> Parent: Level 2 Meeting Scheduled @enduml

I've been using it to refactor my code.

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.

It can explain stuff to you in exactly the way you need to, and walk you through problems. You can ask it to create quizzes / guided questions for you which helps in organizing your own thoughts. Another application I find quite remarkable is asking it for good prompts for image generation models. The biggest benefit is that you're not starting from zero when entering a new problem space / domain.

But how do you assert the information is correct? It could be dangerous to blindly trust an imperfect tool as a teacher.

I am not even in an environment where ChatGPT is very present and yet I've seen that happening sitting right next to a person with exactly that thing happening.

On the other hand, that exact line of thinking could be applied when taking about pair programming or getting advice from a more senior peer. In either case the answer is to critically think about the answers and seek confirmation elsewhere when needed.

If I'm taught by any teacher, even a human one, I will blindly trust them. Also I require these human teachers to be perfect (omniscient, omnipotent, omnibenevolent) so that I avoid danger. This is how I assert the information is correct.

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

We need to interrogate and test all information we receive.

executing the code it's recommending tends to cut the bullshit pretty quick.

You can ask it to specify or point out any logical errors you observed and it will correct itself. If there's a contradiction I know that the information might be incorrect. I'm also just using it as a starting point to prime my brain, of course as of today we still need other sources to verify the knowledge.

Truthfully... what's most remarkable to me is it not saying, "just stfu, you idiot" after me asking it question after question about stuff i want to know more about.

There has been countless times where I'll read it's response of, "sure, here's..." and it just blows my mind that i have the ability to barrage something with questions and it won't get tired of me doing so.

it puts a smile on my face.

I asked it to help explain the Birthday Paradox to a colleague who was struggling to understand the math behind it.

I gave it the requirements and just copied and pasted.

https://jsfiddle.net/k8L9eyf7/5

Give it access to a Python repl, show it how to send HTTP requests, how to take look at images of webpages if it needs to, and how to spawn new instances of itself - give it a budget and tell it to make money for you.

I don't want to underplay how much you've hit the nail on the head here, but...

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?"

I'm using it to write WordPress plugins that can be used in place of expensive yearly subscription plugins.

Why not use it to write expensive yearly subscription WordPress plugins? Make some money with it.

Definitely thinking about it, and open sourcing some too.

Populate open world videogames with real living characters that evolve to your way of gaming.

Apparently somebody tested chatgpt 4 as a d&d DM and worked really well, so this coupled with AI voice acting could lead to exceptional results in the game industry.

Write unit tests, document code, optimize code, explain code. I've tried Python functions and React code. It even converted a class based React component to a functional style component!

Can it really optimize code? And what granularity have you seen it succeed?

Sorry, I really meant refactor here. I did ask it to optimize code, and it came with a suggestion of trimming down my function into a smaller chunk and creating a new function. So it actually refactored it TBH, but I am going to try how I can check if it can optimize code too.

Finding lots of value in drafting emails, especially for topics or asks I'm not as used to dealing with, like asking for extensions or blaming other teams for blockers. Obviously takes a few iterations ('be less verbose!') but I get a lot of value in seeing a couple different ways to word the request and how to frame the background facts to come to the conclusion I want.

Started using it as a coaching tool for new hires who aren't used to the world of corporate memos yet.

I used to build my own little wiki / knowledge graph as a fun project to see what it was capable of.

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.

Neuromancer was published in 1984 and I think we are finally on the verge of creating constructs.

The future of GPT could be like a construct, or an AI embryo. Something you can purchase that will consume every piece of data you have ever produced - every text, comment, journal or blog. Then it will consume every piece of data you have consumed - every book, movie, song or headline. An AI that can be taught to respond as if it believes what you believe and loves what you love. Something that can be a convincing copy of you to represent you to the outside world.

People who only interacted with you online might never know the difference.

William Gibson was the most prophetic sci-fi writer.

When I first saw GPT stuff I thought it was basically Gibson’s ROM constructs, like exactly.

1.: It's the first time I'm seeing a chatbot who could easily replace the human layer.

Chatgpt understands tons of languages.

So I highly prefer to write things up like "you can change a username by going to this ui and click this and than that".

Or anything in-between.

Even if chatgpt only takes care of translating and creating a proper support ticket.

2.: Text understanding. I have plenty of tasks were a human would need to do things for me like "pull out all features of this building and make a list out of it" before I can verify it.

3.: I can already see the use case of fine-tuning my own agent with my own data. Now it's actually worth it to write proper documentation or architecture decision records. Basically doing brain dump because after I can just ask it things like "why do we use angular?" "Because you tried out react and didn't like it"

Two things I find remarkable:

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.

Rewriting text to be more fluent or readable. For all the (many!) people who need to create documents in English but aren't native speakers, GPT-4 is quite effective at rephrasing a technically correct but un-idiomatic text into something that's much better.

There are a lot of use-cases that aren't possible at current pricing and model access.

For instance, we're working on a number of text classification features, but the inability to fine-tune makes it very difficult and expensive.

Sorting by personality. Every search engine I tried was useless. I guess nobody's written that exhaustive article yet.

Sort by personality? Yep. I asked it to pretend to be a clinical psychiatrist and then asked it to tell me about celebrities personalities that closely matched "calm, thoughtful, and introverted".

I've been running similar experiments with creating groups of actors' responses that conform to groups of personality types: the four humours, blood type, etc...

I can't really wait for tooling to visualise complex technical concepts into nice animations and static images.

Imagine describing how some system works, what it consists of, and get architecture images + process animations.

This is what I am missing as well. I'm a pre-sales technical architect that keeps looking for use-cases where I am able to leverage this technology. 75% of my work is in draw.io and PowerPoint.

Write my PhD thesis, even incorporating the written and oral feedback of my advisor, which I record and use as input.

Yes it was a bad joke, one I did not think I could make just 1 year earlier.

don't forget about the made up data

On the occasion I do editing for certain documents on behalf of non-native English speakers. This can take up to two hours per document ordinarily. ChatGPT reduced this to 15 minutes, and almost all of that is just proofreading the output, waiting for it to respond and copying and pasting.

my daughter uses it to explain homework problems. Google search routinely fails for this. Chatgpt gives her the answer, but explains in detail how it got the answer.

marketing team is using it to create baselines for articles/blog posts.

you can do a search for "product manager chatgpt prompts"

my controller is using it to look up finance stuff including answers to questions about tax law. They will then follow up by looking up the first source articles to verify, but chat gpt can give correct answers.

HR is doing the same thing

when I run into problems trying to figure out how to do something in an application chatgpt works better than web searches, watching stupid videos, and especially the built in help for the application etc.

I have using it as my copyrighting assistant, especially when I have to write long copy

Retrieve information from public technical chats to provide Q&A on programming languages, technologies, engineering approaches and so on.

Can you describe how you are doing this, e.g. What is the series of prompts you are using to feed it the chats?

I'm yet not doing this, unfortunately. API limits wouldn't allow me to load in some IRC chat there.

1. Quick code reviews. 2. Generating tests. 3. Good for dumb questions, that I'd hate asking a person.

I'm generally just faster coding.

How are you using it for code reviews?

Are you copy-pasting code that you are working on or code others are working on from you company into ChatGPT?

Look up what is code for plumbing. It can be difficult to find what is code unless if you are a member of IAPMO.

Reply on news.ycombinator.com