Ask HN: I don't use AI. What exactly am I missing out on?

This is a very good question, and I'm sure you will get valuable answers.

My take:

- More efficient than search engines at obscure problems, like configuring an unix tool. Help that you can ask 'explain step by step'

- It expands the range of topics you can read scientific papers in. You just ask it any topic or keyword you don't understand

- Good summaries of text that is valuable, but not worth reading in detail

- Translation

- Code generation/bug fixing. I don't use it this way but other s will tell you whether/how much their coding life improved this way

I never use it for getting answers, for the reasons you mention. A Wikipedia article has a lot more depth than any of the chatbot answers, and is more trustworthy.

Github copilot is slightly useful for creating boilerplate code. But I try to code in ways that avoid boilerplate in the first place.

Image generators are great if you need custom images and can't draw.

If I felt the need to write formal business emails with greetings and full paragraphs I might use it for that. But I mostly just send a few sentences, which saves everyone time and feels more on-brand for me.

I find an LLM is a great way to replace questions that would ultimately end up with me scrolling through a Stackoverflow answers.

For example, "How do you do X in Golang?"

It's also great for translation-like questions like, "Given this cURL command, can you generate a Go function...", or "Can you port this Ruby code into Go", etc.

Try phind[0] - it works without login or payment and provides sources to its claims, I switch to it when I can't find something via standard web search but it has not helped me yet. I am still convinced that AI is better than web search only in Layer-8 related queries.

[0] https://phind.com

It helps for a lot of things that would otherwise require searching the web and/or scrolling long pages: pronounces, grammar, general (and even somewhat specific) history/science/arts concepts, draft formal text, improving phrasing, simple/direct technical questions.

However:

- it requires some level of fact checking

- it subtly reduces my personal retaining of information for technical concepts

- (in my personal experience), it largely fails on highly technical questions

Entering unknown knowledge domain on subject. AI helps get up and running way faster introducing you to terminology and what to follow up (search) afterwards. Helps grasp concepts faster.

I'm strong at PowerShell but haven't touched Hyper-V VM creation - I wanted to automate it. So I asked some questions on how I want my VM configured and it gave out a great starting point of powershell commands and parameters I need to use.

The bash command line chatgpt interface has saved me many many context switches. You can essentially ask "command to convert all files in this folder to X, and do y, z", and it'll output the command and even run it if you like the command it outputs. This should be the standard interface for ffmpeg, which is basically unusable without stackoverflow. This has also helped me learn a ton of new linux commands!

I have recently been using chatgpt to estimate calories from pictures or descriptions. The few times I've checked, it's pretty accurate and is OK about providing ranges where it's uncertain.

I also ask it to tally calories for a day. It's got a chart of calories accumulated for me. It's so much easier than any app.

I often use copilot to avoid looking up syntax for things I should remember. I'll just comment / ask what I'm trying to do and it'll tab-complete that line. I rarely use it for boilerplate or larger amounts of code. But that one use saves me tons of time and context switches and is totally worth it.

I use ChatGPT for rough fermi-style questions. "How much does a cell phone weigh" etc. So much quicker than google, and better.

I often use ChatGPT to have an interactive learning session about surface details of a field I'm interested in. It allows me to find hooks into wikipedia or other search engines for specific topics to read more about, and to get questions answered quickly. It is second to none for this type of interaction and is really incredible.

Midjourney has taken hours of my life the last day or two. I love generating images in the style of this or that. For example, I'm creating schematics / blueprints for flying warships from the game highfleet, and printing them because they look cool. It adds some nerdy flair to my office. When I worked for NASA, my colleagues would do that for rover parts they've worked on, so I like the idea of capturing video game stuff as though it were important. If midjourney had a "buy print" button on it, it'd be dangerous and lucrative.

I feel like there's a couple startup ideas in this post.

I'm not onboard the AI hype train (see my comment history). But I do use ChatGPT on occasion as a sort of specialized/niche/synthesizing search engine. For example:

- English is not my native language, and prepositions give me constant trouble. I've used ChatGPT to check whether a turn of phrase is idiomatic, if I'm using the right preoposition, etc.

- ffmpeg command lines! "I have a MOV file that I want to resize to 720p, taking the audio track from this external MP3 file, and putting the resulting streams in a MP4 container". It gives me the command line, works most of the time. Same for some esoteric bash loop constructs.

- High-level overview of things. "How does my short film become Oscars-eligible?" It tells me things. I ask follow-up questions on the original answer.

- Detailed facts. "Explain why a wide-open lens has a shallower depth of field than if I stop it down".

Basically I don't trust it to generate new ideas, but it seems to be OK at answering specific questions about facts that can be synthesized from existing online sources. I'm sure these kind of questions have been asked online a million times, so I expect it to get the answers (posted by a human) and customize them to my specific use case.

The key to why I use AI is basically their tailored and to the point answers. Google tries to find prewritten blog posts etc. that may or may not hit the mark, and to get there you probably also have to wade through lots of ads and cruft and allow so many cookies to be stored. This often saves me a lot of time.

These tailored answers can even be tailored to different perspectives. You simply ask for the perspective you're looking for. E.g. "what is the problem with solar farms from a Republican perspective?"

But yes, one also needs to take care because due to how an AI is trained, they may have an innate bias and viewpoint even when asking for a different one. I've found that such "personal" points of view in conflict with some other are usually made pretty clear though if you're asking for another view. So it doesn't use to be hard to separate the two as both are plainly and separately made available in the reply if this happens.

But yeah, that's basically it for me. Tailored answers to my questions, as well as the generation aspect of course, where it can generate text or image content which Google can't.

You don’t have to commit to Google OR ChatGPT.

I use ChatGPT for questions with some relatively long wording and also for queries that I know will have tons of ads and SEO-spam.

An example: yesterday I tried searching for board games that are cooperative and best played by two people. Google showed a lot of ads, very old outdated blog and forum posts. I had to navigate for some time to gather a list of games. A lot of open tabs, a lot of scrolling, a lot of ad closing.

I asked the same question on ChatGPT and it immediately gave me a clean list. With a few different good games that wasn’t in my Google-based list.

Then I went to BGG to check the games, as you can’t trust ChatGPT that much, but you also can’t trust that much listicles from Google.

Using both worked better than committing to one or another. Still, in this case (and many others) I think exploratory search on ChatGPt is better than Google. As I check the results later on a site (not Google) that has authority on it

I am a bit of a greybeard around new tools, and even today my usage of ChatGPT is limited, so I will give my 2c. You're right, in most cases, Google + Github search is faster to achieve a certain task vs typing out your problem and waiting for ChatGPT to generate results (personal opinion). Nonetheless, here are a few things where ChatGPT saves a lot of time:

1. Generating config files: Eg, Webpack, Docker-compose. Describe your use-case and it will generate the config file you need. This is usually faster than perusing docs and looking for specific flags/options.

2. Generating CLI syntax: jq, cat, tar, find. The flags/syntax are so hard to remember when you want to do something specific. ChatGPT does a great job here in quickly generating the command you want.

3. Solving problems when you're stuck: When I feel stuck with a problem or an error message, ChatGPT has proven to be an invaluable resource. Even when the answer it gave was wrong, a little hint in there was enough for me to find the solution I needed.

4. Generating functions you're too lazy to write: Often, you know how to write a function but you're too damn lazy to write it. Eg, use axios, go to this URL, get the contents and write it to CSV. Writing a function to do this feels like a chore. ChatGPT is a helpful assistant in this case.

Overall, I think there are plenty of ways the tool can save you time, even if you're skeptical about changing your whole workflow.

Lots of great insights here. Personally, I use ChatGPT as a "sparring partner": I present an idea and talk with it to elaborate it further. Wether it's about programming, science, writing prompts or game design, among other examples. Another great use case is to use ChatGPT to speed up chores, like hard to remember syntax or functions that are easy but boring to think and write, like parsers. Or even thinking about how to build the architecture of a certain application.

I’ve found chat gpt very reliable for helping with coding tedium and avoiding typing. For example, I wanted to refactor 10 or so to tests with repetitive setup to a single parameterized test, all I needed to do was paste in the tests and ask. It’s also great for writing simple happy path tests for code you give it.

I had a go at porting an A*-algo implementation from pure python to a RAM & CPU -optimized Numba version. For this I used the help of Gemini-Advanced (if that's what Google names their paid-for offering today)

It's been a mixed experience, it got me started but did not help much with understanding the error messages, retreading a lot. Still basically from its dawdling I gathered knowledge I did not have before, and gave me a sense of how to progress. To get me over the bump to a functioning solution I had to do my own thinking, the back and forth with Gemini stopped. At the end I used it again to fix some edge cases, there it helped. Without it I would maybe not have attempted the task, or I would have started with a focussed study of numba, and with this old-style approach I do not think would have gotten to a functioning solution faster, not sure though, too late to try that now.

btw, can the ChatGPT "typing animation" be disabled and get the answer immediately?

For ambiguous searches, ChatGPT is a good supplement to traditional search engines. GPT-4 is very good at finding words from ambiguous descriptions or inquiries into whether something exists or not. Real example I needed recently: “thing that maintains a seal even though one side is spinning” => “rotary joint”.

Questions such as „Movie where the main character lives in a virtual world and gets freed from it?“ are much better answered by an AI then a search engine.

Completely useless for me unless I need to find something historical/obscure like an old reading group website for a theory book who's page is now only available on the way back machine. I could only remember half of their official name and chatGPT was able to find it.

I still use ~Google~ Kagi for some searches (people, addresses, items to buy), but ChatGPT is a very different tool for me, and is a regular part of my life now. Some examples.

- My kid's sports schedule is in a table on the school's site. Paste it into chatgpt, give it a few handling rules, and ask for the contents of an .ics file back. Now I have everything in my calendar.

- A job posting looked interesting but I couldn't figure out what the company actually did. Pasted their website copy into chatgpt, asked for an ELI5, and had a five minute conversation, after which I could convincingly answer the "why us?" interview question.

- I haven't done math problems in a long time, so when my kid needs help with something that's totally unfamiliar to me, I can paste a photo of his worksheet into chatgpt and ask for an explanation, and after a few minutes I know enough to help him.

Nothing for me. It is slower than just reading.

A couple things I've used LLMs for:

1. Learning new things - with interactive feedback. While there's a risk of hallucination, genereally for basic concepts on the undergraduate level, the proprietary models (eg. ChatGPT/GPT4) are generally pretty solid. For example, last year, I spent quite a bit of time asking ChatGPT about how the Federal Reserve works, and how it relates to interest rates. The benefit over plain old Google search is that I could ask much more precise questions than what Google searches would give me. You could check your understanding by confirming with the model whether a statement you make is correct. The feedback is invaluable.

2. Miscellaneous language tasks. Language models, unsurprisingly, excel at languages. It can do translation, rephrasing sentences or paragraphs (eg. make it more polite, casual, etc.), and generally converting sloppy writing into impeccable, grammatically correct, fluent prose. (And yes the larger models can even make anything rhyme if you ask them to.) More often though, when I need to express an idea but just couldn't think of the precise word, I can just ask ChatGPT for 10 words that have a similar meaning, and usually I'll get the word that I want.

3. As a coding assistant -- don't use ChatGPT for that, use a solution that integrates with your IDE/editor, and use it as coding autocomplete. It can help you write the (simple) functions you don't bother to write, and even though it usually fails at writing complicated things, sometimes the scaffolding reduces the amount of typing needed. And even if you don't trust the code it generates, you can use it to automatically generate comments and test cases.

I do agree it's probably not great if you think of it as a pocket search engine. LLMs have been called "a blurry JPEG of the web", you probably want to actually check info on the web rather than an inferior substitute.

One of my main uses of ChatGPT or CoPilot is as a less-opinionated StackOverflow. Either way, I still have to verify the solution, but the LLM gives me a more direct answer.

I also use it to generate (hopefully) fictitious names when I need a name. "Elara Nightshade" is more interesting than "Test1 Test" or "Jane Doe3".

You are missing 100x productivity gains if you are an elite engineer.