I regularly use it for menial tasks and it has almost completely replaced searching documentations of major libraries for me. As a senior developer, it's a productivity gold mine if you learn how to use it correctly. But I wouldn't recommend it to any junior dev, because you need lots of experience to see when it starts hallucinating.
No, I'm talking about things like writing SQL queries (even complex ones), CMake files, Docker configs or plotting stuff in Python. Of course, if you're not already an expert in these things, you'll have a hard time distinguishing useful replies from hallucinations - that's why I said it's mostly for senior devs. Without expert knowledge, you will likely not be able to benefit from it in its current state. But if you have that and know how to write efficient queries, it can easily up your productivity by a factor of 10 (i.e. going back and forth for 6 minutes with GPT4 to make it get your requirements can save you an hour of work looking through documentations).
10x seems like a stretch, but I agree. I have a chatgpt window open most of my workday now too. "Hey chatgpt - help me write this mathematical equation in latex. Hey chatgpt - give me an example nodejs program that reads a line-delimited JSON file without loading the whole thing in memory first". "Hey chatgpt - What are the popular & good python libraries for drawing an SVG scatter plot of some data? Give me some example code for each library."
It depends on what you're doing of course. With your example queries I'd also expect a lower boost, but the benefits should still be there. Where it really shines is if you already know exactly what you want and how you want it and you merely need someone to write it in code in a common language.
As a senior developer: Experience. The output from ChatGPT isn't perfect, but it's good enough to point one in the right direction. The code it outputs is mostly garbage that a good developer would never put into production, but what it does fairly well is finding the relevant bits of documentation.
Example: Using a small part of a new, big, unfamiliar library. Rather than digging through the library docs, I can ask ChatGPT about it, which often points me to the relevant parts, which I then can still confirm in the official docs.
Yeah. I got chatgpt to write some code using Indexeddb the other day. The code was absolute rubbish, and I rewrote it to be half as long, while being clearer and with better error handling.
But the code it gave me was a great starting point. I found it much faster & easier to rewrite the bad code it wrote than pore through documentation and figure out how to solve my problem from scratch.
I've found that for questions that aren't super complex on things that are extremely likely to be in its training dataset (such as public documentation of popular Python libraries) the error rate is very close to 0%. Even compared to GPT from a few months ago I've found the accuracy of responses has increased dramatically.
I still believe anyone using these tools on a day-to-day basis should have a sense of "trust but verify."
GPT-4's solutions are usually correct the first time around, or can be corrected by telling it what it did wrong. The ask-clarify-correct loop is still faster and less effort than doing it yourself.
If people continue to learn outside of LLMs and can tell when it is wrong so they can correct it. The slope can be slippery. For those of us who grew up before the rise of LLMs, this isn't really something we have to worry about. But the next generation will be not know what it was like without LLMs, kind of like the pre/post-internet/phone generational splits
I more-or-less know what I expect it to give me, so I can just go somewhere else then. I tend to ask it three broad sets of queries, since it's what it does best at. tl;dr: it can transform text very well, even with lots of ambiguity. also it knows technical vocab
- if I'm working in a field that's new to me or that I don't understand, I ask for help understanding the basics and vocabulary of the field. it does very well at this.
- if I have a well defined problem, but am simply not familiar with the libraries for a given situation, it tends to do a good job translating my english queries into the right code. I do examine and test the code afterwards to make sure it's correct. You can really see this in action when you ask it to do data analysis; and the REPL loop in that mode is also great at catching bugs.
- if I have like a copy-paste from a documentation site, I can ask it to transform that into code or into a better-formatted version. this saves a lot of time, and I don't have to remember regexes or vim keybinds
> if I'm working in a field that's new to me or that I don't understand, I ask for help understanding the basics and vocabulary of the field. it does very well at this
I also do this, but I am careful about being confident that "it does very well at this". We can't actually evaluate what it's putting out, other than that it sounds plausible, which is something LLMs are truly great at.
Would have to disagree. I've writing code for +25 years and counting, a good portion is critical infraestructure code. I've reached the title of chief-expert and quite frankly gets difficult nowdays to find people who can answer my questions or at least help explore possible routes to solve them. ChatGPT is like a buddy that is always available and sitting next to you for a good technical chat.
Sometimes it comes up with really good techniques that are different from my usual approaches, other times is plain wrong and I'm correcting it.
I am more productive with the ChatGPT in my life. Whenever stuck on some weird error, instead of googling I paste the log and discuss with the bot what can be going wrong there. We can talk about pointer analysis, performance bottleneck comparisons. Could I do it all on my own? Sure. However it is boring and quite certainly would require 10x more time to perform all calculations on my own.
In the end of a day it is just another tool. Brings advantage when used properly.
If I'm looking at MDN for JS docs, I don't need to cross-reference anything. Yet I can't be sure it used the same MDN in my answer as opposed to some random SO post.
I could specify for it to use MDN exclusively but at that point I might as well use search.
In addition to that I could judge the quality of search results (a lot vs little mentions of a technology, shady vs reputable site etc.) to make educated guess of the output I'm getting from search. Can't do that with GPT.
Having the context for a recipe makes it a lot easier for me to evaluate whether this is probably a quality source or not. I really don't get the hate so many people seem to have for anything other than a list of ingredients and steps to follow.
In general, the context of search gives some insight into the credibility of the source.
It’s because they are entirely manipulative, and usually have nothing to do with the actual recipe. The “context around the recipe” rarely is written by the person who put the work in to develop that recipe. It’s almost always content-farmed out or more recently, entirely generated by AI.
The only reason those blobs of text exist is to get you to look at more ads. Put more things in your head against your will, sell you more garbage, and manipulate your feelings.
If it wasn’t true, why are the recipes always at the bottom? Why not put the most valuable part right front and center? These websites have no respect for you and likely copy pasted the recipe anyways.
Random life stories are everywhere, and have nothing much in common with cooking well. Even before LLMs could fake that part as easily as the recipe itself.
Only way to know if a recipe is good is to look at it.
> Do you trust the content of the sites Google returns you?
You don’t. But i’dtrust a top rated Stackoverflow answer over whatever LLM spits out.
There is no “confidence score” from an LLM output. You cannot tell whether it is making things up (and potentially make very bad decisions based on it’s output)
There's no confidence score for specifically ChatGPT, other GPT models hosted by OpenAI (let alone the broader research community) have been given that capability.
I honestly never thought of using ChatGPT for recipes. I just asked it for "a simple pizza dough recipe for a medium thick pizza you can cook in an oven" and it returned the exact recipe I have memorized which I think came from a "AirBake" pizza pan I bought 20+ years ago. Thanks for the tip!
Just to be clear, I wasn't complaining. I liked that it came back with the one recipe that I've already settled on (and I've tried quite a few over the years.)
I think I'll ask it for a calzone recipe this weekend. The one I use now makes the dough a little too bready.
What do you trust? I don't trust search or other humans to be 100% accurate about anything so I find it really strange that people presume I take whatever ChatGPT outputs at face value. I parse it just as I parse any other information.
I’ve seen this discussion many times over the past year and have come to think that the disconnect basically arises from the way that we have thousands of years of heuristics built up for interpreting the trustworthiness of what another human is telling us, 30 years of evaluating websites, and less than a year of evaluating LLM outputs.
People have some sense that someone giving them information may be an {expert, charlatan, idiot}, or that a website they’re looking at is run by a university vs a blogspam content farm, but many have not developed a sense for when or how much they can trust LLM output, which is delivered with the same tone and confidence regardless of whether it’s entirely fabricated.
There is probably a component of personality involved in how people approach this. Collectively we are all learning how to interact with this new source of information and people take varying paths.
Exactly. Whether it's a person, a website, or a book, we have a ton of cues that give us some sort of intuitive reliability score. That reliability is essentially never going to be 100%. But, especially if we cross-check sources, we can start to have very high confidence that an answer is true--at least as far as anyone knows. (Or even that no one really knows the answer with any certainty.)
I've had ChatGPT return very serviceable "true" results and I've had ChatGPT return utter fiction.
1. you don't know the answer, but you can check yourself and easily whether a given answer is roughly correct
2. you don't know the answer and wouldn't be able to check how valid a potential answer is
LLM-based tools are great for 1 to synthesize various sources into one coherent answer, since in this case, you won't become a victim of their hallucination. E.g. "write a one-off Python script to do this": you can quickly check if it does the job, even though you couldn't say whether that's idiomatic Python.
When I'm writing (English text) about something I'm passably familiar with, it can be useful for generating some straightforward descriptions and background. Nothing I'd just cut and paste out of the box wholesale but it can be a timesaver, especially if it's somewhat boilerplate.
I would say it is not good at giving a sophisticated answer to anything that requires a lot of nuance. And I've also asked it questions with fairly objective factual answers that it gets hilariously wrong.
I still struggle to find any use to it in my daily life. It is a cool demo, but no one wants to read AI generated text from other people.