I don’t think that is surprising, but I’m curious as to how much it’s inaccuracies will impact our society as people start using it more and more for various things. I use it mostly for menial tasks, such as given it a mermaid table or a copy paste from a data mapping excel file to get it to auto generate the necessary types and CRUD code for me, or as a search engine. Especially in the latter task, however, it’s wise to keep in mind that it’s very good at “telling lies”. One example is that we had a discussion about the name or the horse in Pippi Longstomps at work (don’t ask), which eventually lead to us discussing the colours of the horses dots. So we asked ChatGPT and got three different answers, none of which were correct.
I do consider it a very useful tool in helping me with things that I’m already an expert on, I’m not sure how I would feel about using it for brainstorming. Maybe if you’re really just using it to speak out loud and not really listening to its responses, but if you’re using ChatGPT to learn or refine subjects that you’re not an expert on, then I think there is a very good risk that your output won’t be, well, great. But it’s not like it’s not happening, and I think it’ll be interesting to see what sort of effect that has on our society in the coming decades.
Of course the flip side of this is that search engines and various sources on the wide internet aren’t necessarily any more true than what you might gain from ChatGPT. So maybe it’s not going to be such a huge impact after all.
> helping me with things that I’m already an expert on
This is a very good point. I rarely use ChatGPT (I do not have the reflex yet to go there and some early attempt were miserable) but it helped me a lot to write some bullshit text for a presentation.
It was very good at putting together the buzzwords and the result was excellent. Since there was also some actual content in it I could check for mistakes and there were two. These were not life threatening mistakes (in the context of a corporate speech) but still.
I develop a lot as an amateur and I need to try using it for that. Right now I only use Amazon Whisperer which is great at suggesting code that I do not have to type anymore - but I did not try to have it create longer code. You gave me some ideas, thanks a lot.
Finally, I would like to use it to correct my typing. I never learned to touch type (tries about 165 times over my 30 years career in science, IT and development) but gave up after 10 minutes. This does not bother me too much as I spend a lot of time thinking rather than typing but since I watch my keyboard I have typos. Grammarly fixes some but right now I see that I typed "thater" and not "rather" and i have to go back. I would love to be corrected by the most probable word in the context of the sentence (which is almost always the first suggestion of the spellchecker.
> I don’t think that is surprising, but I’m curious as to how much it’s inaccuracies will impact our society as people start using it more and more for various things.
I think humans can be amazingly knowledgeable in their own niches, but if you talk with the average person about your niche, there seems to be no limit to the amount of slight inaccuracies. I wouldn’t be surprised if ChatGPT is on average more correct when talking with it about random things.
My experience is that chatGPT is at least an order of magnitude less full of shit than the average human.
I notice it listening to NPR though that there is a class of people who are just so worried about safety that they are only able to process the world through the lens of potential dangers.
I know OpenAI is working on some sort of hardware integration with it (designed by Jony Ive). I almost think it’ll compete with HomePod, Nest, and Alexa.
I’m using ChatGPT 4 and love what’s coming back from it. Aside from limitations and some minor inaccuracies, it’s been great.
The mobile app on my iPhone was great on my last road trip. I told it was driving alone and I needed company and if could help me. Sure enough it offered to make small talk with me for about an hour before I got bored trying to push it more to keep me entertained. The entire time I kept thinking of the movie Her and how close this was to reality.
Something tells me Apple is going down this path by the next WWDC.
Ha. My response to this was a very visceral "eww".. but I also realise I'm not normal. Not super keen on small talk with humans either tho tbh.
I somehow fell off the tech nerd train and wandered onto a Luddite bus. I find myself supremely uninterested in the current AI hype, especially given I realised early on that the math was beyond me, so I felt I couldn't get a proper grasp "under the hood"
But it's also very clearly possible my first conversational interactions with these systems might happen without my even realising that I'm not dealing with a human. Might have already happened.
The voice generation is fantastic, sounds like a real person. The follow up, context specific question the AI ends with really keeps the conversation going. One thing I discovered is that even if you pause the conversation, it’s still listening and will take into consideration things said then.
I've used it to brainstorm a presentation while being out on a run. It does work quite well, the latency could be just a little lowe, but it does feel very natural. It's also finally fulfilling the promise of a truly conversational voice assistant, one that can actually hold a conversation and not just like thr glorified state machines that are Siri, Alexa et al.
If they add some memory persistence a la MemGPT and lower the latency slightly I think OpenAI has a huge hit on their hand.
I’ve been talking with ChatGPT to fill in gaps in my general knowledge. I pick a topic that I know very little about—the Holy Roman Empire was one that I tried—and I ask ChatGPT to tell me about it. I then ask about aspects of the answer that pique my curiosity, and the conversation proceeds from there. When ChatGPT’s answers are too long or when it asks me too many questions I’m not interested in answering, I adjust the custom instructions to try to discourage it. Two of the first conversations I had with it are here:
I could also learn about such topics by reading Wikipedia or other references, of course. But somehow talking interactively about them makes them more interesting and the information seems to stick with me better.
> I’ve been talking with ChatGPT to fill in gaps in my general knowledge. I pick a topic that I know very little about—the Holy Roman Empire was one that I tried—and I ask ChatGPT to tell me about it.
This isn’t an ideal use case for AI. You have no way of telling whether you’re learning facts or plausibly sounding fiction.
There’s a tradeoff between seeking perfect information and velocity. ChatGPT is right enough of the time that it’s useful (>90%). I use it the same way knowing that I might have some details incorrect.
It’s also unlikely to hallucinate the same way twice so if you ask follow on questions to verify the accuracy it becomes pretty straightforward to know when it’s hallucinating.
Yes, I'm fully aware of that. I'm not using it as a search engine or to learn specific facts, which are indeed sometimes hallucinated (though less so with GPT-4 than other models). Rather I am using to explore domains of knowledge interactively. It is that exploration and interaction that make it this an intriguing learning method.
It's especially useful, I think, for discussing and thinking about philosophical or controversial issues that do not have right or wrong answers.
> It's especially useful, I think, for discussing and thinking about philosophical or controversial issues that do not have right or wrong answers.
I like the idea that there are no right or wrong answers in philosophy. I have been enjoying dissecting Schopenhauer’s thoughts on hedonistic utilitarianism with chat gpt
This is why I use chatgpt as a type of Google. It's like a stackoverflow almost; generally useful but sometimes the person is wrong, but it helps me get leads.
I think it works pretty well as a programming assistant for that reason, because you can, well, just run the code that it produces to verify if it's wrong or not.
TBD on whether the architectural suggestions it suggests actually pay off in the long run but in my experience if you give it two solutions and ask it to weigh the pros and cons of those solutions it's pretty good at that. It's also decent at coming up with alternatives as well, but you have to prompt it to do that. It is totally happy to not consider any other alternatives unless you ask it to.
It's basically the same risk as learning about the roman empire from random youtubers. When I'm washing dishes and just looking for some edutainment, I'm not too worried about being misled on some random fact about an ancient civilization.
Related, an article on HN [0] a few weeks ago about a disturbed young man who used an AI chat app, which encouraged him on his plans of killing the Queen or England.
> Sometimes you just wanna share your unhinged thoughts with a friend
Oh my, meanwhile I'm too scared to type even the most generic of questions to ChatGPT. Keep in mind that your "friend" in this instance ends up being OpenAI, the company, made of bunch of people, people other than your friend.
Never heard of Youtube rabbitholes? I don't think I would ever converse per se but if ChatGPT is answering well and the inquiries build on each other there's a completely reasonable basis for that to go on for a while.
I’ve found Pi to be useful for this. The app has a voice to voice mode that works pretty well. The speech synthesizer has some tone and inflection variability, which makes it feel more like a conversation. The personality can be kinda over enthusiastic, but it’s way better than just a raw voice to text approach.
It’s pretty good for exploring a topic. You have to double check references, but the basic orientation is useful.
Plenty of usecases left, even if it won't be suitable for spicy political dialog. Slap Whisper and some TTS model on a local model and you can have an unfiltered experience.
Interesting detail about the TTS model generating a cough. Does anyone know what model/approach they’re using? I wonder if it’s transformer based like Bark.
I have seen lonely kids doing this with Siri, which I find amazing as the thing is incredibly bad. GPT3.5-4 are so much better that this must feel like a friend to some.
I do consider it a very useful tool in helping me with things that I’m already an expert on, I’m not sure how I would feel about using it for brainstorming. Maybe if you’re really just using it to speak out loud and not really listening to its responses, but if you’re using ChatGPT to learn or refine subjects that you’re not an expert on, then I think there is a very good risk that your output won’t be, well, great. But it’s not like it’s not happening, and I think it’ll be interesting to see what sort of effect that has on our society in the coming decades.
Of course the flip side of this is that search engines and various sources on the wide internet aren’t necessarily any more true than what you might gain from ChatGPT. So maybe it’s not going to be such a huge impact after all.