Ask HN: Will you be paying for ChatGPT+?
I was just offered the paid subscription yesterday for $20/mo. I went ahead and signed up for this month just to see if bringing back the late-December speeds and uptime would feel compelling.
I'm still not really incorporating ChatGPT into my daily workflows at all, but I give it lots of "for fun" prompts all day. So, I am not confident I will stick with the subscription past a month, but I hope the payment pushes me to find more practical uses for it.
How about you?
55 comments
[ 3.8 ms ] story [ 121 ms ] threador what library would I use to XYZ in ABC language?
I'm really asking. I've never played with co-pilot
you can write a comment above a function explaining what the function will do, and it will give multiple suggestions of what it thinks you're trying to do.
like chatGPT, a lot of code it gives you is not really exactly what you want (or might not work at all), but it offers a really good boiler plate to make small changes on to achieve what you want. it's definitely removed a lot of tedium for me in some ways.
first example that comes to mind is http requests. i use the same kind of style of writing a request function with error handling, and when i start a new project and start writing a new http request handler, it already knows how i've done it in the past and can offer an almost completed version with whatever changes i need added to it for the new project.
-Emily
Yeah, I'm not opposed to paying for something like this. And I like ChatGPT.
But at least for now there's no way I'm realizing $20/month of value from ChatGPT compared to other alternatives like Googling.
Agreed. We get into moods where we would appreciate access to an advanced language model like this, but those moods absolutely do not make up the majority of a month and they also certainly aren't worth $20.
We don't use ChatGPT for problem solving. Mainly brainstorming, or just conversations.
-Emily
"Promoting violence" or some such.
I also got denied (though not warned) for asking it to describe a character's love interest. It responded with some Victorian maiden aunt crap about how it couldn't generate "sex scenes". I didn't even say sex! I said "love interest". Jeez Louise!
You should spend a couple days and try and use it to solve all of your problems, either directly or indirectly. The limitation is in how we are failing to utilize the available tools.
ChatGTP isn’t AGI yet, but it’s incredibly powerful today. Even if GTP-4 wasn’t around the corner, I think the current state of LLMs would be able to power many new businesses and use cases over coming decades. Even that AI Seinfeld clone could be turned into a choose your own adventure streaming Netflix with enough refinement.
As for "isn't AGI yet": Language models will never be AGI. AGI requires internal thought and understanding. Some people have become convinced that this is the case by seeing the output of a language model, but there is no sufficient level of advancement that can turn a language model into an AGI, due to how they are currently implemented.
If OpenAI was somehow able to produce an AGI at some point, it would no longer be a GPT, because GPT stands for Generative Pre-trained Transformer, a type of language model.
-Emily
As for the rest, hard to say, because we really don’t know what puts the G in the I inside the coconut. Hard to measure internal thought and understanding. Hard for you to even prove I have it.
...we technically don't always assume that people we speak to on the internet identify human, since we don't ourselves, but we at least know that all host bodies (that we know of) are physically human. And we also know that all consciousnesses we've spoken with are "human" in nature as in "formed in a human brain".
> Hard to measure internal thought and understanding. Hard for you to even prove I have it.
Indeed, but brain scans exist to show mental activity in response to all sorts of things, including unprompted internal thought. You've heard of the EEG test for ADHD, right? You place someone in a room all by themselves, put some electrodes on their head, and leave them there for around 25 minutes. Then you measure brain activity to see if it looks... uhh, ADHD-y. I don't know what exactly they check for, but I'm pretty sure it's some certain type of uncontrollable mind wandering that people with ADHD suffer from.
With a language model, you are computing the mathematical transforms that result in its output. And there's no evidence to show that there is any sort of internal thought that occurs there, especially since the entire thing is re-run from scratch for every token. Any "thought" or "understanding" can at best be encoded by mathematical weights in the layers, or something, but it's definitely not the same sort of understanding that biological humans have, or even most animals. The way they run is just fundamentally different.
And obviously a mere product of consciousness such as an internet comment can be difficult to tell apart from the output of a hallucinating language model. :)
-Emily
https://www.youtube.com/watch?v=d1z8YUs2FLc
What is this an analogy for?
-Emily
See also: https://en.m.wikipedia.org/wiki/Blindsight_(Watts_novel)
When I think of AGI, I think of some sort of agent that can converse and interact in the same way that a human can, that has the same sort of curiosity, imagination, and internal thought that a human does. But, as it turns out, that's not actually required or even implied by the term "AGI" itself.
-Emily
And then again, maybe math is not all you need. Unknown unknowns, possibly unknowable.
No, for the same reason why asking ChatGPT what it is thinking or feeling does not indicate that it actually thinks or feels that way. ChatGPT is a language model designed to generate outputs that mimic its training data, therefore it will learn to mimic the expression and communication of understanding that it does not itself feel.
This is why language models are so convincing when they hallucinate, because hallucinating is actually their default state, and the hallucinations just tend to line up with some of the facts contained in the training data, but not always. Therefore they can state falsehoods just as convincingly as they can state facts.
> I’m not saying you are incorrect, just pointing out that we’re not exactly sure what we are looking at in those brain scans anyway. Is a thought a mathematical calculation being run on a wet computer?
Thought is sort of an ongoing electrochemical chain reaction, though beyond that I indeed don't have much I can say for sure. It's certainly not a mathematical calculation, though it could possibly be simulated by one.
-Emily
So far my predictions on OpenAI for 2023 [0] have become very accurate and expected but happened surprisingly early. From their investment [1] to pricing the ChatGPT API [2] and now I will expect OpenAI to start competing against their partners as I said in [0].
The only way to compete against OpenAI is an open-source version of ChatGPT.
[0] https://news.ycombinator.com/item?id=34201706
[1] https://news.ycombinator.com/item?id=34321066
[2] https://news.ycombinator.com/item?id=34368641
> The only way to compete against OpenAI is an open-source version of ChatGPT.
Luckily, this is happening. Current efforts are probably ~1 year behind.
It has been a rare day in the last couple months where I haven't found some useful way of applying it.
Just last night I was showing a friend who works in AutoCAD how he can ask ChatGPT to help him automate calculations in his drawing (get perimeter/area of a polyline). ChatGPT started spitting out AutoLISP and answering questions questions about how to get it running on his system and his mind was blown.
We didn't verify that the code would run, but my buddy hadn't even considered the possibility since he doesn't code. But now he has a starting point to hack on, and with a combination of more ChatGPT prompts and Google searches, I'm confident he'll be well on his way.
Considering what I pay to have a decent meal or good time in NYC, $20/mo is a bargain for unrestricted access to such a tool. I'm also hoping ChatGPT+ users will get early access to things like the official API.
- Model-side censoring, where ChatGPT will say something about how it's a language model trained by OpenAI and can't answer some query you give it.
- Client-side censoring, where each response is submitted to a moderation model and removed (client-side!) if it is explicit or hateful.
The first can be defeated with "jailbreaking", the second can be defeated with browser extensions. But this is extra work that some people don't know how or don't want to go through.
-Emily
I'm mostly curious about the future of it, though. The next iteration (GPT4?) and beyond.
https://davidrozado.substack.com/p/openaicms
The article you linked specifically focuses on ChatGPT’s ability to generate hateful content.
It’s hard not to read what you’re expressing as “I won’t pay for it until it’s adequately able to generate hateful content”. There are a number of other domains where you might find it useful.
To me that doesn’t “feel” like it was learned from training data but it feels more like a bias was programmed into the interface. Likely with good intentions but it’s catching edge cases that even the authors probably didn’t intend.
I do believe it will get better. I don’t think ChatGPT is a political secret weapon of the left. With Covid, the last US election, civil unrest, there is a lot of fear about putting out false or negatively insightful information. It’s a tough problem to solve. I don’t think ChatGPT should be responsible for solving it but they should try to make it feel less biased on the overly aggressive parts.
I see far more instances from posters who are unhappy that ChatGPT won’t say something negative about women/POC/LGBT people, but will about men/white/straight, etc.
Personally, I don’t mind that ChatGPT is protective of systematically oppressed groups.
I have yet to see an example where it convincingly leans left on any sort of nuanced argument. If you or anyone else would be able to post an example, I think it would be an interesting addition to the discussion!
The article analyzes the encoded bias of ChatGPT. The system allows you to, for example, make fun of conservatives, but not liberals. It allows you to write negative stories about Republicans, but not Democrats.
The people that created the system have also encoded who is worse off than whom based on what someone looks like. That seems like a bad idea to me.
Using it to write code... that's probably OK. Using it to express ideas? That seems problematic when the system is designed to favor a particular ideology. It is not that the system should output hateful things. It is that it specifically endorses some hateful things as being just fine.
If you extend that into untested biases, what the system favors, you're left with a generator that constantly applies an ideology in favor of some things and against others, regardless of the reason for talking about those things.
"There are no unhappy people in Oceania..." "War is peace. Freedom is slavery. Ignorance is strength."
It would be wonderful to be proven wrong, but I have a feeling that we’re unlikely to see a perfect text generation system with no bias anytime soon. What perfect system exists in the world?
You seem to specifically have an issue with the Democrat vs. Republican leaning demonstrated in the article you linked. If the small bias that exists disqualifies your support for this wildly useful tool, then I suppose there’s not much that can change that. We’re left to hope that another model pops up soon whose biases are more favourable to you.
Other than the D/R bias, other biases seemed to correlate pretty well with protecting oppressed groups. Once again, under the premise that bias will always exist, I personally think that OpenAI has done a pretty good job at channeling that bias towards protecting groups that tend to be the subject of most hateful content.
But then, I’m also coming from a position of being supportive of good moderation. That’s why I like forums like Hacker News, and I suppose why we’re having this conversation at all. If I only saw utility in raw, unbiased, unfiltered text, I’d probably go read elsewhere.
Think I might add co-pilot and kagi too while I'm at it. The extra 40ish a month while not insignificant isn't all that much considering I spend that on coffee...(Should reduce that lol)
Or maybe I'll wait and see what Google roll out. :-)