Ask HN: How much longer will we have free access to ChatGPT?

The more interesting question is how long it will be until we have open, freely available models that are competitive with closed models. It's now clear that everything is going to be disrupted by generative AI - we haven't even begun to think through the consequences. I would much rather have a world where everyone has unfettered access to these capabilities, even if the risks of societal disruption are extreme, than have them controlled by the tech princelings of Silicon Valley.

We urgently need to fund efforts to create the technology to train and publish open models in a distributed way.

Stable diffusion, for Chatgpt, I like it.

Exactly.

It's time for the next Stallman to emerge.

These models are awesome but releasing them in the wild (I.e. not sandboxed) seems really irresponsible.

I hope these modes are prohibitively expensive to train so we can figure out a way to release them responsibly when they can be open.

What would you define as responsible vs. Irresponsible?

I assume along the lines of

    irresponsible = released to the public full of evil <insert slur here>

    responsible = kept in the walled garden of good <insert corp here>

I’m not sure I understand irresponsible vs. responsible in this context. Putting this in the hands of a few for-profit corporations but not releasing it to the open source community sounds extremely irresponsible.

What is irresponsible is not teaching the public what is coming. The press is currently not doing a good job of explaining what chatGPT is and how it works so that people can assess what to expect of it. The adoption and evolution of these systems will probably hit people like a ton of bricks in 2023.

Just curious, in a brief couple sentences of what do you think the first disruptive impacts to occur will be?

Wasn't OpenAI initially going for exactly this until they decided to make it proprietary?

(Hence the term "Open" in OpenAI)

The concern isn't so much the existing model, as impressive and promising as it is, but equitable access to future models.

When this thing gets 10x better, which it will, it will be capable of passing a Turing test against any human interlocutor. It will also pwn the field at any programming competition at that point, given further progress in areas described in the link at https://news.ycombinator.com/item?id=33914122 .

Massive economic advantages will accrue to the company or companies who control these models -- as in "Awesome, now we can fire a bunch of expensive programmers and engineers." A few weeks ago I'd have said that was the dumbest idea ever. Now it's obviously inevitable.

The C-suite shouldn't get too cocky, though, because it's possible that the lowly "language model" will be better at running the company than they are. New funding and business models can be expected to emerge on this basis.

Things are about to change in ways that seemed absurd to expect just a few years ago, and in areas where even the most well-informed people weren't expecting it. Full access to these models will be a very big deal, almost as big a deal as ownership of money and property is now. It will be very hard to compete with entities who have that access.

Can it get 10x better with the current approach though? As it is trained with content written by humans isn't that a natural ceiling ?

Unless you're religious or a UFO cultist, what non-human-authored content is available for anyone, including humans, to learn from?

Individual human knowledge workers aren't limited by the breadth and depth of available knowledge, the way they were up to about 100 years ago, but by their ability to assimilate and leverage it. That's also what limits the machine models. But the models will keep getting better, while we won't.

It's possible we're at the limit, but there's also no inherent reason it couldn't be better than humans. Its limits seem to be tied to conceptual capacity, not training material. We've seen it generate output that was its input combined in novel ways, and not found anywhere else. AlphaGo is a clear-cut case of this.

Of course, the history of AI is one of punctuated equilibrium, so it might be 30-50 years before the next 10x advance shows up, with nothing very interesting happening in the meantime.

But my thinking is that 30-50 years isn't that long, compared to the magnitude of the changes in the nature of knowledge work that may become possible.

I think it needs more work - its very good at information retrieval but it is also returning quite a lot of noise (specifically in the coding space). I have found it invaluable for helping me out with the poorly documented and sometimes esoteric Terraform modules I work with - but it rarely produces something that doesn't have an inherent flaw in. When questioned on it - it often says "yes you are correct - that is an error" and then produces an entirely different set of code.

Its hugely impressive but I think it still has a way to go before I would pay something meaningful for it, maybe the next release will solve a lot of things once the models are retrained now the public has had time to stress test it?

But it is more of a language model rather than a coding model?

If chatgpt cost money, I would pay for it. It's too useful for my work and one beautiful side effect I noticed is that it'll give me momentum.

That analysis paralysis you get sometimes before going on a long code run? Gone when the initial idea burst chatgpt gave me. It's a wonderful wonderful product. I'm floored at how good it is, and how useful it has become so quickly to my workflow.

I'd pay for it _today_.

If you don't mind sharing, would you explain how it is too useful for your work?

Not GP, but I used it the other day to start a simple web app:

Please write a javascript for testing pronunciation. The app should have the following features: 1. Store a static list of Chinese characters. 2. When the user clicks start, show the user a single character from the list, and wait for them to pronounce it out loud. Use the web speech API to recognize the character, and compare it with the character displayed. 3. If the characters match, repeat step 2. If not, give the user one more chance, before continuing. 4. Keep an internal record of which characters the user got right and wrong, and how many times.

Give me the corresponding HTML

<div id="displayCharacter"></div> <div id="prompt"></div> <button id="startButton">Start Test</button>

That's my main takeaway so far. It helps me get started. Having a piece of code that might be wrong but that I can refactor is easier than starting from blank.

Yeah, I'm also convinced that they will remove the model very soon. Thing is, it is very easy to trick the model into producing sexual or other content many or most would consider very immoral. Tell it to write whatever sick fantasy you want and despite OpenAI's best efforts to prevent it, the model will eventually comply. I happen to think that doesn't harm anyone cause it is just one person interacting with a language model, but the OpenAI team thinks otherwise.

If you're worried about losing access, and are willing to pay, you can sign up for an OpenAI account and use the OpenAI Playground with the 'text-davinci-003' model. I haven't tried using that model personally, but the model became available for public use around the same time ChatGPT came online, and they're both labeled as being in the GPT 3.5 family.

I would verify how similar the outputs are, but my trial credits have expired and I'd rather just use the free ChatGPT instead, hah.

davinci-003 isn't the same as the ChatGPT web portal. The web interface is a multi-model system.

I've found the davinci outputs to vary significantly compared to the web interface.

Then perhaps when ChatGPT is taken offline it'll remain available as a paid model on the API.

fyi https://twitter.com/sama/status/1599671496636780546 https://twitter.com/sama/status/1599671496636780546

It could (and probably will be) much lower if they switch to their own custom server builds . I guess he means "single digit cents per prompt" which is still pretty cheap

I have found it helpful, and we could automate some menial tasks. I am ok with paying a subscription fee as long as it is reasonable! After all, many of us have been paying for Spotify and Youtube Premium to avoid ads, and paying Notion and Airtable to simplify our lives. What's another subscription?!

I dont know but i hope they close it for a (reasonably cheap) access soon. Too many people are pushing the boundaries for the giggles which leads openAi to add more and more filters and censorship which makes it less useful down the road.

They’re using us to train the model even further so it will stay free. We’re giving them valuable data for free.

hi idk

seems like it would be a fumble.

most of the cost is probably training the model, right?

so just throw ads in the sidebar and see how much search traffic you can hijack.

or throw ads in the generated output

Based on how these things always play out, it's more likely they'll start charging companies to include tons of advertising materials into its training corpus.

"How do I make X?"

"You make X by ....

By the way, have you tried <promoted product>? People making X love it!"

i would pay for it if it is 2X better and no free alternatives

It needs to be put behind a paywall. It will unleash a massive wave of cheating among students, harming hundreds of millions of students' educations by making cheating super easy and undetectable. We age restrict cigarettes because they harm a children's health - ChatGPT beckons hundreds of millions students to go on autopilot at the expense of learning and growing. A pay wall and age verification is the moral choice. If you want to use it, spring for the $10 a month fee and pay for it.

as reasonable as that proposal sounds, i wonder if that will result in a whole market of asking questions on behalf of students.

As a standard of Ethical AI, all AI companies should require age verification as a pre-requisite for access and mandate blocking services that provide cheating via an API. Some may/will opt-in voluntarily, while others will not. The voluntary opt-in will set a precedent for legal action to mandate it for those that will not. Ultimately, running these large language models won't be trivial, and so you can gate keep access at the API level to prevent unethical usage.

This could lead to inequality, say a dad only cares that his kid gets a good scholarship for Football, having good grades helps with that so he has an account with chatGPT and lets his high schooler use it for all his essays, etc and college admissions essay... Now you have jocks who's parent don't GAF jumping ahead of kids who organically worked hard for their grades, etc...

Cheating is an artificial concept that doesn’t apply to numerous cultures around the world.

Maybe ChatGPT makes cheating easier because the entire idea of cheating was flawed to begin with. Being able to trivially transform some information and put your name on it isn’t a valuable human skill any more, just like switchboard operators or VCR Repairman. Perhaps Western culture needs to change to accommodate AI, not AI that needs to change to accommodate our obsolete measures of intelligence.

Maybe proctored exams (or oral exams) will become more prevalent, and the homework less important. Or the homework will be meaningful, independently of the tools. Like using wolfram alpha to solve equations step by step, this "cheating" device might actually make some students understand the subject better. Exciting times.