69 comments

[ 2.6 ms ] story [ 151 ms ] thread
I always wonder how much I can trust settings like this, but this definitely looks like a step in a good direction!
Well, the FTC has already made companies delete models which relied on deceptive trade practices to acquire data.

So if they offer the setting and then ignore it, they could literally be forced to delete their model. Overall if you trust them to hire lawyers who understand that, and plan accordingly, you can trust the setting.

As a paying user, I wish the toggles were decoupled. For free users, I understand the lack of control. If you want to not have your conversations used in training, you also have to turn off chat history. Or I wish they would allow a toggle on a per chat basis, not on every chat. Step in the right direction though.
I think it’s aimed to be a dummy-proof version of “it’s okay to use chatGPT for proprietary code help because no one will be able to see that I did it, and it wont end up being regurgitated to someone else ten years from now.”

If the controls were more granular people would screw it up leading to another round of news stories like the “Samsung leak” one

Why bundle history and training in the same setting? Most users would probably want to keep history but not allow training on it.
I assume because it's a hassle for them. I've noticed that when busy they have issues with chat history right now. It could also be tightly coupled and just a pain to remove right now. We have no idea how the app was built.
I think you have answered your own question.

Somewhat more charitably though it could be that it’s far easier to ensure you never train on something you don’t store.

You just answered yourself: so people will keep the history on and they can keep training on the data.
Expediency. Perhaps they'll separate the two later on, because this is the fastest way to get to some easy user control right now. They're in a hurry to make this available while GPT is blowing up in usage/popularity with the public, to help reduce the blowback on privacy.

The US Congress for example just purchased a bunch of licenses to start messing with ChatGPT at their offices.

My hope is that if they aren't showing us the history then they literally aren't keeping it _at all_ (though they might keep it 30 days still for LE?).

That's probably just a dream and they just want to push you into keeping it on while they keep it for later just in case?

On the bright side this can be fixed with a client side extension :D

Beyond being a dark pattern, it's far easier to build a product that simply doesn't save things that could cause trouble. You can't have someone "accidentally" use a set of data if that data never exists. You can't have another integration accidentally fail to filter out "bad data" if it doesn't exist.

I've used this pattern in a few side projects. I never, ever have to worry about accidentally storing/moving/integrating data that doesn't exist.

>When chat history is disabled, we will retain new conversations for 30 days and review them only when needed to monitor for abuse, before permanently deleting.

They will still save all conversations. They are just pinky-promising they will not use it for further training and will remove it after 30 days.

But there is no way for us to verify that.

"Conversations that are started when chat history is disabled won’t be used to train and improve our models, and won’t appear in the history sidebar"

A huge step in the right direction to no longer have to use the API to get around training the underlying models -- including ChatGPT4.

Wait, were they using the data for training models? I thought that was the case in the first beta for gpt-3.5 long ago (sigh December 2022 feels like ages ago).

Also, are they training something like GPT-5 or just continuously improving GPT-4?

It seems they are refining 4 more than planning a 5 based on different interviews. I’m not sure if it matters though, could just be a branding issue more than anything.
Agreed. I doubt they're allowed to say out loud that they're releasing 5 soon and it'll be as much a step up as 4 was compared to 3.5. They can call it 4.1 but that won't change the outcome.
they are actively saying 4 represents hitting a limit point for the current models, and that further progress needs different model structure not just more size and training compute. a lot of people are hopeful for the fine tuning but i am sceptical that qualitatively better thinking will happen thru fine tuning, but we'll see
(comment deleted)
OT: Tried to use my paid account earlier today with no success -- login process got stuck in infinite loading. Anyone else affected?
Try clicking on Sign Up and then logging in again. There's a thread named Invalid Authorization Code in the OpenAI community forums and this tip from there helped me solve this same issue 4 hours ago.
I don’t think people appreciate what a sea change it is to have a major new technology launch with a subscription model and no ads.

Google hasn’t figured out how to make money outside of ads which is what’s been keeping them running in circles.

OpenAI is just going with the radical idea of charging people money for a valuable time-saving service.

Meanwhile while Meta and Google flounder Apple just keeps selling premium devices at a premium price and hasn’t laid off anybody.

The age of adtech is closing and the world is better for it.

> The age of adtech is closing and the world is better for it.

YouTube Premium exists and plenty of people still use adblock, or nothing at all. Most people are not going to pay for a subscription.

As for your other point, SaaS is a thing and has been for over a decade...

If anything I'd say ChatGPT is the beginning of next-gen adtech.

> If anything I'd say ChatGPT is the beginning of next-gen adtech.

You're 100% right, and people are gonna hate to hear it. It's like getting people to drink cough medicine.

2 decades ago, we were naive enough to wonder what online video streaming would look like. Now it's an ad-fuelled treadmill that has more promotional content than candid, sincere videos. We were also dumb enough to imagine online search indexing, same problem. Even your app store is not safe from being weaponized.

It's a sad pattern that is 100% fueled by the desperate profit-seeking of shareholders, and sustained by the free market in general.

I’ve got receipts from HBO, Amazon, Disney, Netflix, Curiosity Stream, and YouTube Premium that prove that wrong.
Precisely to the point I made elsewhere in this thread:

> Just don't pretend that you're funding a better world by buying the premium product - it's a bigger trick than advertisement itself. [0]

I'm not going to play the "media companies are the bad guys" card, but they are rent collectors. All of them do this - they take your money, and then keep squeezing. The only way to avoid being manipulated is to not play the game in the first place.

You still can (and should) fund the artists and creators you're fond of. The moment you agree to terms or a contract though, you've been made a business bitch. OpenAI is no different than the rest of them. Their Faustian bargain is worded a little differently, and the flowers on the table are a different color, but it's still the same shitshow.

As a reminder, YouTube also launched without ads. Just you wait, the squeeze is coming. I guarantee it!

[0] https://news.ycombinator.com/item?id=35706083

What are you talking about, there's an entire SAAS industry that exists. I would argue that actually subscription is the default, and "free to use/sell data" model is an outlier.
The biggest players are Google / Facebook, which are not charging for their core product.

You have Microsoft selling software, but I am not sure it is the default anymore.

Google absolutely does charge. Google suite products are sold to corporations a service. Individuals buy gmail/drive storage. There's Youtube premium, Google Cloud...

This is just an objectively wrong take.

> The age of adtech is closing and the world is better for it.

If you read most spending reports by companies, they would disagree. SMBs to large businesses all use advertising to increase the reach of their legitimate products. Your local barbershop likely runs Google ads to reach new customers so that he can run his business.

Furthermore, adtech lead to an infusion of capital into the internet, which led to many wonderful things for society. It's easy to paint adtech in a bad light when you only consider the worst things about it, which are being addressed.

“Being addressed”…yes, by a raft of laws around the world that are making it flat out illegal.
> adtech lead to an infusion of capital into the internet, which led to many wonderful things for society.

Source? Instant global communication is valuable enough by itself to attract massive investments and interest. I highly doubt technological advancements from adtech was in any way shape or form required for that.

> Instant global communication is valuable enough by itself to attract massive investments and interest.

If that's the case, with the advent of email in the early 1990s, why didn't the internet completely take off? It took another 15 to 20 years and only when Google and Facebook came into the picture, did we get open source technologies that spread throughout the world and is used daily.

@smashed, would also like to see your source here too.
“My local…” is less than a mile from home. Why does it need to connect with me via Facebook?

> Furthermore, adtech lead to an infusion of capital into the internet…

Fiat economic capital gave ostensibly free people agency; how generous.

The philosophy we mutter as if it gives rise to human need and doing. Lol

You should spend less time memorizing spoken gibberish that is used as a restraint on human agency. “If grandpa can’t grift on it, no seed round for you!”

How “free.” American culture is nothing but 24/7 propaganda to remind us to cling to the recent past.

No rich people are paying my way. Why are we insulating them from growing their own potatoes? What sort of stupid is this?

I’ll tell you what worked on me.

I feel like used it quite a bit for proofreading and suggesting different ways to say things. (Always subjective, never objective) And my total for all the time I used it was $6.

I feel like the value is definitely there and paid in. I don’t feel like the value is there when looking at most services.

I’m not certain how exactly, but I sense that the era of adtech and the era of Free Money were closely tied together.

The age of adtech has only just begun. Apple advertises their services in-device, and then doubles-down with App Store search result ads a-la Google Search. They know the ad kickback is endgame, they play ball.

What ads do is put wall between a human and some content. ChatGPT does not circumvent that wall - it might be able to tell you something about the content on the other side, but so can a Markov chain or an encyclopedia. The value of wrong information can only be so high, and after a certain point the training material becomes more interesting than the model itself. Your assertion that the age of adtech is closing is like saying that books are over because Wikipedia exists. History suggests that all these things will carve out markets for themselves.

I'm not an ad-apologist in any sense, but ChatGPT and OpenAI are not your savior. They're sycophantic magic-makers, watching your response to minor text illusions and imagining how they can further extort you. Now with Microsoft's support, great...

> The age of adtech has only just begun.

I learned a new word recently: Neuromarketing

https://en.wikipedia.org/wiki/Neuromarketing

Neuromarketing only works if my rhetoric does. I'm just appealing to your logic here - I have no de-facto credibility and obviously don't intend to cast ads in a positive light. I'm not trying to convince you of anything, at least in the traditional sense.

Look at it this way, though: the last 15 years of computing have been characterized by the commodification of computers and the proliferation of advertisement. We're trapped in a stalemate because nothing beats free -- nothing. Ads subsidize marginal visitor fees, which account for the majority of online interaction. It makes a lot of sense for people providing online content, and hardware sales only further churns this content wheel.

So... steal your media. Install adblock. Keep backups and care about what you own digitally. Just don't pretend that you're funding a better world by buying the premium product - it's a bigger trick than advertisement itself.

Chatgpt running on bing is ad tech, no?
Bing uses GPT, but Bing != GPT.
The shitty advertisements in Apple News would beg to disagree that Apple doesn’t think advertising is worth it, or the App Store advertising that has all but ruined organic search on the App Store, which I would think would be an even more dangerous road to walk down, but no they are absolutely jogging towards advertising with gusto.

And I fucking hate it as an apple customer for a couple decades now, I’ve watched the slow slide and I’m not saying it’s as an “if only Steve was still…” I just mean that the company no longer has product people making the ultimate decisions and consequently the product is suffering the slow death by business decision done based on what’s best for this quarter/year profit margins…

There's a kind of reverse money making thing going on with ChatGPT that I haven't seen articulated.

Never ever forget that every time you use chatGPT, you are making chatGPT smarter.

Right now, chatGPT is not priced for how much value it provides to you; it's priced based on how much value you provide to them. Because every time you use it, you are providing high-quality human feedback to the model.

It has already read the entire internet. So where does it get more content? For that, it needs you to use it. So why isn't it free? Better yet, why aren't they paying you?

Having you pay has several advantages:

I) By paying for it, you prove to openAI that the data you provide is more valuable, simply by the fact that you are willing to pay. Sort of like the counter-intuitive fact that if you pay to avoid ads, you are a much juicier target to advertisers. Not paying is valuable -- you're still a human after all -- but paying is even more valuable, because the people who pay provide qualitatively different data. They provide the type of data that has economic value. You can tell, because they pay! So this helps you improve your model far more than just giving it away for free, because you can treat human feedback from paying customers differently than human feedback from non-payers.

II) By paying for it, there is a paper trail that allows openAI to audit their customers. For spam or malicious use purposes. This helps OpenAI fulfill their safety mission.

III) By paying for it, you offset their inference costs.

I is HUGE. It's more important than II (which is, itself, more significant than III). For any AI company, their competitive advantage lies in the dataset. What is the dataset for a chat model? The humans who use it!

Engineering is important, and what OpenAI has shipped is obviously impressive. But it's not like nobody is capable of making an LLM inference frontend. The model? They may have some tricks that they're doing, but at this point, everyone knows how to make a GPT (bard, llama, ...)

Feedback is THE advantage that OpenAI has over everyone else. Feedback is THE advantage that MidJourney has over StableDiffusion. Feedback is THE most important differentiator for any company in this space. Those who have high quality human feedback (paying users) will win. What's the best way to get high-quality feedback? Make people pay.

This is an announcement of literally the opposite.

Of course they could be lying, but the claim is that if you talk to it with history switched off you are not making it smarter.

Ah, yes. I was more responding to the parent comment than the original post.

I would expect that there are two very different pricing that will be at play here.

You help the model -> Pricing determined by value you provide OpenAI -> $20/mo

You don't help the model -> Pricing determined by value OpenAI provides to you -> How much value is having an always-on intelligent writing assistant? Probably a lot more than $20/mo!

Maybe, but aws lambda doesn't work on value based pricing. The question is whether they get big enough to be a utility.
Yet a subscription doesn't buy any sort of semblance of privacy despite the lack of advertisements.
Note that OpenAI is under pressure from the EU to do something about using user data. https://www.technologyreview.com/2023/04/19/1071789/openais-...
The EU is so laughably ineffective and slow that I have no doubt OpenAi could do another 50 gpt-4 sized training runs with PII-pruned data before anything gets to court.

You already are seeing the likes of Adobe using legally-sourced imagery as a key selling point. I imagine the bigger momentum is going to come from companies who want to know what liabilities they have for using the tools.

Italy already banned ChatGPT, and Germany is considering the same. Whilst the EU itself taking a hard stance I agree will take a while, just the fact that EU member states are banning it is forcing OpenAI to take a privacy-preserving stance to prevent regulation on themselves occurring as quickly.
That they moved to ban as a first step is indicative of their regressive mindset on technology broadly. It properly encompasses why Germany is so backwards when it comes to tech and has so few leaders in the field, while supposedly being an industrial and economic superpower. They're way too slow, they're too dogmatic.

The approach doesn't work and it indicates a lot of bad thinking regarding new tech. Which explains why SAP is one of the biggest technology companies in Europe, and they're small compared to the US giants now.

Case in point: Salesforce sales 2020: $17b; SAP sales 2020: $27b. Salesforce sales last four quarters: $31b; SAP sales last four quarters: $30.8b.

Microsoft can unleash a hundred billion dollars of capital investment at AI. Which companies in Europe can and will do that? None. Instead you'll get Macron touting a $30 billion government collaboration program on AI, which will take ten years to get anywhere, and will accomplish nothing in the end.

There are few areas of tech that Europe isn't left in the dust by the US and China. The US is still the world's pre-eminent superpower and it fully intends to aggressively utilize AI to keep its advantages (and it will, just as it did with each era of tech). There's nothing EU bans will do to stop that, it merely provides further advantage to the US as it battles to dominate AI. The giant cloud leaders will have a huge advantage to begin with, so Europe is starting off in a bad way regardless.

Spot on analysis.

Only thing you forgot is Macrons powers of persuasion will surely just convince the USA and China to give France equal footing. Pretty sure after he single handedly ends the Ukraine war he’ll get right on it.

Color me doubtful. The big money for OpenAI is going to come from B2B, the majority of which is likely US-based.

The data they gather is also important to them as GP pointed out. They're getting novel data to train on that their competitors can't.

On top of all that, they're in an arms race with competitors. They're winning right now, but they could end up losing out.

If I were OpenAI, I'd have a long hard think about whether to play hardball here or not. Right now, it looks like they may get to pick winners and losers. I can think of at least a couple of countries that would probably be willing to make up the difference in revenue for a degree of control over the product.

If this happens then people in EU who have access to ChatGPT can potentially gain a big competitive advantage against those that can't.
The EU region managed to promptly fall behind the US in semiconductors, software, PCs, Web, mobile, cloud, and it's repeating history again with AI. Except this time they're going to pre-emptively cripple themselves, making the task of US companies to dominate the field that much easier.

Walling itself off isn't going to help. Europe is going to get crushed if they don't aggressively participate. The AI outside the walls will evolve very quickly and drastically, while everything inside the walls moves at a highly regulated snail's pace. Over a short amount of time it'll put the EU at a large disadvantage, not least of all in labor/talent (which will abandon the EU even faster than it already does, to pursue the best opportunities at the highest pay).

It's hard to regain trust once you loose it.
>before permanently deleting.

I don't believe this. I wouldn't believe this from my parents, my wife, my best friend.

Data is too valuable. It might not be used for ad purposes, but between training and legal purposes, I don't see data being deleted.

This is found on a PR page, not a legal document.

legit question, wouldn't they legally get in trouble if they don't?
If there's anyone here from OpenAI, I'd really appreciate the ability to "print/export/screenshot" the current conversation better.

I find it very frustrating to work through a long conversation with ChatGPT, then be stuck either:

* Copy/Pasting individual messages

* Printing - but only the current page

----

It doesn't need to be fancy. I just want a way to capture the entire transcript for easy reference in other documents.

I’ve gotten around this when I need by deleting the viewport containing the headers and moving out the large chat div, then I use one of the debug screen cap commands built into chrome on that div to save a giant tall image. Kinda bad, they should eventually just make print specific css.
As pure text, I've had success using the following code to be copied in the developer tools:

output = ""; document.querySelectorAll("main .flex-1.overflow-hidden .group").forEach(el => output = output + el.textContent); output;

removing all CSS rules on the chat window DOM node and its ancestors that say either

overflow: ...; or

height: 100%;

works for me

Asking as a Plus user: Why can't I toggle them separately?
Because they want your data, it is a dark pattern.

Also when you disable chat history / data collection they display a big box on the left side about "do you really want to do this" and a big green button to enable it again. There is no button to disable it in the same way, you have to dig into the settings.

Basically they were forced by the EU and using dark patterns to make you not disable it as the data you provide is very valuable for them.

Edit: You also have to enable this setting on every device you work on, it is not stored on the "account"..

The 3.5 prompt/model should probably be updated to inform users of their privacy more accurately.

> Me: is our conversation confidential?

> 3.5: As an AI language model, I do not have the ability to store or share any information discussed during our conversation.

> 4: As an AI language model, I am designed to maintain user privacy and do not store the content of our conversations. However, it's important to remember that I am a software, and your data might be stored by the platform you are using to access me.