Dust XP1 switches to GPT-3.5-turbo, is now free to use

by ukuina • 3 years ago

99 50 3 years ago

They manage a cap on usage for predictability or have other means of converting early/often enough?

bcjordan• 3 years ago

Not sure about their policy/internals myself, but since GPT-3.5-turbo is 1/10th the cost of the previous model it's really hard now to run up a significant (to a funded startup) bill via legitimate usage — especially since at this phase any legit usage may be worth its weight in lessons about how the tool can be used. May be worth rate limiting / DoS protection but even then, could just cap OpenAI spending and spend the time to patch it if/when it happens (e.g. import @upstash/ratelimit or adding a quick Cloudflare rule)

bogwog• 3 years ago

Which raises a concern about privacy, since they're obviously going to be spying on your usage, and the tool has access to potentially sensitive information in your browser tabs.

spolu2• 3 years ago

Completely agreed with that. We'll revisit if needed but we don't expect it to be too much of an issue at this stage.

Our goal is to gather marginally more usage to learn more about productivity use-cases using LLMs.

Here's the email we sent to our users as we removed the paywall: https://twitter.com/dust4ai/status/1633484243228585988

zo12• 3 years ago

Chrome only? Guess I'll just have to wait/avoid for now.

stuckkeys• 3 years ago

Yeah, I noticed that also. I will watch from the sidelines for now.

meghan_rain• 3 years ago

How is this different from Edge the browser?

yunwal4• 3 years ago

Why would anyone trust a free-to-use app that’s grabbing all of the information in their browser and sending it to a 3rd party? Seems insane.

zo12• 3 years ago

I would - and I'm sure a lot of other people.

Are you genuinely curious, or are you asking because you're implying that the people who would use such an app are somehow not understanding something or not intelligent, or don't know something that you do? Like you have to prove your case too, as you're not immediately "right" in your statement. Sure there is some level of "risk" in doing this, but there is risk in a lot of things. It's like me asking people this:

"Why would anyone trust people with 4 weeks driving classes and a test with their lives on a road driving 80mph inside 2-tonne metal cages? Seems insane."

eternalban• 3 years ago

"Why would anyone trust people with 4 weeks driving classes and a test with their lives on a road driving 80mph inside 2-tonne metal cages? Seems insane."

We've run the experiment for n decades with billions of subjects and the results indicate it is actually not that bad.

We have -not- really even discussed, much consider, the implications of having a central system with machine intelligence designed to extract features, patterns, emotions, assumed motives, ..., having access to the entire digital lives of societies.

Is it insane to repeat the same mistakes? Not sure but it's somewhere close in the neighborhood. We could run it by a k-nearest algorithm and see what that suggests as a better category than 'insane'.

yunwal4• 3 years ago

You seem to be willfully ignoring the point.

There is no business model here that doesn't include "We sell your most private and sensitive data to the highest bidder". Because it's free, they can't make money any other way. And while for other browser extensions/software, you can at least audit the requests being made by the extension, this thing is sending away all your data because it has to in order to work.

cloudking• 3 years ago

Because your average user has no idea what you're talking about, nor do they care.

valine• 3 years ago

I installed the extension and asked it to reply to your comment, and thought the response was impressive. I myself am concerned about the privacy implications.

"Hi yunwal,

Regarding your concern about the privacy of Dust XP1, it's understandable to be cautious when it comes to sharing personal information. However, as mentioned in the Discord message shared by spolu, Dust XP1 only sends your requests to OpenAI and stores them for debugging purposes. They do not fetch or store anything else than what is required to process your requests. Additionally, Dust XP1 is open source, which means you can look at the code if needed. If you have any further questions, feel free to ask.

Best regards, XP1"

Veen• 3 years ago

So, they do store requests for "debugging" purposes, which means they have a record of everything you're doing with the extension and probably the page you're on while you're doing it.

I'm not sure why this is necessary. Surely, they're just scraping the page and using it as context in a ChatGTP prompt. They don't have to proxy it through their servers to do that.

sorokod2• 3 years ago

From the site:

Privacy

Only the text content of the tabs you select and submit are sent through our servers to OpenAI's API. Cookies, tab list, or non-submitted tab content are never sent.

A bit less than what the model came up with.

Also, are stements made by an llm model about staments made on discord legally binding?

deet2• 3 years ago

I've been following XP1 and I don't think they have nefarious goals with this. The extension is even open source (though not sure if the proxy is). Check it out here: https://github.com/dust-tt/dust/tree/main/xp1

I'm working on a somewhat related product (except bringing this assistant capability to all apps on your computer, all browsers, and using mostly on-device ML...waitlist in my profile in case you're curious)

What we've discussed internally is having two modes for the cases where we do need network connection:

1) A turn-key, use-our-OpenAI/HuggingFace/whatever proxy that doesn't store anything, just adds our token and pays for it on your behalf

2) Bring-your-own key for each service

The fact is that most users who just want to use these kinds of productivity tools might not have their own OpenAI/Azure/etc account, so offering option 1 and even defaulting to it is right for most end-users.

I think XP1 is making the right call here with this default, though offering #2 would be nice!

(edit: added Github link to XP1)

yunwal4• 3 years ago

Does the extension being open-source actually help here? I don't have time to look through the whole thing, but essentially it sounds like I can verify for myself that all of my requests are being sent to dust for them to... store for debugging purposes and not make any money while paying everyone's OpenAI fees? Doesn't seem believable to me.

sureglymop2• 3 years ago

I had a good experience testing out "gpt-3.5-turbo". Made a bot for the Matrix server we use at work so that co workers can query and work with the AI assistant collaboratively and in any group chat. Works pretty well so far and is fairly inexpensive. The markdown sent by the AI gets rendered nicely in Element and Conny. Want to open source it soon.

apetresc• 3 years ago

Did you somehow get it to play nice with multi-party chats, or does every user have their own private 1-1 room with the bot?

sureglymop2• 3 years ago

Both work. I have an SQLite db where i save "rooms", "sessions" and "messages". So multiple users can contribute in the same chat session, where i pass the whole conversation history to the AI as context. Then there are commands to switch and delete sessions. When the bot is mentioned in a message, it gets the answer from the API in a different thread and shows a writing indicator. It also plays super nice with the Element Matrix client which supports tables and codeblocks. The openai API sends back markdown which i can parse into html and send like that. I want to clean up my code a bit and then open source the bot.

2Gkashmiri• 3 years ago

please bring firefox versions from the start. You are only playing into the chrome monopoly by treating firefox users as second class citizens then complaining there is no "demand"

dang• 3 years ago

Show HN: XP1 – A GPT-based Assistant with access to the browser Tabs - https://news.ycombinator.com/item?id=33970570 - Dec 2022 (9 comments)

bfeynman• 3 years ago

Certainly cool, but it's getting a lot to see just how good GPT is at generating utterly useless cheap content. Making content is a moving target, and GPT by nature produces uncannily generic generations.

shp0ngle2• 3 years ago

I realized.

I could use GPT to generate Wikipedia articles about random villages and people I know nothing about.

Sure, those will not be at all correct, but who checks correctness on wikipedia, especially about random people and villages?

GPT, give me 10 paragraphs about New East Lake, Michigan

sorokod2• 3 years ago

Would that not qualify as an act of vandalism?

Centigonal• 3 years ago

You really think someone would do that? Just go on the internet and tell lies?

schroeding• 3 years ago

You can, and some people will do exactly that.

Ideally, new versions of GPT (and other LLMs) will be also trained on the newly added, mostly hallucinated data when using Wikipedia as part of the training set, baking those past hallucinations into the model, allowing other users to get additional hallucinated facts that fit the existing fictitious data, slowly creating a new "GPT cinematic universe" for e.g. small towns in Nebraska.

And if people fact check it with Bing, there is a good chance it will say "Yeah, that checks out, here's Wikipedia and <some Wikipedia mirror> as a source". A creative derivation of XKCDs citogenesis[1] :D

[1] https://xkcd.com/978/

benopal64• 3 years ago

Well that is true. However, I used Phind and was able to get 10 paragraphs about New East Lake with sources! https://www.phind.com/search?cache=8ff1dcc7-7587-4df8-9e46-6...

At the same time, it uses Wikipedia as a source as well as marketing and blog sources so Phind results should be carefully critiqued as well.

shp0ngle2• 3 years ago

I got ChatGPT to hallucinate actual sources that I can readily copy-paste on wikipedia:

“Give me 3 citations about New East Lake, Michigan, that I can use on wikipedia, in mediawiki syntax. Use actual book names and URLs instead of examples.”

Note that this town doesn’t exist. But now I can pretend it does.

Note, I don’t want to do this of course. But it shows that you really need to check sources on wikipedia, especially nowadays, when it’s so easy to generate realistically looking nonsense.

pornel• 3 years ago

You can already do that by hand (hoaxes have been done, e.g. https://en.wikipedia.org/wiki/Henryk_Batuta_hoax)

The only thing GPT changes here is speed at which you can do that, but if you start mass-spamming with such articles that will actually make it easier to spot.

valgaze• 3 years ago

"- Does not work with Google Docs and Google Sheets (work in progress)"

I wonder if security issue or because the canvas/ui is special

sva_2• 3 years ago

Only the text content of the tabs you select and submit are sent through our servers to OpenAI's API.

But they're still sent through your server for pretty much no reason?

sodality2• 3 years ago

So they don't have to send the API keys to the client, where they can be abused

lordpankake• 3 years ago

OpenAI could begin to accept pre-signed URLS to accommodate this use case. Would reduce the amount of people with access to your private data by 1.

sva_2• 3 years ago

Ah, I thought you'd provide your own key.

If it is free, how do they not make losses?!

sebastianconcpt• 3 years ago

This kind of apps make me wonder... is AI going to learn how to treat people better than people?

ur-whale• 3 years ago

is AI going to learn how to treat people better than people?

The bar isn't very high TBH.

m3kw9• 3 years ago

Thought I saw Dell XP1 switching to use GPT 3.5 turbo.

When you come out of beta, probably need a rebrand

fire• 3 years ago

Unfortunately Dust.tt doesn't use a modern email validation method and so fails on "modern" TLD's like .media, making it impossible to sign up for an activation key with it.

aa-jv• 3 years ago

Information needs to start having a "Free of AI-generated content" seal of approval, akin to the GMO battle being fought elsewhere.

I mean, we already process high fructose digitally-altered content, that ship has sailed - but we ought really start to having an "AI / not-an-AI" tag we can put on things...

(The moment I detect you've sent me an AI-created email about our project, I will replace you with a shell script.)

/disclaimer: I may have generated this comment without any intelligence, whatsoever.

mdrzn• 3 years ago

Tried it in this page with the following prompt: "read the comments and write a summary of what people think of this product"

GPT: Users are impressed by the capabilities of the product and its ability to generate content quickly. However, some users are concerned about the privacy implications of the product and are hesitant to use it due to the fact that it sends personal information to a third-party. The fact that the extension is open source helps to alleviate some of these concerns, but others remain skeptical about the intentions of the developers. Overall, there is a mix of excitement and caution surrounding this product.

mmacvicarprett• 3 years ago

"Mr X is an experienced software engineer with over 500 connections". ChatGPT learns from the best bullshit as well.

Dust XP1 switches to GPT-3.5-turbo, is now free to use

Related Stories