How do DDG AI and Brave Leo compare to MS Copilot in terms of quality?
Also, there’s no way to use Leo without the Brave browser, right?
There are many models for both. Big difference is they aren’t fetching the web like Copilot.
No, although everyone can use Brave Search AI - which is roughly similar to Google’s AI overview.
2 Likes
Edit : PR reopened.
Was wondering, should we remove DuckDuckGo ? They haven’t answered my email and I am really concerned about the “30 days (except for limited legal & safety compliance)” data retention period.
Plus, their marketing is pretty unresponsible. In this blogpost DuckDuckGo AI Chat: anonymous access to popular chatbots, they state
Even if you enter your name or other personal information into the chat, the model providers have no way of knowing who typed it in – you, or someone else.
This is misleading. Who is going to enter your name if not you ?
I am not saying Duck is not going a good thing. Allowing anyone to acess a chatbot anonynously is great. But we shouldn’t pretend the use of the chatbot is anonymous. If chats are stored up to a month, then they aren’t anonymous.
3 Likes
As pointed out they don’t search the internet. Yet, hugging chat might be an alternative though you should be aware the option to disable sharing chats for training.
But they have a far better privacy policy than Microsoft.
1 Like
I think this will probably come eventually. The thing is right now both the cloud and the self-hosted side of AI is evolving very rapidly, a lot of what is true today wasn’t true 6 months or a year ago, and may not still be true a year in the future.
In my opinion, making specific authoritative recommendations at this point is maybe a little premature. It sounds like you have some experience with the self-hosted LLMs so I’m sure you’ve experienced some of this first hand.
I think the TL;DR of the current unofficial recommendations would be:
- Where privacy is the priority, use a permissively licensed model (e.g. Mistral, Llama 3) that you host and run locally, with an open source privacy respecting frontend (E.g. Ollama).
- Where running a local model is not an option, consider using Duckduckgo Chat or Brave Leo, and understand that they can’t offer perfect privacy and your queries will be shared with the model providers in cases where Brave/DDG aren’t hosting the model themselves.
There is a work-in-progress draft proposal you can check out here. I want to stress that it is a proposal/draft not finished, and not official or necessarily agreed upon.
1 Like
This is super helpful, thank you. I haven’t self-hosted anything yet but that is how I expect I will interact with LLMs. Is Ollama your perferred frontend? As far as backends go, do you have a preferred default LLM (for general every-day stuff)? I also wanted to make y’all aware of this interesting project if you weren’t already aware of it. It’s pretty niche, but still pretty cool.
My NVDIA 4070 GPU is on an ubuntu machine so I guess that limits me somewhat. It also runs my jellyfin lol. May have to dual boot with windows 
I’m not really sure how to handle AI tools personally, since even putting privacy aside none of them have been developed in an ethical way. It’s basically like automated copyright infringement, and so it’s a topic I’m not super interested in. For the same reason that we don’t recommend like… pirating movies to avoid streaming services, for example.
I’d imagine we’ll end up writing something about it though, since it seems like people are more interested in following the latest fads than all that I wrote. We ended up adding cryptocurrency eventually too I suppose, albeit with a lot of warnings about how it’s a scam 90% of the time lol
Just from a privacy perspective, AFAIK there is currently no technology that allows AI to be processed in the cloud while also keeping your data secure,[1] and therefore it is only possible to use privately if you run it yourself.
edit: oh you deleted the post I’m replying to 
Supposedly Apple has just recently developed this for the upcoming iOS 18, but really only if you take their word for it ↩︎
2 Likes
All these AI chat tools are basically very convincing nonsense generators. I don’t see much value in writing about them.
4 Likes
It’s funny, I think we all feel this same way. I’ve resisted AI and LLMs for the very reasons you stated. But it seems like no one cares and this is what we’re doing now so I either need to figure it out or get left behind and become the old curmudgeon in the corner talking about the “good old days”.
Aside from all of the very valid reasons you stated for avoiding ai, the biggest one I know of is that literally no one ever asked for this. No one asked for AI. So why are we doing it? But, it seems this is what we’re doing now, so here we are…
1 Like
Yeah, It just said, “add an ai section” which is the subject of this whole thread so I deleted it because it was duplicative and just voted for the OP instead.
I understand the reluctance, but I think the way to reconcile this is to think in terms of harm reduction. (e.g. when I recommend using a youtube frontend as a more private way to watch youtube, I’m not endorsing YouTube or their business model)
All these AI chat tools are basically very convincing nonsense generators.
To be fair, it sometimes feels like most humans fit this description as well… 
But yeah, I think thinking of AI in its current form as “intelligent” is not a good idea. Its useful in many ways and contexts, but its not anything like what its hyped to be.
To me I view frontends as a means to access knowledge that is otherwise gatekept, which is a key aspect of privacy and anti-censorship. There is some duty fundamentally to make sure people can access any form of knowledge and do so in a privacy-respecting way.
Whereas in this case, no knowledge is to be found behind the doors of OpenAI
3 Likes
I was just using frontends as an example of a Harm Reduction approach. The obvious (non-privacy) related example of harm reduction is in the context of drug use/addiction. But another example more on topic here, would be blogposts written for Windows and Mac (which I think are not intended to be endorsements)
Whereas in this case, no knowledge is to be found behind the doors of OpenAI
The other day, I was testing Perplexity. I don’t recall the question, but it was privacy related. It actually paraphrased and linked to a Privacy Guides discussion, which on the one hand is good I think, out of all the resources I would hope a search engine or LLM would rely on for Privacy topics, this website would be in my top 10. BUT the mildly terrifying part… was it linked me to one of my own comments… 
This scared me not for tinfoil hat reasons, but because my deeply imperfect opinions and understandings given informally in the context of a discussion forum could be ingested and spit out somewhat authoritatively by an LLM (albeit a search-llm hybrid in this case). Still a little terrifying, I’d like to think I get many things right, but I am well aware I get many things wrong.
This has happened about 3 times now. My best guess is my specific writing style/how I phrase my prompts or maybe my obscure questions and interests lead the LLM to link me back to myself
3 Likes
WIRED concurs with your assessment lol
1 Like
I think we should recommend the most privacy respecting AI tools, those that are ran locally, as well as DDG and Brave Leo, and tell users to strictly use them for paraphrasing (i.e., language related purposes), and not any logic (e.g., reasoning, observation, fact, empirical) related tasks, and call it a day.
1 Like
Cory Doctorow calls LLMs garbage generators. I agree with that designation, with the caveat that garbage is not always useless.
For an on the nose example, look at Fallout, Cyberpunk 2077, and other games or shows set in filthy dystopias. They are set dressed with garbage; professionals are literally paid handsome wages to prepare fake garbage for their backgrounds.
Obviously, LLMs are not good only for generating literal garbage. My point is just there are many (non-fraudulent) use cases where factual accuracy is not important, and quick production of low-quality output is helpful. I see the technology as a useful tool for assisting (not replacing) artistic processes.
None of this is to defend the sorry state of the industry, where megacorporations are massively profiting from and copyrighting the fruits of their rampant copyright infringement. But just because we’re doing it wrong now that doesn’t make it intrinsically wrong to do.
2 Likes
Perplexity halucinates a lot but it can also gove me good answers sometimes.
2 Likes
While I disagree that AI outputs are “garbage” , it is true that some models are more prone to hallucination than others. For example, Google’s Gemini hallucinates a lot and invent statements from thin air. I will edit the pull request to mention that everyone shouldn’t trust AI models blindly and that you should learn which model is reliable and which model isn’t and even when you think a model is reliable, you should always fact check it for crucial information.
There is already a draft in writing, see
Anyway, let me know if you all agree on decision to remove duck.ai
Personally I think DDG and Brave Leo should stay listed. I think that non-local LLMs are closer conceptually to search engines than they are to standalone tools. And I think that realistically, most people who interact with LLMs do (and will continue to) use cloud based models. In that context both DDG and Brave are improvements over the status quo (OpenAI, Copolit, etc) because they add a (limited but meaningful) layer of privacy compared to using OpenAI, etc directly. It seems to me that the vulnerability of users voluntarily or accidentally undermining their own privacy by sharing personal information, is not conceptually different for Duckduckgo AI Chat than it is with recommended private search engines (or VPNs or Tor for that matter).
Afaik, DDG AI is the only private option being considered that openly available online. The other options either require (1) Some technical proficience and moderately robust hardware, or (2) or are locked to a single browser.
I could see demoting DDG AI at some later date if more private options emerge, but as it stands right now DDG AI fills a niche that the other options do not, and does meaningfully improve privacy over the status quo (by hiding IP, by having terms that prohibit chats being used to train models, or be stored for >30 days, and because DDG AI is accessible via tor).
I think probably the more important thing should be educating users that Locally hosted > Hosted by a “semi-trusted” 3rd party (e.g. Brave, if you trust them) > Hosted by an untrusted 3rd party (e.g. OpenAI) And formatting the page to promote a local-first approach, and more prominently display the local options.
But I’d be interested to here more about your reasoning, for not listing DDG Chat?
2 Likes
We also need to keep in mind that prompting matters significantly. Choosing the right model for the purpose with good prompting changes the results. It’s not only about hallucinations.
For instance, if you need to do generative tasks such as writing, generating sth, you should not use models connected to internet.
If you ask questions about the war in Ukraine and use chatgpt 3.5, of course it can hallucinate since its data cutoff is in 2021.
That’s also why Perplexity has focus modes. For generative tasks, you need to select writing. If you are concerned with academic sources, that’s another option.
For privacy questions, I generally select Reddit as focus, and it gives me incredible results.
But if you expect a very good result from LEO without a good prompting, than you have no idea about how to use AI tools rightly. (Since Leo free tier is powered by mistral and subpar to chatgpt-4 and other sophisticated models.)
At the end, they are tools, and the user needs to know how to make the most of it.
Edit: additional info about Leo is added.
2 Likes