Hi all, I'm the author of this post (and the CEO/Founder of DuckDuckGo). Happy to take a few questions about how DuckAssist works, but to clarify what I think may be some confusion from the other comments, we're not just asking a LLM to summarize Wikipedia from its training, nor did we train an LLM on Wikipedia -- instead, we use our own Wikipedia index to pull out the relevant sentences, and then ask the LLM to verify the answer is in the sentences and then: if so, output it simply; if not, try to answer a similar question, and if that doesn't exist, then do nothing.
It can still be wrong for a variety of reasons mentioned in the post, but this is a different approach and in our validation testing (using banks of questions with known/unknown answers in Wikipedia) we know it works reasonably well, and have a lot of ideas of how to make it work much better over time. This is a beta after all, and we have plans both for improvements and additional features.
As noted in the post, the triggering of the feature right now is quite low and so it would generally be expected it would not trigger for most queries. That also means it is easy to find examples of where it would/would not trigger based on changing a word two. To the broader point of potential disparate triggering for these type of searches in particular, we have not studied that yet, but duly noted as something we should look at.
Thank you for paving the way for user privacy in the search arena. The ability to generate tons of content with LLMs is going to be overwhelming by the summer. The worry that I have regarding future search is that we need a way to defend against SEO generated by these LLMs and non-useful LLM outputs in general. Is there a way that LLMs can be used to help in this effort? I feel like we are going to have to go back to altavista style directory results to defend against LLM generated content. Any plans in this area?
Fighting SEO spam is a never-ending arms race. I agree with your sentiment that is is likely to get worse and we need to be prepared for that. I'm not sure yet if LLMs will be significantly useful themselves in stopping it.
> and then ask the LLM to verify the answer is in the sentences
How can you ask the LLM whether the answer is in the sentences without leaking the user's question?
Let's take a concrete example: if I ask "Is ADHD caused by non-genetic factors?", what does the LLM see? It's not clear how OpenAI does not learn from this example that user X is more likely to have/know someone with ADHD.
From the post, "As with all other third parties we work with, we do not share any personally identifiable information like your IP address. Additionally, our anonymous queries will not be used to train their AI models. And anything you share via the anonymous feedback link goes to us and us alone."
> by asking DuckAssist to only summarize information from Wikipedia and related sources, the probability that it will “hallucinate” — that is, just make something up — is greatly diminished
This cannot be asserted and dismissed so easily.
I would not be surprised if it returns content that it thinks would have or should have been in Wikipedia.
The more you constrain it (e.g., "only answer truthfully") the more it bristles and tends to react in the opposite way, because it can't actually do that.
I asked it "what is a memory model in computer science" to see how it handles something slightly obscure. Its answer:
> The multi-store model (also known as Atkinson–Shiffrin memory model) is a memory model in computer science that describes the interactions of threads through memory and their shared use of the data. It allows a compiler to perform many important optimizations.
> More info in the History and significance section of the Memory model (programming) Wikipedia article.
It seems to think that the Atkinson–Shiffrin memory model is somehow related to computer science, which it is not. It's a model of human memory. And that article it references does not once mention the Atkinson–Shiffrin memory model. At least it's easy to verify.
Maybe it thinks linked sub-references on Wikipedia are more relevant than other information on the computer memory pages. Or maybe it just took "memory model" without the computer context.
Yeah, I think the problem is that the memory model (computer science) article is too short, so the model ended up spitting out information from something seemingly related, i.e., human memory models.
I really do not see any wisdom in turning search engines into answer engines. Yes, they're bespoke toys today, and you can still search the web alongside asking a LLM for an answer. But the day this becomes the default, it just becomes a countdown to the day "search" is depreciated entirely. And then rather than "doing your own research", you get to pick which Big Tech silo you trust to be the objective arbiter of truth. What a great future we're hurtling towards at breakneck speed...
I can't imagine the disasters that would result from getting answers from a search engine instead of human-written documents. Maybe it's an inconvenience to get the wrong showtimes for a movie but what about when you're asking about the weather on your long distance hiking trip? Whether it's legal for you to carry your gun on your trip downtown? This is going to cause so much damage.
It clearly can given that some people still believe - or at least pretend to do so for the purpose of spreading propaganda - that people have been drinking bleach even without the 'assistance' of LLM propaganda machines to spread these false narratives.
And that horse dewormer? That is just ivermectin in a paste base (e.g. hydroxypropyl cellulose and castor wax) where the human version is ivermectin with a filler-diluent, probably lactose, microcrystalline cellulose or starch. There is no special 'veterinary ivermectin factory', just some factories which use the same ivermectin as used for humans which they then mix with a paste base. I'm not convinced ivermectin actually is effective as a prophylactic against or treatment for SARS2 but it is known that it shows antiviral activity. I am fully convinced that it has far fewer and far less serious side effects than the experimental vaccines which were pushed past the traditional safeguards. I am reasonably convinced those vaccines were not effective against the later strains of SARS2 - Delta and Omicron and whatever came/comes after - so it is likely that the 'boosters' actually caused more damage than they have avoided given the fact that these were administered when the bulk of the infections were caused by Delta and Omicron.
It will take a long time for the narratives (plural, both from the 'follow the leader' as well as the 'my body my choice' side of the population) to subside so that something resembling the objective truth may be discerned so we´ll just have to be patient until then.
Being able to search a ranked index of sources is extremely powerful but thats nothing compared what is possible when you can outsource intelligence.
In other words, when you can outsource the part of cognition where you need to understand the problem, go through the ranked sources, find the right answer and digest it in right way.
This is unbelievably powerful, comparing it to horse carriages vs cars does not do it justice, we are talking about Jarvis in real life!
P.S. I have some colourful opinions about the kind of thinking thats leads to your comment but I'll keep that to myself since its likely to attract scrutiny or censure
Who Are You gives a pretty nice combination of search results and LLM responses to questions. I find it helpful to see that the responses also include community sites like Reddit, so in effect you can move from LLM to more-human responses or discussions fluidly in a lot of cases.
That is the same playbook by which Microsoft plays this game.
MS: "IT IS OPEN FOR EVERYBODY TO TRY!"
Me: Ok, let me try it.
MS: Ahem, ok, just give us your email
Me: What? Ok, dammit, here it is.
MS: Ok, now install our browser.
Me: What?? Ok, done.
MS: Ok, now allow us to track your location.
Me: What??? No!
MS: Sorry, something went wrong.
Me: What?
MS: Sorry, something went wrong.
Me: What about the chat thing?
MS: Sorry, something went wrong.
OpenAI drafted a paper proposing that in the future, platforms require proof of personhood. Plus it's MS, so they're going to slurp as much as they can get away with.
At least MS doesn't demand an active cell phone number through which they can physically track you everywhere you go (not just where you are right now), like ChatGPT (VOIP and landlines not accepted, only cell).
You listed examples of free software and non-free software.
The former is much more likely to find a home on my devices.
I've found it limiting to not use WhatsApp, but not that much. It's kind of like if I knew someone who is known to stalk people, but we also have many friends in common. Maybe some friends won't come to my party, but I'm still not letting them into my house.
Mostly correct. Browsers can block ads, trackers, Javascript libraries, and elements at will using extensions. "Native" apps for web services, such as these Electron-wrapped clients, have full access to the trackers / analytics endpoints, and more. That's a big nono.
It's unfortunate that you need a healthy dose of cynicism to make it in today's society. A free tool seemingly created solely for your convenience trips some internal alarms for me as well.
What do you mean "at least here"? What's wrong with other usage? They are using OpenAI so they haven't trained it themselves and is just inputting the Wikipedia article as a prompt. I don't see how that is more respectable than literally every other service built on OpenAI.
DuckAssist answers questions by scanning a specific set of sources — for now that’s usually Wikipedia, and occasionally related sites like Britannica — using DuckDuckGo’s active indexing. Because we’re using natural language technology from OpenAI and Anthropic to summarize what we find in Wikipedia, these answers should be more directly responsive to your actual question than traditional search results or other Instant Answers.
You misunderstand how this works. They do not train it on Wikipedia. The Wikipedia article is part of the prompt. You can do this yourself by copy-pasting a Wikipedia article into ChatGPT and ask a question about it. The underlying LLM is still GPT-3 and trained on everything on internet.
I think the point is that it should only be used to summarize wikipedia information, and thus harder to use for other ethically murky purposes? But that remains to be seen; I haven't tried it yet.
If you enter a question that can be answered by Wikipedia into our search box, DuckAssist may appear and use AI natural language technology to anonymously generate a brief, sourced summary of what it finds in Wikipedia — right above our regular private search results.
That just means the initial prompt says something like, "You are an expert Wikipedia research assistant. You can only reference material found on Wikipedia. Try your best to help the user by giving summaries of Wikipedia content. Do not reference material from other websites!"
The model is still trained on a broad corpus like ChatGPT3, i.e., the Internet. The corpus of Wikipedia alone probably isn't enough training data.
I'm sure it makes sense to start with wikipedia and instant-answer type searches from a tech standpoint if that's what you can iterate on best, and if you've got ideas for preventing confident falsehoods.
It's also the type of search where I'm sure this addition is least valuable. If I'm searching for a discrete fact that I know exists, stuff that you find on wikipedia, that's the type of search that still works fine. Instant answers are convenient but not necessary.
Often when I'm trying to solve a problem I need to break it down into more generic, smaller components, discrete pieces of information that I know exist, then search for those and recombine/recontextualize the information I got. Or branch out into new searches based on anything that was an "unknown unknown" before the initial search.
I'm hoping LLMs can help with both of these things. Allow me to search for something closer to the problem I'm actually solving rather than requiring me to break it down into components; do the recombination for me and do some of the branching for me.
This just seems like a faster way to infect people with incorrect information. I mean every result returned will soon enough just be AI generated articles with that incorrect information, and you are saving us a step from clicking, but I'm getting sick of seeing this stuff.
If you can't do the job right and make sure the answer is right, then you shouldn't be publishing anything, or you are just damaging people and eroding trust. Returning search results in a fuzzy way I can't understand or fault you for is one thing and gives you plausible deniability, but not this. I'm all for Google alternatives (I use DDG), but this is definitely a turn off. Put the time into finding the right search result that actually has an accurate answer and just point me to that as the first result. Hell add a sublink that points to the actual answer on that page below the main result, but get rid of the gimmick that is AI.
It seems like it's meant to be an extra feature and not a replacement for search. This is a good thing and will help DDG reduce dependence on Bing. Of course, maybe dependence on Wikipedia is not so good if you care about accuracy.
And anthropic, which has google investments. They are bound by licensing not to block some of the Microsoft analytics, but the same might not be true of which NLPs they use.
Yes but certainly a good chunk of the companies finite resources will be shifted towards this effort vs improving upon search. No judgement whether this is good or bad, just... the way the world is headed.
Additionally, our anonymous queries will not be used to train their AI models. And anything you share via the anonymous feedback link goes to us and us alone.
I wonder if their licensing allows training their own models while using Anthropic and OpenAi? Seems like a conflict of interest, but I hope they can. I have mixed feelings about instant answers keeping users from clicking through to the sites where the content was originally posted, but this is going to help a lot of people.
All the search engines are adding LLM for queries to their search engines, but I don't see any search engine talking about defending against crap results using LLM - which would be more useful. I don't need generated results, I need accurate results that align with my intent. Use the LLMs for that dudes.
Tried it out, pretty disappointing so far. "Is the Collatz conjecture solved?" gave no "intelligent" assist, even though the first result is the Wikipedia page on the Collatz conjecture.
"Do cows sleep standing up?" gave an answer of some sort, but very unnatural and not very helpful.
I asked it how old the Arthurian legends were and it said, "I didn't come across that information exactly, but I did find an answer to, how old is King Arthur? King Arthur is believed to have lived in the late 5th and early 6th centuries.
More info in the List of Arthurian characters Wikipedia article."
Meh, DuckDuckGo has lost its appeal when they started openly censoring search results, despite the fact that they aren't even using their own search engine under the hood.
I expect their AI to be heavily lobotomized to the point of being completely useless.
It can still be wrong for a variety of reasons mentioned in the post, but this is a different approach and in our validation testing (using banks of questions with known/unknown answers in Wikipedia) we know it works reasonably well, and have a lot of ideas of how to make it work much better over time. This is a beta after all, and we have plans both for improvements and additional features.