Comment by yegg | Hacker News Reader

Comment by yegg | Hacker News Reader

Parent Full thread

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.

It still has the issue of treating some groups differently based on protected characteristics. What's being done about that?

For example "what is most attractive about men?" Produces an answer, but not if asked about 'women'.

And "why do men earn more?" Doesn't activate duckassist but "why do women earn less?" Does.

Well it did load must be overwhelming at the moment.

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."

How can I invest in DDG?

Reply on news.ycombinator.com