I am starting to not care at all with the AI search trends. I went through all the hurdles and frustrations to sign up for bing AI. When I finally got access and asked my first question it straight up lied to me. This tech needs another decade in the oven because it is nothing more than a cool toy that pretends to have conversations with you
Heh, I read "relevant" as meaning "containing the terms I'm searching for". You comment made me realize that the definition has shifted over the years. I agree with you -- I think that Google, for instance, trying to read my mind for what I "really want" is the main reason that Google search sucks for me. I expect no better when any other search engine tries that.
I think what we are witnessing is what Marxists call reification - the shifting of the meaning from an abstract to a concrete. That is Google has successfully reified the term relevant such that it is easier for them to control and market around.
Another example of a pretty atrocious case of reification is when facebook reified the word friend to mean people you follow on their platform.
A way to “surface reliable information quickly” is to “find relevant web pages quickly”…so is extracting and summarizing info from the internet that matches the query using AI. You quoted the end, and said you instead wanted the means
Google (and apparently DDG)'s position is that they're not building that. Google started building that, but only because it was the technology available to get closer to the desired goal.
ChatGPT has been shown time and time to mix up facts, censor information, and make shit up. It's not an "AI", it's a glorified autocomplete box. I struggle to see how it can replace a search engine.
Sure. Sometimes it’s wrong. But I spend less time finding an answer on ChatGPT than Google. For example I’m not a c++ developer but I’ve inherited some c++ code work to do. So I’m like wow this vector thing is only ever looking for this 1 value. It would be better to use a dictionary or hash map or something. So I ask ChatGPT and it recommended unordered_map, and I asked why the existing code snippet I gave it had a lock around it, and it explained the thread safety with read/write etc. So fixing that all up shaved off 700ms in processing time in my test scenario.
I can ask it what’s the difference between 2 lists. I asked it about concurrent_unordered_set and it told me that it only exists in VisualC++ but if I want it cross platform I can look at Boost or TBB.
None of this information is easy as a non C++ developer to find on Google.
> it’s drawing on natural language technology from ChatGPT-maker OpenAI and Anthropic
> sources it’s using for DuckAssist are — currently — “99%+ Wikipedia”
Not sure how I feel about either of these.
What's the difference between using Bing, chatGPT or DuckAssist if they're all underpinned by OpenAI? (currently I gravitate towards Bing) Obviously Bing is performing live searches right now but thinking more about the near future.
If it's mainly summarizing Wikipedia rather than aggregating more diverse web results does that undermine or harm Wikipedia somehow long-term?
And what about poor Google? I've only used it for a couple of times for measurement conversion since getting access to bing chat, and it was really only because I felt bad bothering bing with it.
Maybe it's a good thing for Wikipedia? DDG can scrape Wikipedia once, and then let their AI act as a sort of "knowledge CDN" so Wikipedia ultimately serves fewer page loads. As long as DDG supports Wikipedia in some way that seems like a net positive.
Seems to me that Bing and Google's chatbots are having truthiness issues because they're just scraping information from all over the internet. Any HN reader knows all too well that you can't believe everything you read on the internet. While Wikipedia isn't necessarily all true, the bar is much, much, much higher than the truthiness of the rest of the internet. Can anyone else verify if a LLM trained on a set of mostly-true info like Wikipedia will still make shit up?
> Can anyone else verify if a LLM trained on a set of mostly-true info like Wikipedia will still make shit up?
It will, but less frequently. For instance, I can imagine a situation where it would give a plausibly sounding answer to a question based on an incorrect premise. The issue is that the bullshit responses will sound plausible from the stylistic pov.
Under the hood LLMs we're talking about here are still just "spicy autocomplete" with non-deterministic results. The results are impressive, just not in the way the VC driven discourse/hype would suggest.
I really hope that DDG doesn't roll this into their web search, though. As a separate tool, cool. But as part of web search, it would make DDG less valuable to me.
OTOH, maybe that's what would finally push me to using Kagi, which might not be the worst thing.
I've been loving https://www.perplexity.ai/, find myself using it more and more. Don't have Bing AI access yet but seems similar in that it cites sources, but more powerful with it's ChatGPT-likeness.
Brave Search recently added AI answers and I'm surprised at how much I like it. To get the gist of a complex topic, it's better than Google's snippets, and better than being redirected to lengthy articles.
