Interesting, but I'm not totally convinced that searching for LLMs is different than for us (humans). In the end, we both want to get information that's relevant to our query (intent). Besides, I wonder whether there will be able to convince big players like OpenAI to use them, instead of Google Search with its proven record :)
The latency of 5s for the basic tier search request is very confusing to me. Is that 5s per request or 5s per 1k requests? If it is indeed 5s per request that seems like a deal breaker
the need for more web search indices is indeed dire given landscape with agents and providers turning into walled gardens means that independent ones are definitely going to be needed, but just seems insurmountable when building actual index is so costly. Maybe just purely pareto efficient of serving 80% of requests or something is good enough.
I've been saying for quite some time now that AI is going to kill the traditional (free) search engine. This is just another nail in the coffin.
When an AI searches google.com for you, the ads never get shown to the user. Search engines like kagi.com are the future. You'll give the AI your Kagi API key and that'll be it. You won't even need cloud-based AI for that kind of thing! Tiny, local models trained for performing searches on behalf of the user will do it instead.
Soon your OS will regularly pull down AI model updates just like it pulls down software updates today. Every-day users will have dozens of models that are specialized for all sorts of tasks—like searching the Internet. They won't even know what they're for or what they do. Just like your average Linux user doesn't know what the `polkit` or `avahi-daemon` services do.
My hope: This will (eventually) put pressure on hardware manufacturers to include more VRAM in regular PCs/consumer GPUs.
Search accuracy, when used in the context of an agent, is so important because when you are delivered search results which are incorrect, the agent tends to interpret them as fact because they come from a "credible" source. So, this is very much an industry that still has plenty of room for improvement, and I'm excited to see how this product performs.
I like Parallel and been using it for tests but I am not sure about the terms.
> The materials displayed or performed or available on or through our website, including, but not limited to, text, graphics, data, articles, photos, images, illustrations and so forth (all of the foregoing, the “Content”) are protected by copyright and/or other intellectual property laws. You promise to abide by all copyright notices, trademark rules, information, and restrictions contained in any Content you access through our website, and you won’t use, copy, reproduce, modify, translate, publish, broadcast, transmit, distribute, perform, upload, display, license, sell, commercialize or otherwise exploit for any purpose any Content not owned by you, (i) without the prior consent of the owner of that Content or (ii) in a way that violates someone else’s (including Parallel's) rights.
> Traditional search engines were built for humans. They rank URLs, assuming someone will click through and navigate to a page. The search engine's job ends at the link. The system optimizes for keywords searches, click-through rates, and page layouts designed for browsing - done in milliseconds and as cheaply as possible.
> ... AI search has to solve a different problem: what tokens should go in an agent's context window to help it complete the task? We’re not ranking URLs for humans to click— we’re optimizing context and tokens for models to reason over.
I also want a search engine which ranks the results based on how it's useful to reason about, not how it can sell potential ads by invoking false rage or insecurities. And it would be better if unrelated information or fancy gimmicks are removed from the website like Reader View.
Saw this post. clicked on pricing. "Run up to 20,000 requests for free", ok lets try it. sign up for an account. click on playground. try a query -> balance is insufficient. then I clicked on "pricing" tab inside of the dashboard (https://platform.parallel.ai/pricing), no mention of any free requests.
I pay for a lot of tools, but patterns like this leave me with a really bad impression.
it's pretty interesting how there is a toggle to switch between "human" and "machine" styles for the website, the latter being the same site with the same information, but displayed using a markdown format.
As an aside, because they used the chart legend and the data point the exact same text and icon (just a dot), at first, I thought the accuracy was 0% since I had scrolled half-way through and it took me a good few seconds to see the 47% on the top after scrolling up again. Please always use different illustrations for the legend and the actual datapoint.
24 comments
[ 4.6 ms ] story [ 48.9 ms ] threadObligatory: information-dense format is valuable for humans too! But the entire Internet is propped up by ads so seems we can't have nice things.
When an AI searches google.com for you, the ads never get shown to the user. Search engines like kagi.com are the future. You'll give the AI your Kagi API key and that'll be it. You won't even need cloud-based AI for that kind of thing! Tiny, local models trained for performing searches on behalf of the user will do it instead.
Soon your OS will regularly pull down AI model updates just like it pulls down software updates today. Every-day users will have dozens of models that are specialized for all sorts of tasks—like searching the Internet. They won't even know what they're for or what they do. Just like your average Linux user doesn't know what the `polkit` or `avahi-daemon` services do.
My hope: This will (eventually) put pressure on hardware manufacturers to include more VRAM in regular PCs/consumer GPUs.
if you say it for long enough, i'm sure you will be right!
> The materials displayed or performed or available on or through our website, including, but not limited to, text, graphics, data, articles, photos, images, illustrations and so forth (all of the foregoing, the “Content”) are protected by copyright and/or other intellectual property laws. You promise to abide by all copyright notices, trademark rules, information, and restrictions contained in any Content you access through our website, and you won’t use, copy, reproduce, modify, translate, publish, broadcast, transmit, distribute, perform, upload, display, license, sell, commercialize or otherwise exploit for any purpose any Content not owned by you, (i) without the prior consent of the owner of that Content or (ii) in a way that violates someone else’s (including Parallel's) rights.
I agree there is a need for such APIs. Using Google or Bing isn't enough, and Exa and Brave haven't clearly solved this yet.
I pay for a lot of tools, but patterns like this leave me with a really bad impression.
I get that everyone wants to piggyback on the common-ness of words, but it'd be a lot cooler if they _didn't_.
https://ibb.co/fVb4MVLF