Show HN: Web Search Powered by GPT and Bing | Hacker News Reader

Show HN: Web Search Powered by GPT and Bing

perplexity.ai

37 pointsaravindsrinivas11 comments

We built a search engine interface on top of OpenAI GPT 3.5 and Microsoft Bing that summarizes and cites top search results in response to natural language questions.

By using search results, the AI is able to reference recent news and provide citations for specific facts. Our interface offers concise answers, without having to click through links, scroll past irrelevant content, or read ads. No login is required; no personal data is collected.

We believe in the power of combining the intuitive UI of web search with the intelligence of large language models. The search engine does indexing work and interpretation is made by the LLM (a la OpenAI WebGPT https://openai.com/blog/webgpt/). We are also working on using LLMs to reword queries before searching and to make searches more conversational.

Additionally, we are building new features that allow you to curate the result of your query, e.g., by omitting references that are not relevant and expanding the list of references that are relevant. New references could be added either “breadth-first” (include more results from the original query) or “depth-first” (summarize more content from one of the pages in the search results and find related links).

Curious what the community thinks, and what features come to mind when you use this interface.

Twitter: https://twitter.com/perplexity_ai/status/1600551871554338816 Discord: https://t.co/R4G21AmwQ7

Open articleView on HN

The idea of using the first search results as an input for GPT is a good compromise between the static/"smart" (static in time) GPT and the dynamic/"dumber" search engines.

I also agree it is a good idea to give control to the user to choose

- How to expand : between “breadth-first” and “depth-first”

- Then, how to filter : using dynamic keywords/references

This is what I’ve been telling people all day long these days. It is not an either/or. Good UX design (I like the citations system in your result!) can take the best of both words: The classic search engine (or any other classic UI), and AI functions/conversational UI.

The business model is unclear, but you are definitely into something here.

Good luck !

Good ideas.

Business model wise, beyond this interface we have a demo of natural language to data-code (SQL and Python pandas/matplotlib/etc. APIs).

> Good ideas.

Well, I just reformulated some key ideas in your post. Nothing added really.

> natural language to data-code (... Python pandas/matplotlib/etc)

You mean something like this? https://news.ycombinator.com/item?id=33876826

hal9 is cool, and in a similar spirit. One thing we've observed with many approaches being explored lately is that they generate code into an IDE like experience. We are more of the camp of making the code gen and code run happen under the hood, so that you get an end-to-end experience ("ask question" -> "see result of running generated code").

Then still allow the user to inspect the code generated, fix it if necessary, and re-run. It's a bit of a subtle distinction but we think that UI/UX makes a big difference in the utility of these LLM applications.

Yeah, it is more in the spirit of "no code first" [0]

I suspect generic data-driven low code platforms like pipedream [1] and autocode [2] could do it this way at some point.

[0] https://zupa.hu/posts/no-code-first [1] https://pipedream.com/ [2] https://autocode.com/

Yeah, exactly "no code first", which I hadn't heard of before. Cool.

I tried Perplexity this morning with the following prompt:

"What is the evidence for and against systematic phonics instruction?"

It not only gave me the stuff 'everyone knows' but also pointed to a 2020 paper that goes against the conventional wisdom. (I had read that paper a few days ago, and was curious whether the AI would mention it.).

Pretty cool that I get a simple summary of information with citations. Would love it if it were a little bit more conversational considering that's a huge strength of GPT. Would love to see how this develops further.

+1 to having it be more conversational. Would be cool to have perplexity (optionally) ask me follow up questions to better narrow/home in on the most relevant answer I'm looking forward

The biggest obstacle is gpt cost per query. How are you planning to scale this?

Good question. We are betting that the cost per LLM inference will continue to go down for given level of performance as the market matures over the next few years.

Reply on news.ycombinator.com