One thing that the various Deep Research tools are demonstrating is that there is tremendous value in a RAG strategy that just consists of giving an advanced model a search engine and letting it look up whatever it feels is necessary in order to respond to a prompt. - ThreadSky

emollick.bsky.social • 56 days ago

One thing that the various Deep Research tools are demonstrating is that there is tremendous value in a RAG strategy that just consists of giving an advanced model a search engine and letting it look up whatever it feels is necessary in order to respond to a prompt.

Comments

drposts.bsky.social•56 days ago

100%. It’s one of the reasons I am shocked Claude has not enabled native web search yet

Forcing people to use Brave Search via MCP means only a select few highly technical users will get the benefit

recursiverealms.com•55 days ago

You mean, like a human? That's crazy, you'd just GIVE access to things to artificial intelligences? but they could be evil! and they could.. search evil things! and ummm... search MORE evil things! *scary vampire look* 😅🤣

Yah totally.

un1crom.bsky.social•56 days ago

the history of knowledge is the history of index and retrieval. (image i think i original got from fielden at sf state university. can’t find live link anymore)

tomsharpsays.bsky.social•56 days ago

I enjoy seeing how the labs are figuring out this tech can be made the most useful — in public.

A bit similar to the continuous beta status of Gmail at that time

mergesort.me•56 days ago

The big downside to this approach is that Deep Research will use sources like https://statista.com, as it does in your screenshot, as if they’re not an SEO farm with tons of incorrect data. Personally I’m much more excited about running Deep Research over your own data, where you know the data’s integrity.

joshuashew.bsky.social•56 days ago

Well, that might be true now, but I’m hopeful that they’re gonna be able to figure this out. Like maybe run a more efficient model on search results to be able to identify the ones that are more likely to be credible. Or just provide it a better search engine—I hear there are alternatives.

mergesort.me•56 days ago

I follow along with AI news very closely and read papers about the subject, but haven’t heard any discussion about that. Any chance you can link me to some sources so I can better understand what approaches they’re using or considering?

joshuashew.bsky.social•56 days ago

Oh and to make it clear! I don’t think they are doing this yet. Just trying to express that I’m hopeful they can figure out the low quality input issue and improve its output.

mergesort.me•56 days ago

It’s definitely possible and I think pretty likely that we’re in the v1 of this feature, so I appreciate the thoughts, I just haven’t seen anyone publicly discussing or working on this and would like to read something authoritative before jumping to a conclusion about how it would work. 😄

joshuashew.bsky.social•56 days ago

Yeah for sure! I’m looking forward to the improvements and it’s fun to think of ways it might happen.

joshuashew.bsky.social•56 days ago

Then you follow this much more closely than me! Let me share my thinking anyway and lmk what you think.

(dictated; please excuse errors)

joshuashew.bsky.social•56 days ago

I’m borrowing from ideas on AI alignment that talk about running a smaller (less intelligent) model on the outputs of a more intelligent model to determine if they are scheming or not. The idea is that detection is easier than generating the scheming itself.

joshuashew.bsky.social•56 days ago

Similarly, I wonder if determining the credibility of a source (like an SEO optimized source) is easier than taking a wide range of credible sources and incorporating those into a meaningful report.

isolyth.dev•55 days ago

Would be interesting to see how a more specialized model designed to effectively search for information being called by a more general but not-as-good at searching model would perform

arivero.bsky.social•56 days ago

Honestly this looks more agentic that ragged.

Comments

Posting Rules

Reply