스크립트
00:00:00-
00:00:12- Okay, today I'm gonna talk about
00:00:14where reinforcement learning will take search.
00:00:18And some background on me.
00:00:22I'm the founder and CEO of Cit.AI.
00:00:24We're a stealth-ish AI lab for search.
00:00:27We're backed by some pretty amazing people,
00:00:30and we're hiring.
00:00:35Okay, agents are a new paradigm for search.
00:00:40You can now get vastly higher quality results.
00:00:45Twice as likely to find the right documents.
00:00:49But it's incredibly expensive.
00:00:52About 100 to 1,000 times more expensive
00:00:54than what you'd get out of a classical search query.
00:00:58And it is extremely slow.
00:00:59You're looking at minutes and not milliseconds,
00:01:02like Turbo Puffer.
00:01:05And what this means is that agents spend 30 to 50%
00:01:10of their tokens on searching.
00:01:11And this is usually at the beginning of some task.
00:01:14It finds the right context to then do whatever
00:01:17what you ask it to do.
00:01:19And the idea here is quite simple.
00:01:21First, instead of having the main agent do the searching,
00:01:25you pass the searching to a sub-agent,
00:01:28and you train a model to be a great sub-agent.
00:01:32And the question here that we'll answer today
00:01:35is how much cheaper and faster can we make this
00:01:38with reinforcement learning?
00:01:40And this is really our target.
00:01:42So this is a benchmark across legal, finance,
00:01:45knowledge bases, science, email,
00:01:48a bunch of different tasks, some academic benchmarks,
00:01:50some internal benchmarks.
00:01:52And this is where you currently are.
00:01:58You see Reranker and Vector-only performance
00:02:00at the bottom.
00:02:01And you can see frontier models essentially
00:02:03kind of like go through here at the cost of spending
00:02:07many, many minutes per question.
00:02:10And can we get a model to kind of like be extremely accurate,
00:02:14have extremely high recall,
00:02:16but also be incredibly fast and cheap?
00:02:19And let's quickly look at classical search.
00:02:23This is the pipeline that many of you guys
00:02:25will be familiar with.
00:02:26A question comes in, you might have an LLM
00:02:29that rewrites the question.
00:02:30You then execute that on a search backend.
00:02:33You might have a Reranker
00:02:35and you get your results at the end of the day.
00:02:38It is essentially a pipeline of chained,
00:02:40locally optimized models.
00:02:43And all of the decisions are baked in at design time.
00:02:47And you expand a fixed amount of compute per question.
00:02:51And this one is really important.
00:02:53The Reranker might know that the results
00:02:55are insufficient at answering the question,
00:02:58but the Reranker can't take action.
00:03:00It can only essentially return the results
00:03:02even when they're bad.
00:03:03And what this means is that in practice,
00:03:06a pipeline like this accrues a long tail of failure
00:03:09where unexpected questions come
00:03:11that the designer didn't have something for.
00:03:15And in practice, this usually means people add
00:03:17lots of edge cases to essentially fix these.
00:03:21But of course, you can't design infinite edge cases,
00:03:23and you can't add infinite tweaks.
00:03:27And so the strategy here is one that we've seen before.
00:03:32Machine design outperforms human design.
00:03:36And we saw this in computer vision
00:03:38where you had your primitive edge detection algorithms.
00:03:42You then had the box-around-a-dog generation of models
00:03:45that were very good at this very narrow task
00:03:49and locally optimized for it.
00:03:50And then you had VLMs that were extremely good
00:03:53at all parts of the search pipeline.
00:03:54You saw this again with chess, with IBM Deep Blue
00:03:58being largely a collection of human-written rules,
00:04:01Stockfish bridging the two, and then Alpha Zero and Mu Zero
00:04:04essentially completely just putting it all inside of the model.
00:04:07And we're gonna see something similar happen to search
00:04:11where we have our existing algorithms
00:04:13like BM25 and PageRank.
00:04:16Then we had an evolution from that with small models
00:04:19that did some task very well, like vectors and re-rankers.
00:04:23And now essentially this new paradigm of pure RL
00:04:28where we actually don't make any design decisions
00:04:30into the model.
00:04:32And what this looks like in practice is something like this.
00:04:37So you have one model, it goes back and forth
00:04:41with the database, it can search, it can read results,
00:04:43it can iterate, it can search again until it is happy.
00:04:47It can set metadata filters on the fly,
00:04:50it can constrain its search, it can try as much as it wants.
00:04:55And in the end it produces a ranked list of results.
00:05:00And what you get is a model that makes all of the decisions
00:05:03and can adapt to any question on the fly,
00:05:06and for example, use much more compute
00:05:07if a user asks a very difficult question.
00:05:11And what helps us here is that search is verifiable.
00:05:15And reinforcement learning needs rewards
00:05:17that are verifiable and grindable.
00:05:19Verifiable here means for a given question,
00:05:22did you find the correct document?
00:05:23And we can design this and tell this quite easily.
00:05:28And is there an environment where the model can attempt this question
00:05:32loads and loads of times?
00:05:33And in practice for us this means thousands of times per second
00:05:37during a training run.
00:05:40And the second part is kind of like,
00:05:42can we turn the models that are currently very general
00:05:45and very general purpose into something that is much more specialized?
00:05:49And it turns out we don't actually need most of the parts
00:05:51of a language model to be extremely performant at search.
00:05:55And similarly, with like CPUs and GPUs and ASICs,
00:05:58a CPU in theory can do anything that a GPU can do,
00:06:02but you would never want to use a CPU to do LLM inference, for example,
00:06:06because the much more specialized version is much more effective.
00:06:10And this really makes search an ideal target for RL.
00:06:15And this is what happens when you train a model on this task.
00:06:19And so again, here we added the vector and Reranker-only baselines.
00:06:22This is of an earlier task.
00:06:24And what we see is that search quality increases
00:06:26very predictably with compute.
00:06:28And we can mix in other rewards like latency
00:06:30and different kind of like retrieval strategies
00:06:33to make it even more performant.
00:06:35Importantly, we don't really tell the model what to do.
00:06:38Much like in AlphaZero and chess,
00:06:40we want it to discover its own strategies
00:06:42and its own tricks to essentially use search well.
00:06:45And we don't know whether this method has no ceiling,
00:06:50but we're definitely not yet seeing a ceiling to this approach.
00:06:54And these are the results.
00:06:56So this is the same chart as before,
00:06:58and this is SID1 and then SID1
00:07:00with some parallel execution on the left-hand side.
00:07:04And so what this ends up meaning is
00:07:06you're about 20 times faster.
00:07:09So instead of taking around two minutes,
00:07:10you take around five seconds on average,
00:07:12and it's about 100 times cheaper
00:07:14than using a frontier model for this task.
00:07:18There's some more detail here,
00:07:19but like, yeah, the cost and kind of like speed
00:07:22are just completely incomparable.
00:07:25You can, yeah.
00:07:27It's not quite at the latency of a vector and re-ranker pipeline,
00:07:32but in practice, we think we can get there quite quickly.
00:07:36And how does this look like in production?
00:07:38So this is a usual kind of like agent execution trace.
00:07:41The agent does some searching here.
00:07:42It finds some good stuff.
00:07:43It finds some bad stuff.
00:07:45But all of the bad stuff that it finds
00:07:47is essentially polluting its own context window.
00:07:50And what we can instead do
00:07:51is use a sub-agent here
00:07:54that does all of the searching and thinking
00:07:56and iterating for the main agent,
00:07:57and the main agent only ever sees great results.
00:08:00And this means that the main agent sees more good stuff,
00:08:03which means it's more likely to be correct,
00:08:06and it is also extremely cost-effective,
00:08:10where those 30 to 50% tokens
00:08:14that were earlier used by the main agent
00:08:16to do searching can now be passed off
00:08:17to this just, you know, 100x cheaper search sub-agent.
00:08:25And where will this take us?
00:08:29Scaling RL will give us arbitrarily good search in any domain.
00:08:35And RL models will become even faster,
00:08:38which will allow them to be used
00:08:39in things like voice and e-commerce.
00:08:42They'll become even cheaper than we are currently.
00:08:45So the charts that you saw there,
00:08:47I think we can move even further.
00:08:50And better search will unlock more knowledge work tasks.
00:08:55The web is actually quite small
00:08:56in comparison to the entirety of data that is out there.
00:09:03And the most valuable information is not on the internet.
00:09:06For example, how to run JP Morgan is nowhere on the web,
00:09:09but it is deep inside of the databases at JP Morgan.
00:09:14So that's it for me.
00:09:16Thank you.
00:09:16Thank you.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기