This is a super helpful breakdown and really helps me understand how the RL step is different than the initial training step. I didn't realize the reward was delayed until the end of the response for the RL step. Having the reward for this step be dependent on the coherent thought rather than a coherent word now seems like an obvious and critical part of how this works.
I had a very similar idea a while back. I wanted to rank news by "impact" which might be more concrete than "significance."
For an LLM prompt, it would be something like:
"estimate the number of people who's lives that will be materially changed by this news."
and
"estimate the average degree of change for those impacted."
Then impact is roughly the product of those two.
Additionally, I want a version that is tailored to me specifically "estimate the degree of change this will have on my life." + context of my life.
Tangentially, I've found that getting ratings out LLMs works better when I can give all options and request relative ratings. If I ask for rankings individually I get different and less good results. Not enough context length to rate all news from all time in one go though. Any thoughts on that? Maybe providing some benchmark ratings with each request could help? Something I'm exploring.
> 3. Hypothetical answer generation from a query using an LLM, and then using that hypothetical answer to query for embeddings works really well.
I've been wondering about that and am glad to hear it's working in the wild.
I'm now wondering if using a fine-tuned LLM (on the corpus) to gen the hypothetical answers and then use those for the rag flow would work even better.
This looks cool. Here's my feature request list (can't tell if these are there yet):
- import existing html/css. This would allow me to use it for non-greenfield projects, and allow for back-and-forth workflows.
- mark nodes with my own semantic ids that are included in the html export. I would use these to post process exported html to create dynamic templates/components/views. I don't want to be tied to a specific framework.
Agreed. I really like the concept of the now page as a succinct quarterly update. I want it to only change once per quarter.
I also really like consolidated news feeds of one persons activity across all platforms (which seems more like what these "auto now" page attempts are going for).
I think we need a new term. I would like to call it /feed. It could even be implemented as an rss/atom/activitypub aggregate feed of all a user's activity across publishing platforms.
/feed
Let's make it a thing.
And let's keep /now as the quarterly update.
My broad sweeping generalization was primarily meant as a counterpoint to this from parent comment: "The pandemic showed us exactly what children would prefer to do, when they don't have a physical teacher standing over them, which is bugger all."
My point is more that kids, when left to their own devices (with basic needs met), will find ways to occupy themselves that they find interesting that are not outcome oriented (I call this playing).
And I personally have never met a kid that didn't like playing in some form or another, though the form of playing is highly, highly individualized.
I completely agree. As language is the preferred encoding method for intelligent thought (at least in our species) it could very well be that a sufficiently accurate language model is also a generally intelligent model.
I'd argue that what you're talking about in fiction is coherence (internal consistency) not factual accuracy (consistency with an externally verifiably ground truth).
I'd also argue that the economic value of coherent bullshit is ... quite high. Many people have made careers out of producing coherent bullshit (some even with incoherent bullshit :-).
Of course, in the long run, factual accuracy has more economic value than bullshit.
LLMs would be better nomenclature than AI in this context.
LLMs are not factual databases. They are not trained to retrieve or produce factual statements.
LLMs give you the most likely word after some prior words. They are incredibly accurate at estimating the probabilities of the next word.
It is a weird accident that you can use auto-regressive next word prediction to make a chat bot. It's even weirder that you can ask the chatbot questions and give it requests and it appears to produce coherent answers and responses.
LLMs are best thought of as language generators (or "writers") not as repositories of knowledge and facts.
LLM chatbots were a happy and fascinating (and for some, very helpful) accident. But they were not designed to be "factually correct" they were designed to predict words.
People don't care about (or are willing to accept) the "wrong answers" because there are enough use cases for "writing" that don't require factual accuracy. (see for instance, the entire genre of fiction writing)
I would argue that it is precisely LLMs ability to escape the strict accuracy requirements of the rest of CS and just write/hallucinate some fiction that is actually what makes this tech fascinating and uniquely novel.
Ditto. I somewhat recently learned that Safari only keeps the history log for one year unless you explicitly set "Remove history items" to "manually." Which I've done so that I at least have a list that I could crawl in the future to build a full text index on.