I like to joke at work that soon, when my co-worker sends me an email, they’ll just write 5 bullet points and let an LLM in their email client turn it into a nice professional email. That will save them a lot of time. But guess what? I already get too many emails as it is, so when I receive theirs, I’ll also ask an LLM to summarize it and extract the 5 key bullet points.
In essence, their LLM (assistant) will be talking to my LLM (assistant).
It’s not as silly as it sounds though, because the 5 bullet points I’ll read may not be exactly the ones they typed. It’s possible that their LLM takes into account local context that it knows about them. And it works the same on the receiving end: my LLM may extract information in a way relevant to me that my co-worker may not be aware of. So there is value in these LLMs talking to each other, each on our behalf.
The 1-Page Book
This silly observation is part of a larger theme I have been pondering for a while. I have read too many business, tech and startup books that are too long and repetitive. Often it takes multiple chapters for the author to introduce a topic I’m already quite familiar with.
I would daydream of a time where all books fit in one page, perfectly tailored for my current level of expertise and understanding.
Has that time come? The author usually has a few specific goals in mind for their reader: convince them of some idea or concept, and take them from their current state to enlightenment. Except the author is in the dark as to exactly who their reader is, what they already agree and understand, and what arguments they will find difficult to accept. So the author makes the book longer, trying to cover all bases. And in practice, losing me, the impatient one.
It’s just summarization
Isn’t this the same as browser summarization?
Not quite. Browser summarization today is too generic: it doesn’t take into account the reader’s views. That might change, but even if it does, it has no insights into the author’s intent. Sure, if it had a whole book to summarize, it could theoretically do a great job. But in practice, given a shorter blog post or article, it doesn’t know enough about the author’s views.
It’s a privacy nightmare
Sharing reader’s information is a privacy nightmare.
I certainly don’t want websites to get more personal data about us to track us even more. So the customization needs to happen on device somehow. I’ll dig into that topic in another post.
Work in Progress
There’s existing literature on the topic: Multi-Agentic LLMs for Personalizing STEM Texts or Tailoring Scientific Knowledge: How Generative AI Personalizes Academic Reading Experiences.
A reddit user makes a good point about summarization (AI summarizes content but doesn’t preserve how ideas connect) which I think agrees with the core of my argument.
I’m very interested in brainstorming ideas around that space. I have some ideas about how that might work and what it might look like, but the more opinions, the better.
Reach out on @[email protected] with your thoughts, questions, objections or encouragement!