Browser, Collaborator, Synthesizer: The Evolution of a Thesis

As an educator and life-long learner, I’ve thought a lot about how to capture what I am learning and craft it into something useful. I’ve dug deep on PKM, note taking and ways to allow the ideas I come across to interact with each other.

I’ve been reading and following along with the work of Linus over at thesephist.com for a while now and this post is an attempt to follow his train of thought shared in posts from 2021 to 2025.

Stage 1: Personal knowledge tools, pre-LLM (2021)

I first came across Linus in 2021 when he was building two projects - Monocle and Revery. At that time, I didn’t have a full grasp of what semantic search was or how to implement it but this post made me very curious. Linus argued that:

“the future of the web browser is the best tool – nay, medium – for thought” — The web browser as a tool of thought, Aug 2021

At the time, I’d started using a (then) new tool called Obsidian and I was building a Digital Garden. This is something I’ve bounced up against continually since then but have struggled to make it consistent. A tool that could surface articles and quotes from my PKM while browsing the web seemed very exciting and still does now. Even as I write this I feel like my browser and RSS feedreader would both be enhanced by helping me link my thinking - enrich or supplement this new information with things I’ve written and encountered before.

My PKM exists separate from my browsing experience. Bringing those together in a tangible way where they could interact live with the web is something really compelling. I do have the technical and practical knowledge to generate embeddings and build semantic search over my PKM.

If the browser is to be a tool for thought then this needs to be multi-directional and easy. Encounter an article, have notes that are semantically similar surface in a logical way and then have the ability to add or enhance notes as I go.

What is interesting to me, is that Linus was thinking and exploring this space before the explosion of LLM which is the impetus, really, of following his chain of thought through to the present.

Stage 2: Tool vs. collaborator (2021)

In 2026, it doesn’t feel like a novel idea to consider AI as a collaborator but I really like the framing that Linus brings here.

As a tool, there is a specific action we’re asking AI to carry out - tidy up this picture, highlight stiff turns of phrase, format the document in helpful ways.

As a collaborator, there is a sense of working alongside.

“the human and the AI are two independent, autonomous agents” — AI as a creative collaborator, Nov 2021

We know that LLMs don’t produce output without input but having a series of skills that would kick off agentic loops as you write sounds really interesting. Linus has ideas about responding to color you’re using, questions that you wrote, noticing patterns in your code - we’re asking the LLM to look for patterns, provide suggestions and links to previous writing and thinking.

This builds on top of the PKM semantic search from before. Maybe a tool call to a semantic search that takes the writing that is going on and subtly provides launching off points. I don’t know what that would look like graphically but I feel like it needs to be in the background and not get in the way of actual creation. A side panel that can be closed or ignored, tool tips that can be switched off when in focus mode - this isn’t replacing my thinking or creation, instead it’s using my previous thoughts to stimulate new ones. I don’t like “type-ahead” as a paradigm for writing. I find it distracting and can often bring me down a path I didn’t want to go. Subtle suggestions and optional views would fit me much better.

Stage 3: Chat is the wrong default (2022)

Almost every LLM product is built around a chat interface. This makes sense as that’s how the API calls to these models are shaped but Linus makes the point that:

“Conversations are a terrible way to keep track of information” — Imagining better interfaces to language models, May 2022

It’s the easiest one to implement and often the only one that does. The challenge is that while there are knobs that we can twist (temperature, top-k, etc) the vector representations are often black boxes. Where would we “drag” this too in multi-dimensional space to make it more formal or funnier? It would be great to visualise and move things tangibly but this doesn’t seem to be something that is currently possible. Maybe it’s like democracy right now? The least worse forms of interaction?

Stage 4: Interpretability makes it concrete (2024)

Skip forward a few years and Linus is exploring how to leverage embeddings more directly as a tool for thought.

“with breakthroughs in mathematical understanding of a medium, come new tools” — Synthesizer for thought, Jun 2024

When we understood sound as a waveform, tools could be build that altered the wave and allowed for new and interesting sounds. Since we can generate embeddings for units of thought, could we alter the embedding and generate new units of thought? This feels similar to what LLMs do but potentially with some more control. At the moment though the black box of embeddings makes this hard to build. I imagine some AI labs are looking at this but I don’t know much about it but the thought of decomposing the features of a piece of writing and steering the output seems pretty interesting.

I really like Linus’ framing of what a document is:

“documents aren’t a collection of words, but a collection of concepts” — A notebook for seeing with language models, Jul 2024

Creating and building more human interfaces - beyond chat - to help build and experiment with content, expose more semantically meaningful “knobs” and push thought further seems really interesting.

Are DJs who create music with synteziers musicians? I think so.

Are folks who generate embeddings and play with them authors? I don’t know.

Stage 5: From wonder to pragmatic systems engineering (2025)

So much of our cultural narrative around AI and tech are placed in a context of productivity. But, some of the more interesting, engaging and life-giving things we can engage with grow out of wonder.

“Wonder is what guides my search for meaningful problems to work on” — Wonder engines, Apr 2025

I teach a lot in this space, use these tools for thought and exploration. I always want to write and think for myself and I want to focus on meaningful problems that have an impact on my life and the lives of others. Not to make us more efficient cogs in a machine but to help us be more fully human. Maybe that’s just touching more grass but maybe it’s by helping us think more deeply, imagine more creatively and create more diversely.

I’m excited to think and explore these concepts more. Maybe make my own versions of Monocle, explore Linus’ Prism or just continuing to create ways to help me build learning experiences more “sticky”? I don’t know what it looks like but reflecting on Linus’ work helps give shape and name to these thoughts and have a been a useful lens to reflect.

Linus has a newsletter and continues to publish articles. If you’ve connected with these ideas then make sure you subscribe and check them out.