I don't understand the point of automating note taking. It never worked for me to copy paste text into my notes and now you can 100x that?
The whole point of taking notes for me is to read a source critically, fit it in my mental model, and then document that. Then sometimes I look it up for the details. But for me the shaping of the mental model is what counts
First of all, this is more than just note taking. It appears to be a (yet another) harness for coordinating work between agents with minimal human intervention. And as such, shouldn’t part of the point be to not have to build that mental model yourself, but rather offload it to the shared LLM “brain”?
Highly debatable whether it’s possible to create anything truly valuable (valuable for the owner of the product that is) with this approach, though. I’m not convinced that it will ever be possible to create valuable products from just a prompt and an agent harness. At that point, the product itself can be (re)created by anyone, product development has been commodified, and the only thing of value is tokens.
My hypothesis is that “do things that don’t scale”[0] will still apply well into the future, but the “things that don’t scale” will change.
All that said, I’ve finally started using Obsidian after setting up some skills for note taking, researching, linking, splitting, and restructuring the knowledge base. I’ve never been able to spend time on keeping it structured, but I now have a digital secretary that can do all of the work I’m too lazy to do. I can just jot down random thoughts and ideas, and the agent helps me structure it, ask follow-up questions, relate it to other ongoing work, and so on. I’m still putting in the work of reading sources and building a mental model, but I’m also getting high-quality notes almost for free.
The few scientific studies out there actually show a degradation of output quality when these markdown collections are fully LLM maintained (opposed to an increase when they’re human maintained), which I found fascinating.
I think the sweet spot is human curation of these documents, but unsupervised management is never the answer, especially if you don’t consciously think about debt / drift in these.
Are you referring to the one (1) study that showed that when cheaper LLM's auto-generated an AGENTS.md, it performed more poorly than human editted AGENTS.md? https://arxiv.org/abs/2602.11988
I'd love to see other sources that seek to academically understand how LLM's use context, specifically ones using modern frontier models.
My takeaway from these CLAUDE.md/AGENTS.md efforts isn't that agents can't maintain any form of context at all, rather, that bloated CLAUDE.md files filled with data that agents can gather on the spot very quickly are counter-productive.
For information which cannot be gathered on the spot quickly, clearly (to me) context helps improve quality, and in my experience, having AI summarize some key information in a thread and write to a file, and organize that, has been helpful and useful.
The space of self building artefacts is interesting and is booming now because recent LLM versions are becoming good at it fast (in particular if they are of the "coding" kind).
I've also experimented recently with such a project [0] with minimal dependencies and with some emphasis on staying local and in control of the agent.
It's building and organising its own sqlite database to fulfil a long running task given in a prompt while having access to a local wikipedia copy for source data.
A very minimal set of harness and tools to experiment with agent drift.
Adding image processing tool in this framework is also easy (by encoding them as base64 (details can be vibecoded by local LLMs) and passing them to llama.cpp ).
It's a useful versatile tool to have.
For example, I used to have some scripts which processed invoices and receipts in some folders, extracting amount date and vendor from them using amazon textract, then I have a ui to manually check the numbers and put the result in some csv for the accountant every year. Now I can replace the amazon textract requests by a llama.cpp model call with the appropriate prompt while still my existing invoices tools, but now with a prompt I can do a lot more creative accounting.
I have also experimented with some vibecoded variation of this code to drive a physical robot from a sequence of camera images and while it does move and reach the target in the simple cases (even though the LLM I use was never explicitly train to drive a robot), it is too slow (10s to choose the next action) for practical use. (The current no deep-learning controller I use for this robot does the vision processing loop at 20hz).
Put AI in your product name, make billion dollars. Put Karpathy in your blog article, get hired by Anthropic as Principal engineer. Milk money as long as fad last. No one is thinking about customer needs, everyone is trying to wash hands in the wave as it last.
LLM models and the agents that use them are probabilistic, not deterministic. They accomplish something a percentage of the time, never every time.
That means the longer an agent runs on a task, the more likely it will fail the task. Running agents like this will always fail and burn a ton of token cash in the process.
One thing that LLM agents are good at is writing their own instructions. The trick is to limit the time and thinking steps in a thinking model then evaluate, update, and run again. A good metaphor is that agents trip. Don't let them run long enough to trip. It is better to let them run twice for 5 minutes than once for 10 minutes.
Give it a few weeks and self-referencing agents are going to be at the top of everybody's twitter feed.
The BM25-first routing bet is interesting. You mention 85% recall@20 on 500 artifacts, but the heuristic classifier routing "short lookups to BM25 and narrative queries to cited-answer" raises a practical question: what does the classifier key on to decide a query is narrative vs short? Token count? Syntactic structure? The reason I ask is that in agent-generated queries, the boundary is often blurry - an agent doing a dependency lookup might issue a surprisingly long, well-formed sentence. If the classifier routes those to the more expensive cited-answer loop it could negate the latency advantage of BM25 being first.
Any particular reason for BM25? Why not just a table of contents or index structure (json, md, whatever) that is updated automatically and fed in context at query time? I know bag of words is great for speed but even at 1000s of documents, the index can be quite cheap and will maximise precision
I read the durability thing as markdown files are very open, easy to find software for, simple and are widely used. All of this together almost guarantees that they will he viewable/usable in the far future.
I don't understand the point of automating note taking. It never worked for me to copy paste text into my notes and now you can 100x that?
The whole point of taking notes for me is to read a source critically, fit it in my mental model, and then document that. Then sometimes I look it up for the details. But for me the shaping of the mental model is what counts
First of all, this is more than just note taking. It appears to be a (yet another) harness for coordinating work between agents with minimal human intervention. And as such, shouldn’t part of the point be to not have to build that mental model yourself, but rather offload it to the shared LLM “brain”?
Highly debatable whether it’s possible to create anything truly valuable (valuable for the owner of the product that is) with this approach, though. I’m not convinced that it will ever be possible to create valuable products from just a prompt and an agent harness. At that point, the product itself can be (re)created by anyone, product development has been commodified, and the only thing of value is tokens.
My hypothesis is that “do things that don’t scale”[0] will still apply well into the future, but the “things that don’t scale” will change.
All that said, I’ve finally started using Obsidian after setting up some skills for note taking, researching, linking, splitting, and restructuring the knowledge base. I’ve never been able to spend time on keeping it structured, but I now have a digital secretary that can do all of the work I’m too lazy to do. I can just jot down random thoughts and ideas, and the agent helps me structure it, ask follow-up questions, relate it to other ongoing work, and so on. I’m still putting in the work of reading sources and building a mental model, but I’m also getting high-quality notes almost for free.
[0]: https://www.paulgraham.com/ds.html
The few scientific studies out there actually show a degradation of output quality when these markdown collections are fully LLM maintained (opposed to an increase when they’re human maintained), which I found fascinating.
I think the sweet spot is human curation of these documents, but unsupervised management is never the answer, especially if you don’t consciously think about debt / drift in these.
Are you referring to the one (1) study that showed that when cheaper LLM's auto-generated an AGENTS.md, it performed more poorly than human editted AGENTS.md? https://arxiv.org/abs/2602.11988
I'd love to see other sources that seek to academically understand how LLM's use context, specifically ones using modern frontier models.
My takeaway from these CLAUDE.md/AGENTS.md efforts isn't that agents can't maintain any form of context at all, rather, that bloated CLAUDE.md files filled with data that agents can gather on the spot very quickly are counter-productive.
For information which cannot be gathered on the spot quickly, clearly (to me) context helps improve quality, and in my experience, having AI summarize some key information in a thread and write to a file, and organize that, has been helpful and useful.
I think there‘s a serious issue with people using AI to do an immense amount of busywork and then never look at it again. Colossal waste.
Then you have never worked at a large enough codebase or across enough many projects?
The space of self building artefacts is interesting and is booming now because recent LLM versions are becoming good at it fast (in particular if they are of the "coding" kind).
I've also experimented recently with such a project [0] with minimal dependencies and with some emphasis on staying local and in control of the agent.
It's building and organising its own sqlite database to fulfil a long running task given in a prompt while having access to a local wikipedia copy for source data.
A very minimal set of harness and tools to experiment with agent drift.
Adding image processing tool in this framework is also easy (by encoding them as base64 (details can be vibecoded by local LLMs) and passing them to llama.cpp ).
It's a useful versatile tool to have.
For example, I used to have some scripts which processed invoices and receipts in some folders, extracting amount date and vendor from them using amazon textract, then I have a ui to manually check the numbers and put the result in some csv for the accountant every year. Now I can replace the amazon textract requests by a llama.cpp model call with the appropriate prompt while still my existing invoices tools, but now with a prompt I can do a lot more creative accounting.
I have also experimented with some vibecoded variation of this code to drive a physical robot from a sequence of camera images and while it does move and reach the target in the simple cases (even though the LLM I use was never explicitly train to drive a robot), it is too slow (10s to choose the next action) for practical use. (The current no deep-learning controller I use for this robot does the vision processing loop at 20hz).
[0]https://github.com/GistNoesis/Shoggoth.db/
Put AI in your product name, make billion dollars. Put Karpathy in your blog article, get hired by Anthropic as Principal engineer. Milk money as long as fad last. No one is thinking about customer needs, everyone is trying to wash hands in the wave as it last.
LLM models and the agents that use them are probabilistic, not deterministic. They accomplish something a percentage of the time, never every time.
That means the longer an agent runs on a task, the more likely it will fail the task. Running agents like this will always fail and burn a ton of token cash in the process.
One thing that LLM agents are good at is writing their own instructions. The trick is to limit the time and thinking steps in a thinking model then evaluate, update, and run again. A good metaphor is that agents trip. Don't let them run long enough to trip. It is better to let them run twice for 5 minutes than once for 10 minutes.
Give it a few weeks and self-referencing agents are going to be at the top of everybody's twitter feed.
Karpathy's original post for context:
https://x.com/karpathy/status/2039805659525644595
https://xcancel.com/karpathy/status/2039805659525644595
The BM25-first routing bet is interesting. You mention 85% recall@20 on 500 artifacts, but the heuristic classifier routing "short lookups to BM25 and narrative queries to cited-answer" raises a practical question: what does the classifier key on to decide a query is narrative vs short? Token count? Syntactic structure? The reason I ask is that in agent-generated queries, the boundary is often blurry - an agent doing a dependency lookup might issue a surprisingly long, well-formed sentence. If the classifier routes those to the more expensive cited-answer loop it could negate the latency advantage of BM25 being first.
Any particular reason for BM25? Why not just a table of contents or index structure (json, md, whatever) that is updated automatically and fed in context at query time? I know bag of words is great for speed but even at 1000s of documents, the index can be quite cheap and will maximise precision
Couldn't you instruct your LLM to make the starting dir configurable?
I love that so many people are building with markdown !
But also would like to understand how markdown helps in durability - if I understand correctly markdown has a edge over other formats for LLMs.
Also I too am building something similar on markdown which versions with git but for a completely different use case : https://voiden.md/
I read the durability thing as markdown files are very open, easy to find software for, simple and are widely used. All of this together almost guarantees that they will he viewable/usable in the far future.
So markdown will be great for distribution in the future.
Don’t know if Karpathy even wrote this version. Where are the citations?
need to try out asap. love the „the office“ vibe
love the bm25-first call over vector dbs. most teams jump to vectors before measuring anything
I was looking for something similar to try out. Cool!
Cool idea. But is anyone actually building real stuff like this with any kind of high quality?
Every time I hear someone say "I have a team of agents", what I hear is "I'm shipping heaps of AI slop".
+100 for this comment.
[flagged]
Feels like disliking musician for fanaticism towards musical instruments.
[flagged]
I have the same feeling ever since his infamous LLM OS post
Probably just envy.
Obviously it is envy, and not scepticism over a guy who practically lives on Twitter and has unhinged[1] follower base.
1 -https://x.com/__endif/status/2039810651120705569
why not an Obsidian vault with a plugin?
what plugin are you using?
srsly tho this looks slick & love the office refs / will go play with it :)
[dead]