I write more notes than full-length articles. Writing a note helps me structure a fleeting thought, create a narrative around it and explore epsilon-extensions around it. These notes appear as LinkedIn Posts and/or Substack Notes — both great places to share.
I find it useful to reflect back on what I've written to see broader patterns. Also helps reach readers again.
So here goes the monthly harvest from the October month.
The new Model2Vec series of models were created by distilling large models into light weight embedders: fixing the vocabulary, getting token embeddings and storing them. They are 500x embedders, take much less space and promise a wonderful sweet spot for semantic information retrieval / RAG.
I tested out the model by quickly adding a plugin to Ragpipe — a lean, flexible, configurable RAG framework I’m building, and embedding 40k documents. The figure in the note clearly show how drastically you can reduce the ingestion time by using Model2Vec. Now, they promise even better models — looking forward to them.
While Swarm is minimalistic, I got pretty excited about how they introduced a single new programming primitive to enable building complex agents. Took a deep dive.
Wrote a full article on the topic.
Building AI Agent Swarms
·
October 17, 2024
Building AI agents has become the holy grail of modern AI software development, with everyone from startups to tech giants eager to harness the power of communities of AI agents to automate complex tasks. However, as exciting as the prospect of creating autonomous, intelligent systems may be, the reality is that getting them to work reliably is incredib…
They rightly say the framework is educational and initial, but I’m curious what abstractions are missing in Swarm. Compared to other agent frameworks, they are missing a couple of layers/features:
memory layer. agents are stateless. developer faces the additional responsibility of maintaining history and filtering history into per turn context. In comparison, Crew has short- and long-term memory.
no explicit execution graphs. hard to steer control if want to enforce global communication patterns, say round-robin among agents on some condition. Autogen has external manager to orchestrate.
no message passing. many agent frameworks carry out orchestration via sending messages between agents. Do we lose something by not having explicit messages between agents?
what else?
There is plenty of hype around copilots, and how 25% of Google’s code is now written by AI. However, I don’t find many thinking about the end-game, how will it look like at the end of the tunnel?
There is considerable excitement around Cursor, Zed and other AI programming assistants. OpenAI released Canvas with the ability to edit-with-LLMs in place. Seemingly minor integrations of LLMs in the coding/editing workflow seems magical, move us towards the AI's new UI. I'd like to decode what the magic is about.
The ability to modify what you write by a combination of:
External instructions.
Inplace selection and instructions.
Write notes in the margin.
Translating to another domain
The above atomic actions, when integrated seamlessly into your writing tool (not just coding, but writing in general) augment your productivity and creativity by leaps and bounds. Are there other atomic actions that are coming up soon?
Who is the long context RAG leader? O1-preview or gemini-1.5pro or .. ? read here.
John J Hopfield won the Noble Prize for Physics. I knew about Hopfield networks but had never tried to learn more about Hopfield earlier. I found this , very insightful, elegantly written, note by him from 2018 (when he won the Franklin medal).
The note is framed around the problem of "Now what".
Every scientist, every scholar, every writer, every artist ..... faces this vexing issue of what to work on each day. Most respond by working today to extend a little yesterday's line of thought, or line of measurement, or plot development or ... . Most, if forced to ask "Now What" at a more fundamental level, or for a long-term commitment of effort, endeavor to not much change the sandbox in which they have been happily playing.
Now what is fundamentally a question about explore-exploit tradeoffs, to restart or to continue. Read my note about their note.
This continues my rant on a topic close to my heart. Inspired from Chris Molnar’s post. Read here.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.