cody collier

Thinking and building ⌁ Always learning :: Making mistakes and moving forward

[cody collier @cmcollier]

@reedes Good luck and safe travels!

[cody collier @cmcollier]

As agents become the new search customer, there's so much room for re-thinking previous assumptions and so much room for creative new thinking and building. It's quite refreshing.

[cody collier @cmcollier]

Retrieval is so back.

[cody collier @cmcollier]

@signulll It's the narrow waist lesson:
https://t.co/toryeYZnoX

[cody collier @cmcollier]

MCP and the narrow waist lesson

Is MCP making a mistake? Why are they trying to expand beyond the simplicity of tool call and response?

I don't debate the benefit of creating a bridge between the unstructured language world of LLMs and the structured world of programming and data. This ability is unlocking many new paths in software and system interactions.

However, if you look at the most recent version of MCP docs, you'll notice increasing complexity. The proposed fundamental server primitives include tools, resources, and prompts. The client primitives are sampling, elicitation, and logging. The definitions may not be what you think and are sometimes vague and overlapping.

The adoption and evolution of the Internet taught us some important lessons in interoperability. Notably the narrow waist of IP (Internet protocol) at the middle of the TCP/IP hourglass model.

Maybe MCP should draw from this lesson and focus on the narrow waist. Tool call and response with arbitrary content. It's the simple clarity of tool name, params, response, and description. The beauty of a bridge from unstructured to structured, akin to the simplicity of IP.

[cody collier @cmcollier]

There's lots of talk in tpot about mcp being a mistake. Here's my view from last Fall.

[cody collier @cmcollier]

MCP and the narrow waist lesson

Is MCP making a mistake? Why are they trying to expand beyond the simplicity of tool call and response?

I don't debate the benefit of creating a bridge between the unstructured language world of LLMs and the structured world of programming and data. This ability is unlocking many new paths in software and system interactions.

However, if you look at the most recent version of MCP docs, you'll notice increasing complexity. The proposed fundamental server primitives include tools, resources, and prompts. The client primitives are sampling, elicitation, and logging. The definitions may not be what you think and are sometimes vague and overlapping.

The adoption and evolution of the Internet taught us some important lessons in interoperability. Notably the narrow waist of IP (Internet protocol) at the middle of the TCP/IP hourglass model.

Maybe MCP should draw from this lesson and focus on the narrow waist. Tool call and response with arbitrary content. It's the simple clarity of tool name, params, response, and description. The beauty of a bridge from unstructured to structured, akin to the simplicity of IP.

[cody collier @cmcollier]

@softwaredoug .-- .-- .-- / -.. --- - / ... --- ..-. - .-- .- .-. . -.. --- ..- --. / -.. --- - / -.-. --- -- / ... .-.. .- ... .... / ... . .- .-. -.-. .... / .---- ----- .----

[cody collier @cmcollier]

Agents have induced lots of conversation around search implementation details. This has been useful for revisiting all kinds of assumptions baked into search systems thinking.

I keep coming back to how agents have led us to relax the contraint of low latency.

If I have more time, then I'm allowed more compute and a more diverse set of algorithms at query time. Instead of pre-computing work, which is currently expressed in leveraging trained embedding and encoder models, I can explore new ways to solve the same problems at query time.

You might still end up with a query time model, but I suspect all the work and resources for managing embeddings would start to look like a poor tradeoff. I think we can likely get the incremental relevancy boosts in different ways while still building on the more efficient core of a lexical index.

[cody collier @cmcollier]

The recent attention to skills is starting to rival the noise around mcp from months back.

I suspect software engineers are optimizing workflows too early, adding too much complexity, and not being precise enough with context building.

[cody collier @cmcollier]

Agents have induced lots of conversation around search implementation details. This has been useful for revisiting all kinds of assumptions baked into search systems thinking.

I keep coming back to how agents have led us to relax the contraint of low latency.

If I have more time, then I'm allowed more compute and a more diverse set of algorithms at query time. Instead of pre-computing work, which is currently expressed in leveraging trained embedding and encoder models, I can explore new ways to solve the same problems at query time.

You might still end up with a query time model, but I suspect all the work and resources for managing embeddings would start to look like a poor tradeoff. I think we can likely get the incremental relevancy boosts in different ways while still building on the more efficient core of a lexical index.

[cody collier @cmcollier]

Context is the new customer for search

[cody collier @cmcollier]

The recent interview with Jensen Huang is so good. I've been listening to it repeatedly over the last day or two. It's definitely worth the repetition.

I really appreciate Jensen's speaking style, and even seeing him drift a little outside the normal bounds. Sure things got intense, but I can hear in his thinking a real sense of depth, framing around different timelines, and maturity that comes with experience.

[cody collier @cmcollier]

@Plinz Concise. Tight. Generalizes with the abstraction ladder. Allows for reframing based on definition of distribution. Interesting. I like it.

[cody collier @cmcollier]

It's ok to have a nuanced opinion about AI.

[cody collier @cmcollier]

@habibislop @Steve_Yegge So you want a filter for only AI native jobs? Hm.

[cody collier @cmcollier]

"I found it by systematic exhaustive search" -- paper author

This abstract concept has been on my mind a while. It seems like enumerating all possible permutations of some domains could help with understanding and lead to innovations. It's straight forward but I don't see it discussed much. It's even more interesting in our age of compute and LLMs.

the paper where I saw the quote, interesting in its own right: https://t.co/zeUF8vAVM1

[cody collier @cmcollier]

There are some interesting points here about state categories. Of particular note is the memory state and the organizational data state.

I wonder though if it's clear cut. I suspect much of agent operations can draw on ephemeral indexes of the org data. Then you can inject most of the memory state as artifacts back into the org data layer. It's like the benefits of the Google play with more model flexibility.

I've been thinking about this awhile. Maybe it's time for some building...

[Jaya Gupta @JayaGup10]

https://t.co/Lsuyx2hvpA

[cody collier @cmcollier]

@JayaGup10 Interesting points re: state. Where do you think modern enterprise search players fit (Glean, Atolio, etc)?

[cody collier @cmcollier]

I suspect success with coding agents is inversely correlated to the average number of tokens per session.

From observing online conversations, it seems like many people are having lengthy multitask sessions, use numerous skills tools mcps, and far too many tokens per session. Many people are also complaining about usage limits and pointing fingers outward, which is a symptom of the same.

I'm a bit puzzled. My intuition is popular agent harnesses and general usage are veering off track.

[cody collier @cmcollier]

It was an acciedent of sorts, that I ended up working at a couple of ISPs early in my career. It was beneficial to learn the fundamentals while working support and spending time in a noc. There's a lot to learn working as a network engineer and linux admin. The value of this experience has ebbed and flowed over time. With the AI code generation world it seems like the experience is increasing in value again.