Twitxy Saved

Matt Arderne 🌊

Verified account This account is verified. Learn more

@mattarderne

Builder
Co-founder @ sea.dev
AI for SMB Underwriting
~GMT sea.dev Joined April 2020
2.2K Following 1.1K Followers 7K Posts

Latest posts 5

Matt Arderne 🌊 Verified account @mattarderne
context layer / semantic layer / ontology layer = lap 2 of prompt engineering

bitter pill β†’ πŸ’Š
Jennifer Smith Verified account @scribeceo 1d
Tons of companies are building β€œcontext layer” tools to give agents information that will supposedly help them perform better.

We decided to test what type of context makes a difference and found that more =/= better.

I wrote about the growing context layer market, what info agents actually need (and don't), and what our internal benchmarking showed.
View quoted post on X
2 replies 0 reposts 30 likes 3.1K views View on X
Matt Arderne 🌊 Verified account @mattarderne
you don't need knowledge factories, you need Knowledge Infrastructure β„’
Harrison Chase Verified account @hwchase17 1d
having inconsistent ways to define revenue is exactly why you'd need knowledge factory for your company!

centralize tacit or implicit knowledge, make those readily available to everyone

a big part of ours at langchain is a bunch of definitions of metrics/skills for fetching data from different systems
View quoted post on X
Matt Arderne 🌊 Verified account @mattarderne
in the early days of ocean navigation, all the officers on a ship would perform a daily ritual called "the noon sight" where all the officers would measure the angle of the sun to the horizon and then calculate using a very sophisticated algo their estimated latitude (no longitude)

They did this only daily because it was tedious, and they all (all officers that is) did it in parallel because it was inaccurate.

This resulted in a "Source of Truth" which was the agreed position estimate.

They then acted in accordance with this being held to be true, because they didn't know any better.

This is how I think about the semantic layer fwiw

When measurements where expensive, and inaccurate, it was important to agree, it was important to triangulate, and it was important there was a single agreed best effort.

BUT! If you go onto a ship today, there are at least two GPS devices per person, nevermind the ship nav system, which can measure it's location multiple times per second, at the front back and sides. There are so many different vertically integrated systems all measuring everything they could possibly need, and not a single one of these systems gives one single shit about what the other system measures.

Each vertically integrated system has a jtbd, of which GPS position is an input, but they are NOT worried about what the other ones have to say about it.

GPS itself has many inputs, which triangulate to a best guess location, More accurate systems push this further. The Chef doesn't need much more accuracy than an estimated hour of arrival. A rocket needs more detail.

Vertically integrating systems ask questions and the inputs are defined accordingly, usually deeply integrated for resilience.

There is no noon sight anymore.

Map this to the current moment as you see fit.
Nelson Lee Verified account @NelsonXLee 2d
The more I deploy AI in enterprises, the more I realize that the biggest productivity gains come from decoupling AI from existing human workflows entirely.

In our work at The GCC, I keep coming back to one question: "If you were to rebuild a bank in 2026 to be AI-native, what would still need to be true?"

And quite frankly, once you stop designing around human workflows, you realize that much of an enterprise can be broken down into a few hundred discrete systems that can operate independently and coordinate when needed.

We're starting to explore this more deeply with some of our customers.

But I'm increasingly convinced that the future of enterprise AI isn't about augmenting how humans work but about building organizations that no longer need to work the way humans do.

I can already see some interesting new human roles come out of this at FIs.
View quoted post on X
Matt Arderne 🌊 Verified account @mattarderne
+1 this point with all I have

"once you stop designing around human workflows, you realize that much of an enterprise can be broken down into a few hundred discrete systems that can operate independently and coordinate when needed."

If you roll forward a little bit from this idea, then almost everything in circulation looks like it was built wrong and shaped wrong
Nelson Lee Verified account @NelsonXLee 2d
The more I deploy AI in enterprises, the more I realize that the biggest productivity gains come from decoupling AI from existing human workflows entirely.

In our work at The GCC, I keep coming back to one question: "If you were to rebuild a bank in 2026 to be AI-native, what would still need to be true?"

And quite frankly, once you stop designing around human workflows, you realize that much of an enterprise can be broken down into a few hundred discrete systems that can operate independently and coordinate when needed.

We're starting to explore this more deeply with some of our customers.

But I'm increasingly convinced that the future of enterprise AI isn't about augmenting how humans work but about building organizations that no longer need to work the way humans do.

I can already see some interesting new human roles come out of this at FIs.
View quoted post on X
Matt Arderne 🌊 Verified account @mattarderne
i'm actually amazed that there isn't surge pricing for moving further away from kids. the airline knows who has kids, what age, etc.

Efficient Market Hypothesis? more like lazy clowns at the airline data team leaving $100 notes on the floor
Mike Bird Verified account @Birdyword 2d
Huge shoutout to the guy with the seat next to us in the economy bulkhead (where the bassinet goes). Boarded for a 7hr flight, saw the baby and his face dropped. He called the stewardess over and plaintively asked if it was still possible to pay for an upgrade. It was not.
View quoted post on X