\(
\def\cuberoot#1{\sqrt[3]{#1}}
\def\fourthroot#1{\sqrt[4]{#1}}
\def\abspartial#1#2#3#4{\left|\,{\partial(#1,#2)\over\partial(#3,#4)}\,\right|}
\def\absdeltal#1#2#3#4{\left|\,{\d(#1,#2)\over\d(#3,#4)}\,\right|}
\def\dispop#1#2{\disfrac{\partial #1}{\partial #2}}
\def\definedas{\equiv}
\def\bb{{\bf b}}
\def\bB{{\bf B}}
\def\bsigma{\boldsymbol{\sigma}}
\def\bx{{\bf x}}
\def\bu{{\bf u}}
\def\Re{{\rm Re\hskip1pt}}
\def\Reals{{\mathbb R\hskip1pt}}
\def\Integers{{\mathbb Z\hskip1pt}}
\def\Naturals{{\mathbb N\hskip1pt}}
\def\Im{{\rm Im\hskip1pt}}
\def\P{\mbox{P}}
\def\half{{\textstyle{1\over 2}}}
\def\third{{\textstyle{1\over3}}}
\def\fourth{{\textstyle{1\over 4}}}
\def\fifth{{\scriptstyle{1\over 5}}}
\def\sixth{{\textstyle{1\over 6}}}
\def\oA{\rlap{$A$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obA{\rlap{$A$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obX{\rlap{$X$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obY{\rlap{$Y$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obZ{\rlap{$Z$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obc{\rlap{$c$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obd{\rlap{$d$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obk{\rlap{$k$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\oba{\rlap{$a$}\kern2pt\overline{\phantom{\dis{}I}}\kern.5pt}
\def\obb{\rlap{$b$}\kern1pt\overline{\phantom{\dis{}t}}\kern.5pt}
\def\obw{\rlap{$w$}\kern1pt\overline{\phantom{\dis{}t}}\kern.5pt}
\def\obz{\overline{z}}\kern.5pt}
\newcommand{\bx}{\boldsymbol{x}}
\newcommand{\by}{\boldsymbol{y}}
\newcommand{\br}{\boldsymbol{r}}
\renewcommand{\bk}{\boldsymbol{k}}
\def\cuberoot#1{\sqrt[3]{#1}}
\def\fourthroot#1{\sqrt[4]{#1}}
\def\fifthroot#1{\sqrt[5]{#1}}
\def\eighthroot#1{\sqrt[8]{#1}}
\def\twelfthroot#1{\sqrt[12]{#1}}
\def\dis{\displaystyle}
%\def\definedas{\equiv}
\def\bq{{\bf q}}
\def\bp{{\bf p}}
\def\abs#1{\left|\,#1\,\right|}
\def\disfrac#1#2{{\displaystyle #1\over\displaystyle #2}}
\def\select#1{ \langle\, #1 \,\rangle }
\def\autoselect#1{ \left\langle\, #1 \,\right\rangle }
\def\bigselect#1{ \big\langle\, #1 \,\big\rangle }
\renewcommand{\ba}{\boldsymbol{a}}
\renewcommand{\bb}{\boldsymbol{b}}
\newcommand{\bc}{\boldsymbol{c}}
\newcommand{\bh}{\boldsymbol{h}}
\newcommand{\bA}{\boldsymbol{A}}
\newcommand{\bB}{\boldsymbol{B}}
\newcommand{\bC}{\boldsymbol{C}}
\newcommand{\definedas}{\equiv}
\newcommand{\half}{\frac{1}{2}}
%\newcommand{\slfrac}[2]{\raisebox{0.5pt}{$\scriptstyle{}^{#1}\!/\!_{#2}$}}
\def\slfrac#1#2{\raise.8ex\hbox{$\scriptstyle#1$}\!/\!\lower.5ex\hbox{$\scriptstyle#2$}}
\newcommand{\bsigma}{\boldsymbol \sigma}
\newcommand{\abs}[1]{\left|\,#1\,\right|}
\newcommand{\Rectangle}{\sqsubset\!\sqsupset}
\newcommand{\rectangle}{{%
\ooalign{$\sqsubset\mkern3mu$\cr$\mkern3mu\sqsupset$\cr}%
% How to do fractions: for 1/2 use this in place: $\raise{0.8pt}{\scriptstyle{}^{1}\!/\!_{2}}$
% for n/m use this in place: $\raise{0.8pt}{\scriptstyle{}^{n}\!/\!_{m}}$
}}
\)
Agentic Internet and Architectures
Patrick Reany
Starting in May 2026
The links below deal with the topic of how artificial intelligence (AI)
and how agentic architecture will affect internet commerce, browsing,
and finance.
How small-to-medium LLMs can be used on a personal computer in agentware.
The Great Attractor Program (GAP):
GAP defined.
On this website, GAP has evolved into a huge collection of expositions
of science
and mathematcatics topics as view by the organizing principles of:
- Start with a huge space of possible objects/representations under interest
- Apply legitimate projections (coarse-grainings) that remove irrelevant detail
- Track a monotone (entropy, Lyapunov, Fisher information, etc.)
- Arrive at a stable attractor: the representation that cannot be simplified further
Chats with Copilot -- Articles:
Copilot Predicts Four Evolving Computer Technologies to Align With
According to Copilot, the evolving computer/cloud technologies that will dominate
the next ten years are:
- C $\rightarrow$ Rust $\rightarrow$ memory-safe AI-generated code
- SaaS $\rightarrow$ agentic computing $\rightarrow$ on-device AI
- Cloud $\rightarrow$ edge $\rightarrow$ hybrid neurocomputing
- Apps $\rightarrow$ workflows $\rightarrow$ autonomous agents
And Linux would do well to take these agentic/NPU recommendations to heart.
Copilot also provided us with an NPU-era concept of Moore's Law to replace
the traditional one based on faltering improvements in CPU technologies.
Copilot and I do a follow-up on the last chat: Four Evolving Computer Technologies to Align With
How Microsoft's architecture fits
This is Windows Copilot Runtime.
It includes:
- ONNX Runtime
- Windows Semantic Index
- NPU scheduling
- Local model hosting
- System-level AI APIs
- Presence, vision, and audio models
- Recall (vectorized system memory)
This is a follow-up to Four Evolving Computer Technologies to Align With,
in which Copilot and I look at the NPU and the Singular Value Deomposition
(SVD) from the viewpoint of GAP + Entropy.
How Microsoft's architecture fits
This is Windows Copilot Runtime.
It includes:
- ONNX Runtime
- Windows Semantic Index
- NPU scheduling
- Local model hosting
- System-level AI APIs
- Presence, vision, and audio models
- Recall (vectorized system memory)
Conclusion: Yes -- we are entering the NPU-era of Moore's Law
The new curve isn't about transistor density.
It's about tensor throughput per watt.
This is the new exponential:

This scales much faster than CPU IPC or clock speed ever could.
In other words:
Moore's Law didn't die -- it moved.
It migrated from the CPU to the NPU.
The NPU is a tensor engine wearing a neural-network Halloween costume.
What is Stripe?
According to Copilot:
Stripe is the transactional layer that agents use to:
- find products
- compare options
- check availability
- complete purchases
- handle returns
- manage subscriptions
HTML vs JSON vs Markdown?
Which is good for agents, which for humans, which for both? How does
cost enter the equation by tokens?