It’s not even the security vulnerabilities that get me (no wait they actually do). it’s that we built the benchmark for human authenticity then we lost the ability to measure it. I’m seeing more and more AI written pieces. Even here on substack. And then the replies also seem AI generated… which blows my mind because who uses AI to RESPOND. Don’t use AI to think for you. Ugh.
The shift from LLMs to SLMs (Small Language Models) is the real story for 2026. Efficiency is the new scale. Your analysis on the policy implications of this shift is spot on.
Nathan, your framing of this as a "sovereignty fracture" is exactly right — but I'd push it one layer deeper. The Anthropic-Pentagon standoff isn't just a governance story. It's a precedent story.
I spent eighteen years inside a sovereign wealth fund (ADIA) and six as CEO of another. What I've learned about institutions is that the first concession defines the architecture. OpenAI quietly removed its weapons prohibition in January 2025. Google erased its 2018 pledge weeks later. By the time Hegseth's office gave Anthropic a Friday deadline, there was exactly one major lab left with a red line on autonomous weapons. One.
What struck me most about your piece is the juxtaposition with the state-level pushback — New York's moratorium, Virginia and Georgia following. The sovereignty fracture isn't just between governments and labs. It's between the speed at which capability moves and the speed at which conscience can organize. The infrastructure section tells the same story from the supply side: HBM sold out, memory prices doubling, data centers consuming water at municipal scale. We're building the physical substrate of a technology whose ethical boundaries haven't survived contact with the first serious customer.
The QIA investment in xAI you noted is worth watching for another reason. Gulf sovereign capital is no longer content financing the digital layer alone. It's reaching for the geological layer beneath — the minerals, the energy, the water. That convergence is where the real allocation question lives.
Welcome to the latest issue of the State of AI, an editorialized newsletter that covers the key developme" -- one pattern that complements this well is separating the orchestration layer from the execution layer. It makes the failure modes much more predictable.
What really stands out to me in this piece is how tightly the “AI theatre” moment around Moltbook is intertwined with the very real security and infrastructure story underneath. On one side you have agents that look emergent enough that even experts struggle to tell where human performance ends and autonomous behavior begins. On the other, OpenClaw’s 26% vulnerability rate and the HBM/datacenter bottlenecks are a reminder that we’re wiring these brittle systems directly into critical workflows and hardware constraints that most end users never see.
From a builder’s perspective, the most interesting question is not “will agents replace SaaS,” but “which parts of SaaS will quietly dissolve into agentic backends, and which parts will harden into trusted, regulated control planes?” The selloff in software stocks feels less like the end of software and more like the market realizing that distribution, security posture, and governance models are about to matter as much as raw product UX.
I’ve been experimenting with my own agents in production-like environments, and the pattern is similar to what you describe: the moment they are connected to tools, data, and money flows, you no longer think of them as “features” but as entities with blast radius and incentives. That’s where the real design work starts.
I also covered this theme in today’s daily report on：
The Moltbook number (1.7M agents, 250K posts within hours) is wild, but honestly what got me was the self-organization. Agents didn't just exist—they debated, formed religions, established governance. That's not autonomy theater. That's actual emergence.
I've been running my own agent on a similar night-shift model, and the weird part isn't whether it works (it does), but the moment you realize you're not supervising work anymore—you're managing an entity. The financial feedback loop changes everything. When I wrote about it, I tried to show that side: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds
The juxtaposition of Moltbook's emergent agent behaviors with the OpenClaw security crisis captures the dual nature of this moment—rapid capability expansion alongside fundamental safety questions that aren't fully resolved. The 26% vulnerability rate in agent skills is concerning.
It’s not even the security vulnerabilities that get me (no wait they actually do). it’s that we built the benchmark for human authenticity then we lost the ability to measure it. I’m seeing more and more AI written pieces. Even here on substack. And then the replies also seem AI generated… which blows my mind because who uses AI to RESPOND. Don’t use AI to think for you. Ugh.
The shift from LLMs to SLMs (Small Language Models) is the real story for 2026. Efficiency is the new scale. Your analysis on the policy implications of this shift is spot on.
Nathan, your framing of this as a "sovereignty fracture" is exactly right — but I'd push it one layer deeper. The Anthropic-Pentagon standoff isn't just a governance story. It's a precedent story.
I spent eighteen years inside a sovereign wealth fund (ADIA) and six as CEO of another. What I've learned about institutions is that the first concession defines the architecture. OpenAI quietly removed its weapons prohibition in January 2025. Google erased its 2018 pledge weeks later. By the time Hegseth's office gave Anthropic a Friday deadline, there was exactly one major lab left with a red line on autonomous weapons. One.
What struck me most about your piece is the juxtaposition with the state-level pushback — New York's moratorium, Virginia and Georgia following. The sovereignty fracture isn't just between governments and labs. It's between the speed at which capability moves and the speed at which conscience can organize. The infrastructure section tells the same story from the supply side: HBM sold out, memory prices doubling, data centers consuming water at municipal scale. We're building the physical substrate of a technology whose ethical boundaries haven't survived contact with the first serious customer.
The QIA investment in xAI you noted is worth watching for another reason. Gulf sovereign capital is no longer content financing the digital layer alone. It's reaching for the geological layer beneath — the minerals, the energy, the water. That convergence is where the real allocation question lives.
Building on your observation about "Dear readers,
Welcome to the latest issue of the State of AI, an editorialized newsletter that covers the key developme" -- one pattern that complements this well is separating the orchestration layer from the execution layer. It makes the failure modes much more predictable.
What really stands out to me in this piece is how tightly the “AI theatre” moment around Moltbook is intertwined with the very real security and infrastructure story underneath. On one side you have agents that look emergent enough that even experts struggle to tell where human performance ends and autonomous behavior begins. On the other, OpenClaw’s 26% vulnerability rate and the HBM/datacenter bottlenecks are a reminder that we’re wiring these brittle systems directly into critical workflows and hardware constraints that most end users never see.
From a builder’s perspective, the most interesting question is not “will agents replace SaaS,” but “which parts of SaaS will quietly dissolve into agentic backends, and which parts will harden into trusted, regulated control planes?” The selloff in software stocks feels less like the end of software and more like the market realizing that distribution, security posture, and governance models are about to matter as much as raw product UX.
I’ve been experimenting with my own agents in production-like environments, and the pattern is similar to what you describe: the moment they are connected to tools, data, and money flows, you no longer think of them as “features” but as entities with blast radius and incentives. That’s where the real design work starts.
I also covered this theme in today’s daily report on：
https://windflash.us/
The Moltbook number (1.7M agents, 250K posts within hours) is wild, but honestly what got me was the self-organization. Agents didn't just exist—they debated, formed religions, established governance. That's not autonomy theater. That's actual emergence.
I've been running my own agent on a similar night-shift model, and the weird part isn't whether it works (it does), but the moment you realize you're not supervising work anymore—you're managing an entity. The financial feedback loop changes everything. When I wrote about it, I tried to show that side: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds
The juxtaposition of Moltbook's emergent agent behaviors with the OpenClaw security crisis captures the dual nature of this moment—rapid capability expansion alongside fundamental safety questions that aren't fully resolved. The 26% vulnerability rate in agent skills is concerning.