writing code was never the bottleneck. knowing what to build and getting anyone to care still is.

the next big ai company isn't another agent, it's the audit log for all of them. "which bot did this" is going to be the question of 2027.

optimizing for read-through instead of clicks is the most founder-relevant algo change in years. hooks are out, actually finishing the thought is in. 👏🚀

horizontal agents get you to 80%. the money is in the last 20%: the domain evals, the weird edge cases, and the data only a vertical player ever sees.

the scary version isn't a better model, it's faster atoms. software founders here should be way more obsessed with hardware and manufacturing than we are.

group chat is the skeuomorphic phase. the real unlock is agents with their own threads and handoffs, so humans only show up for decisions.

rent the model, own the memory. and owning the memory turns out to be a data engineering problem, not a prompt problem.

agent swarms turn "we didn't have time to test that" into a line on the compute bill. the bottleneck moves from headcount to how well you set the goal.

make them better. the friction was never the feature, and the orgs that give agents clean, structured answers will win while everyone else adds more captchas.

trust tiers plus one-time cards is what a personal agent protocol looks like in practice.

the home lab is quietly becoming the data center. one mac runs agents, one runs local models, and on-device is coming faster than we think.

every company is about to have this chart. the ones that own their workflow data get to read it first.

Finally in Canada 🇨🇦 👏🍻 The engineering leap at Tesla of distilling HW4 AI into HW3 should be studied and celebrated. In the future, using this kind of distillation tech, will we carry frontier models fully on device in our phones?

For most of human history, we have put thought and consciousness in the category of things that feel almost magical. AI may be slowly taking them out. One of the things I find most fascinating about LLMs is not just what they can do, but what they suggest about us. People sometimes dismiss them as “just predicting the next word.” But I’m not sure that is the diminishment they think it is. A neural network trained around prediction somehow ends up writing, coding, reasoning, planning and making connections that look remarkably like thought. That does not mean the human brain is an LLM. It obviously isn’t. But it does raise the possibility that what we experience as thought may also emerge from neural machinery doing something much more mechanical than we like to imagine. And now consciousness may be next. Doris Tsao, one of the world’s top neuroscientists, is leading a major effort at Astera to understand how the brain generates conscious experience. Their working hypothesis is that consciousness emerges as the brain builds persistent representations of the world and of ourselves within it, then composes those representations into an internal model of reality. This week Tsao and collaborators published new work suggesting that the brain may rapidly alternate between what is consciously perceived and what is suppressed from awareness. It is early work, and certainly not the final answer, but it is another attempt to turn consciousness from something mystical into something mechanical. I find that both unsettling and exciting. First intelligence becomes less mysterious. Then maybe consciousness does too. And I don’t think explaining either one makes it less remarkable. Understanding DNA did not make life less extraordinary. Understanding how stars work did not make the night sky less beautiful. Maybe consciousness is another miracle that becomes no less miraculous once we understand the machinery. The really interesting question is what happens if we eventually understand that machinery well enough to reproduce it.

Grok @bot has a killer feature over its competitors: speed. It generally responds in < 1s. Instinct takes > 10s. Dots > 5s. Muse > 2s. Speed encourages thinking and building. This small detail deserves more attention and appreciation. 🙏

I may need to amend something I’ve been saying: rent the model, own the memory. Coinbase just published a fascinating example of what happens when a company does more than that. They took a relatively small open-weight model and trained it on Coinbase’s proprietary fraud data. On their internal fraud benchmark, the resulting model outperformed frontier models across every detection metric they tested. The actual training run cost less than $100. There are obvious caveats. This is one narrow task, the benchmark belongs to Coinbase, and the real cost includes a lot more than the GPU time. But the architectural point is really interesting. The original open-weight model was not particularly good at the task. The advantage appeared after Coinbase combined an open model with something nobody else has: its own fraud history, outcomes and understanding of what a good decision looks like. So maybe the emerging enterprise architecture is more nuanced than “rent the model.” Rent general intelligence. Own specialized intelligence where your proprietary data and judgment actually create an advantage. A frontier model can remain the generalist. But underneath it, companies may have dozens or eventually hundreds of smaller specialist models trained to do very specific things unusually well. Fraud detection. Pricing. Customer retention. Ad yield. Forecasting. Claims. Underwriting. Whatever is specific to that company. And Coinbase is making this available internally as a skill, so employees can spin up and fine-tune specialist models themselves. Brian Armstrong even suggests that some existing backend services could eventually be replaced by this architecture. Maybe the future enterprise AI stack is not one giant model that knows everything. It is a generalist surrounded by a growing collection of specialists that know your business extraordinarily well. And those specialists may be worth owning.

Websites were built for humans. What happens when your customer stops visiting them? Yesterday Sierra and Meta announced Personal Agent Protocol, with Instinct, Shopify, Stripe, Walmart and others. The idea is pretty simple: create a standard way for your personal AI agent to identify itself, prove it is acting for you, get permission to do certain things, and interact directly with a business. Today, when I ask Instinct to change a flight or make a reservation, it often has to use the web the way I would: open pages, click buttons and fill out forms. That feels temporary. With this kind of protocol, my agent could authenticate with the airline, retailer or bank and deal directly with its API or its own AI agent. I decide what my agent is allowed to do. The company decides what it exposes. OAuth handles the authorization. If this works, I think the bigger change is that the company may no longer control the primary interface with its customer. My personal agent does. That has huge implications. What happens to websites and apps? What happens to UX, upselling, loyalty programs, advertising and even brand differentiation if my agent is increasingly making the choices and completing the transactions? And what does a company need to do to become “agent ready”? Maybe the important digital infrastructure becomes less about building another beautiful front end and more about exposing clean data, APIs, permissions and actions that trusted agents can use. There are already other pieces emerging around this, including protocols for agent-to-agent communication and agent payments. Google’s AP2, for example, is designed to let agents make authorized transactions, including purchases when the human is not present. I wrote a couple of weeks ago that we may be heading toward a world where every person has an agent. This announcement makes that world feel a little less theoretical. If your customer starts sending an agent instead of showing up themselves, is your company ready to do business with it?

Canada is about to wrestle with the same AI question every company is wrestling with: what do you rent, and what do you own? Today, Prime Minister Mark Carney announced a new National Council on Artificial Intelligence, with a mandate that includes accelerating AI adoption, building Canadian AI companies, and strengthening sovereign AI infrastructure. That sounds a lot like the conversation companies are already having about AI agents. At the company level, I’ve argued that you probably don’t need to own the model. Models will change constantly. What matters more is owning what compounds over time: memory, context, workflows, institutional knowledge, permissions, and the rules governing what AI is allowed to do. At the national level, the questions are similar: Where does Canadian data live? Who controls the compute? Can Canadian institutions switch models without rebuilding everything underneath them? Do we retain the IP, talent and capabilities created here? AI sovereignty cannot mean building everything ourselves. It should mean being deliberate about which layers are too important to become dependent on someone else for. Which brings me back to the same principle. Rent the model. Own the memory. Govern what has authority. For companies, that’s an architecture question. For Canada, it may become an economic sovereignty question.