This Week

This week I’ll talk about my AI orchestrator project, then dive into the books and movies, plus an NBA playoff preview.

Technology

Anthropic is back in the headlines, this time in the fear-mongering variety. Anthropic has always leaned hard into AI safety. Last month they took on the US government, refusing to allow certain military and surveillance uses of their models without protections in place — the concern being that their models are now capable enough to enable autonomous killing or a surveillance-state apparatus. This week, they announced that their latest model, Mythos, is strong enough at detecting code vulnerabilities that it poses a serious security threat in the wrong hands. Mythos won’t be released publicly. It’ll only be available to a small set of US companies running critical infrastructure — Google, Nvidia, a few others.

I have no doubt Anthropic’s models are extremely capable at this point. But I also can’t shake the sense that there’s something else going on underneath. The cynical read is that this is a PR play — lean into doomsday framing and you’re implicitly claiming you’re the top model company. The credulous read is that Anthropic is simply that principled. I think the interesting answer is that both are true at once. Anthropic’s safety posture is sincere and commercially convenient, and when a company’s values and its market positioning line up this perfectly, that’s worth noticing. It’s not a scandal. It’s a tell about where the frontier actually is.

A detailed profile of Sam Altman also came out this week. Nothing new really surfaced, but it re-litigates the question of whether Altman is a good person or a shrewd businessman. The more that comes out, the clearer it is he’s the latter — and honestly, if you’re running one of the most important companies in the world, there’s only so much altruism on offer. Your job is to make money.

The most interesting thread, for me, was the theory of why OpenAI was structured as a non-profit in the first place. It gets called OpenAI’s original sin, and I almost agree — but I think the more honest framing is that the non-profit structure wasn’t a sin so much as a necessary fiction with an expiration date. In 2015, a for-profit OpenAI wasn’t going to land Ilya, Dario, and the rest of that cohort. You needed the mission. The research-lab framing was the price of admission for the talent that built the thing. The sin wasn’t choosing it; the sin was pretending the fiction could survive contact with a product doing a billion dollars a year.

Once the ChatGPT moment hit and OpenAI suddenly owned one of the fastest-growing applications in history, Altman had to reckon with their identity. He doubled down on what was already obvious: OpenAI would remain a “research company” only insofar as that framing could attract top talent and keep them from defecting to Google. From a pure business lens, it worked. OpenAI is beating Google in the AI race, in my opinion. But it cost them their earliest believers, the ones who felt betrayed — most notably Dario Amodei, who went on to found Anthropic and take a good chunk of the research bench with him.

And while ChatGPT jumped to a commanding early lead with hundreds of millions of weekly users, the dust is settling and the economics are coming into focus. AI is brutally capex-heavy for the frontier labs, and the question becomes: who pays? There are two markets, consumer and enterprise. AI has opened consumer wallets somewhat — people are paying for the $20 and $200 tiers — but most users stay on the free version. Anthropic saw this early and went straight at enterprise. They built a moat around coding, where no one has beaten them, and businesses have both the budget and the investor pressure to spend on AI. Once an enterprise wires its workflows into a specific model, that’s a real moat. The question is how long it lasts — and I think the honest answer is two to four years, not forever.

Which brings me to the piece I’ve been turning over all week: where should the intelligence layer actually live? Should the model company own both the intelligence and the harness, or should the intelligence layer sit locally, inside the business, agnostic to whichever model is behind it?

The more I sit with it, the more this feels like the iPhone and App Store moment all over again. The iPhone changed how we navigated the internet. We were no longer tethered to a desktop or laptop at home — we could carry the internet with us everywhere. But the device itself was only half the story. To make it genuinely useful, companies built apps, and “there’s an app for that” became a cultural through line. Once you had the next-generation interface in your pocket, you wanted it to do everything for you, and every “I wish my phone could do this” became a product.

AI is sitting in the same spot right now. The new interface is agents. Whether it’s Claudebot or Clawbot, you’ve got agents working locally on your machine, texting you updates, and you text back what to do next. And the apps in this world are Skills files. The same impulse that gave us “there’s an app for that” is reappearing as “I wish my agent could do this” — and a Skills file is the answer. It’s the new unit of extensibility.

That reframes the moat question entirely. Apple made enormous money on the device and on App Store fees, but the value that third parties built on top of the platform — Uber, Instagram, Spotify — dwarfed anything Apple shipped itself. If Skills are the apps, the durable value accrues to the orchestration layer and the Skills that live inside specific businesses, not to the model underneath. The model becomes the device. Your business logic, your PRD disciplines, your evaluation thresholds, your internal tools and data access — that’s what makes it useful for your context, and none of it belongs to the model company. Which is why I think Anthropic’s enterprise moat is real but finite. It’s a lease, not land.

One thing my company is exploring is exactly this: an orchestration engine local to the business, with the model acting as the harness rather than the brain. The disciplines of how to write a PRD, the requirement to define evaluations and success thresholds before a line of code ships — all of it is being codified into Skills files. The model runs against them. Business logic lives in markdown files that spell out the goal (say, 90% note completion or 90% first-pass clean claims), and Skills teach the model how to reach into internal services and data to get there.

This reinforces something I keep coming back to: institutionalize the thinking. Before you do anything, ask how it becomes part of a system you can iterate on indefinitely. That’s where AI’s real leverage is — cheap, continuous execution. Models work 24/7 because they don’t get tired and don’t care about labor rights. Give one a clear goal and the Skills to reach the tools and data it needs, and that’s where the magic happens.

It’s fascinating to think about how work reshapes in this era, and unsettling to place your own role inside it. Once you’ve built the system, what comes next? Will workers become system operators? That feels replaceable. What I think holds up is judgment — but I want to be specific about what I mean, because “judgment is the moat” has become the universal cope of the 2026 knowledge worker, and most people saying it can’t tell you what they mean.

Here’s what I mean. The work that survives is encoding judgment into systems, not operating them. For a PM, that’s knowing which hypotheses are worth testing in the first place, which evaluations actually measure the thing you care about, which workflows deserve to be codified and which are still too early. It’s taste applied to process design. The people who win are the ones building the encoding layer — the Skills, the evals, the orchestration — not the ones running it. That’s automatable eventually, sure. But the boundary keeps moving, and the work is to stay on the right side of it by continually building the next system.

So that’s what I spent the weekend on. I built a system that ingests raw material from my inbox: articles I read, articles I meant to read, YouTube videos I queued up and never watched, Slack threads from coworkers, call transcripts, customer data, sales calls. All of it becomes raw input. The system processes each piece against the domains I care about — healthcare, technology, the macro economy, being an AI-native PM, building the new billing platform — and asks a structured set of questions. Does this provide evidence for my top hypotheses, or against them? Is it a strong input at all? If not, file it away — not delete, just shelved until it becomes relevant.

What I’m building is a knowledge base with a sharpened point of view baked in. I can ask it what I know about technology, or what the top themes are right now, and it answers from my synthesis rather than from its generic training data. I review, push back, and sharpen. It helps me draw connections I wouldn’t have seen and does the mundane filing work of tying each source to a live hypothesis.

If the thread isn’t obvious yet: this is how knowledge compounds. Everything I take in every day gets layered into a base only I can query. When I ask what the top challenge in healthcare is, I don’t get a generic answer — I get one built on my material, my framing. No one else can replicate it because no one else has my base. That’s the differentiator I’m betting on.

I’ll also name the obvious failure mode, because I’ve been thinking about it: a system like this is only as good as the discipline I bring to reviewing it. If I stop reviewing three weeks in, the synthesis drifts, and I end up with a confident-sounding knowledge base built on my own unexamined priors — which is worse than having no system at all, because it launders bad thinking as good thinking. The hedge is to keep the review loop tight and the system small until I know the habit sticks. Over time, I want it to become a living thing that pings me proactively — this looks like a trend you should investigate. But I don’t want to over-engineer out of the gate. Use it, learn what it actually needs, iterate. Don’t build the perfect system on day one.

Books

This week I started working through the 2025 Booker Prize short and long list. For books, my two anchor award cycles this year will be the Booker and the Pulitzer. I began with Susan Choi’s Flashlight.

Flashlight is a 464-page family drama and historical mystery about identity, memory, and political exile. It opens with the disappearance of Serk, a Korean immigrant, on a Japanese beach during an evening walk with his ten-year-old daughter Louisa, who is later found half-dead on the shore with no memory of what happened. The novel then moves across decades, focused on the strained triangle of Louisa, her estranged mother Anne, and the absence of Serk. It’s billed as multigenerational, but in practice it stays tight on these three.

The backbone is Serk’s history — born to Korean parents in wartime Japan, emigrating to the US rather than North Korea, living a version of the American dream as a green-card holder in the 1960s and 70s. He meets Anne, they marry, and the marriage is profoundly unhappy — two people carrying heavy scars, rarely speaking, and when they do, speaking in sharp edges. Louisa inherits all of it, and only at the end learns to accept her parents on their own flawed terms. The novel closes with the revelation of what happened to Serk and a quiet hospital-bedside scene of forgiveness.

The book opened strong for me — genuinely mysterious, with flawed characters navigating a hard situation. But the length didn’t serve it. The plot never thickened, emotionally or suspensefully. Resolutions arrived quickly and felt unearned. It wasn’t a bad read, but at 400+ pages it didn’t reward the journey. 3/5, B.

Movie

This week I watched Train Dreams on Netflix as part of my Oscars 2026 run — it was the only eligible film on Netflix, and my subscription was about to lapse.

Train Dreams is a poetic, melancholic film following Robert Grainier, an Idaho logger navigating loss and upheaval in the early 20th century, framed by a third-person narrator recounting his life. Grainier is introduced as a directionless young man working in logging, who finds his first real spark meeting his wife Gladys and later their daughter Kate. Money is tight — the default condition of blue-collar life in that era — but Robert and Gladys still dream of something more: a horse, some cattle, a farm, a peaceful life together without the long stretches apart while he works. Then a fire comes through while he’s away and takes the cabin. It’s never made explicit that Gladys and Kate die, but their absence becomes the shape of the rest of his life. He rebuilds in the same spot and waits, quietly, until the end.

The film is slow-paced, atmospheric, and meditative — memory, grief, and the transformation of the American West. You move with Robert from innocence to love to loss to acceptance. It flows, even if it never quite reaches the highs it gestures at. Still, it’s soulful, and at moments it touches something ethereal. 3.5/5, B+. I’d slot it between last week’s Hamlet and Frankenstein.

Sports

With the last day of the NBA regular season this week and several matchups still unresolved, I’ll use this post as a combined end-of-season and play-in prediction. Next week will be full playoff predictions.

East — top 6: Pistons, Celtics, Knicks, Cavs, Hawks, Raptors.

In the 9/10 game, I’ve got the Heat’s veteran savvy overcoming a hot but young Hornets team. It’s wild that the Hornets have been this good and are still stuck in the 9/10. In the 7/8, I’m taking the Magic. The Magic have been disappointing, but the Sixers feel more so — Maxey’s breakout is real, but George and Embiid’s injuries and age, plus Edgecombe plateauing, don’t inspire confidence. Heat then beat the Sixers for the 8 seed.

Magic 7th seed. Heat 8th seed.

West — top 6: Thunder, Spurs, Nuggets, Lakers, Rockets, Wolves.

The 9/10 is Clippers–Warriors, playing tonight. I don’t expect Golden State to play their starters; Clippers and playoff Kawhi take it. In the 8/9, Suns–Blazers — the Suns have been the more pleasant surprise all season and I don’t see them faltering. Clippers vs. Blazers then comes down to experience: Ty Lue and Kawhi. Clippers take it.

Suns 7th seed. Clippers 8th seed.