Christopher Mims's Keywords column on B2 this weekend — "Three Reasons AI Is Now More Reliable Than Ever" — pulls apart the last nine months of quiet engineering progress inside the frontier labs. It is short, non-hyperbolic, and its conclusions are the most important thing a client needs to understand about AI this year: the chatbots are not smarter, but the systems around them are — and that is what's finally made them useful.
Mims's three mechanisms, in plain English:
- Models look things up. Modern AI agents don't rely on what they memorized in training; they search the live web, proprietary databases, and internal documents at inference time. Anthropic, OpenAI and Google all report double-digit drops in fabrication rates since this shift.
- Models call tools. Instead of trying to "do math" in their head, the models now invoke calculators, code interpreters, spreadsheets, and domain-specific APIs. Coding agents like Claude Code and Cursor run tests against the code they write before showing it to the user.
- Models check each other. Multi-model "council" architectures run the same question through several different models (often from different providers) and only return an answer when they agree. Mims reports leaked Anthropic source code that confirms this pattern is already in production.
Why This Matters For The AI Sleeve
Our 20% AI / Cyber / Data theme has been the best-performing bucket in the book year-to-date (+38% cap-weighted through Friday's close). The bear case on this sleeve has always been "the frontier labs will commoditize each other and margins will collapse." Mims's column is the single clearest piece of evidence I've seen in 2026 that the bear case is early.
Here's the read-through:
- The inference layer is eating the training layer. The economic value of a frontier model increasingly lives in the orchestration around it — tool use, retrieval, multi-model checking. That favors companies that own the deployment stack, not just the weights.
- Claude Code is real. Anthropic's coding agent is, per Mims, already generating substantial revenue and the leaked source code shows a level of engineering sophistication that is unusual. Anthropic is not public, but its compute partner (AMZN) and its chip supplier (NVDA) are.
- Cybersecurity rises. If every knowledge worker gets an AI agent that can read their company's data, the security surface explodes. CRWD, PANW, ZS stay core defensive AI infrastructure names.
Our Current AI / Cyber / Data Holdings (20% Theme)
MSFT, GOOGL, META, AMZN
Platform owners who both sell inference and deploy it in their own products. META gains on open-weight infrastructure.
NVDA, AVGO, TSM, VRT
Picks-and-shovels. Wins regardless of which frontier lab is ahead. VRT adds the data-center thermal story.
CRM, NOW, PLTR, SNOW
Application-layer AI. Agents and copilots that plug into existing enterprise data.
CRWD, PANW, ZS, S
AI expands the attack surface. Endpoint + network + identity. Core defensive layer.
What We're Watching
Two companies are mentioned by name in Mims's column without being public: Anthropic (frontier lab behind Claude) and NineTwoThree (consulting shop implementing the "council of models" pattern for clients). Anthropic is reportedly raising at a $350B+ valuation; Amazon is the largest outside investor. If Anthropic IPOs in 2027 it will be the most-watched AI listing since the OpenAI filing telegraphed this week.
How Clients Should Think About The AI Theme In Q2
The AI sleeve has done a lot of work in your portfolio over the last 15 months. Three things to check in your next review:
- Concentration. If NVDA or MSFT has drifted above 6% of your total account on price alone, we're going to trim and rebalance across the sleeve.
- Tax lot management. Wins are mostly long-term now. If you have short-term lots from a recent add, let them age into long-term before rebalancing where possible.
- The private layer. If you've asked about getting Anthropic / OpenAI / Scale AI exposure pre-IPO, the two accessible proxies today are AMZN (Anthropic's cloud partner) and publicly-traded custodians of secondary shares. We'll walk through them.
Mims's column does not tell you to sell the AI sleeve. It tells you the thesis is maturing from "will it work" to "who captures the margin." We stay invested, we stay diversified within the theme, and we manage the sizing.
Bottom Line
The AI trade is shifting from model hype to deployment economics. Mims's three mechanisms — lookup, tool use, multi-model checking — are the first rung of a real productivity story that has been promised for three years. We remain at 20% of the model portfolios, cap-weighted across the four buckets above.