ICYMI 2026-06-20: Structural Risk
Our weekly roundup of signals from the AI noise — for humans leading change.
The AI Capex Ledger
On the broad implications of AI on the economy. Not necessarily that there’s a bubble; it’s more complicated than that. Consider four interconnected ledgers: infrastructure, clouds and hyperscalers, token buyers, and “macro.” Each layer’s revenue is a cost for the one above it. Orgs buy compute so long as it produces results. But where are the results? That’s the bottom line (literally) for you: ultimately, you’re paying for this. IOW, unless you architect intelligence, the bottom layers are counting on your margins. (H/t Tyler Cowen)
A frontier without an ecosystem is not stable
Satya Nadella offers a refreshingly adult perspective on the relationship between AI and human capital. The two are interdependent; AI isn’t a replacement for people. My take: in a world where AI labs want to lock you into intelligence-on-tap (at an initially subsidized cost), architecting context around your data and IP becomes your moat.
Doom Trolling
Cal Newport on frontier labs’ incessant fearmongering as a way to gain attention or regulatory capture. Or perhaps they believe their rhetoric, which would put them in an ethical bind. Anthropic’s fearmongering has now led (perhaps wittingly) to the U.S. government slapping export restrictions on them. This is a risk if you build solutions around these proprietary systems. (NY Times gift link)
The Heart of LLMs
Melanie Mitchell on the intrinsic limitations of LLMs. Experts have differing opinions on these systems’ capabilities and constraints. I’m with Yann LeCun: “A system trained on language alone will never approximate human intelligence, even if trained from now until the heat death of the universe.” That doesn’t mean they’re not useful, but the degree of intelligence they demonstrate at performing day-to-day tasks is almost entirely dependent on how you architect their context.
What it feels like to work with Mythos
Last Friday, a client and I experimented with Anthropic’s Fable. I’m glad we did: a few hours later, the U.S. government slapped export restrictions, forcing Anthropic to pull the model for everyone. It’s still unavailable, but if you want a sense of how Fable/Mythos differs from other models, check out this post by Ethan Mollick. My take: these advanced frontier models are very expensive and energy-intensive for most day-to-day tasks. But for a certain range of problems (e.g., research, which is what my client and I did with it,) they’re unparalleled. The risk: you lose visibility and agency in the decision-making process.
See you next week!
— Jorge