Fintech is now ~4% of global financial services revenue and growing 4x faster than incumbents. So what should a bank do about it? Founder at MIT Cryptoeconomics Lab, @ccatalini on Tokenized recently said that banks should disrupt themselves before Fintech does. “I think if I was leading a global bank, I would just embrace it and rather disrupt myself than waiting to be disrupted.” “Right now, customers will want more choice. They will want better products and services. You can be ahead of that game.” “If you look at the music labels, the ones that embraced streaming and understood that the economics of distribution of music were going to zero, and they had to win in a different game.” “The blockbuster late fees in the financial services industry are going away, and we should play by new rules.” 🎙 Listen to the latest episode on Tokenizedpod(dot)com

1/ Somewhere in your company today, an expert rejected a perfectly plausible answer. A controller caught a revenue figure that counted deferred bookings. A security engineer blocked a release that passed every test but would have handed an AI agent write access to the production database.

🚨 You can't safely automate what you can't reliably measure and verify. Verification means checking that the agent did what you intended. As execution gets cheaper, proving the task was done right becomes the bottleneck to safe scaling.

The marginal cost of trying one more attack has been falling for years. Attackers only need one. Mythos was not a discontinuity. If anything, we now have better tooling (open weights included!) to replace attention-constrained humans with machines and close the bugs we inevitably leave behind. H/T @alighodsi for surfacing the long-term trend.

5/ Fix two is the superpower. The models that ate junior tasks also collapsed the cost of learning and experimentation. Explore more fields. Test more ideas. Discover what you’re unusually good at, then use AI to do more with it. Superpowers, if you deliberately use them.

Safer, more reliable AI requires closing the measurability gap between what humans care about and what agents can measure. If our preferences and context don't shape what agents optimize, being "in the loop" is just rubber-stamping.

At this moment in Q4 2026 it's net good for cyber-defense that we have open weights models that are good at dual use cybersecurity tasks. - If there's been an attacker productivity bump due to AI 2022-26 it's small and close to being in the statistical noise - In contrast, there's a ton of measureable progress in defenders using models to find and fix bugs in highly leveraged codebases - Many defenders have struggled to get access to the best unguardrailed closed weights cyber models - ... and when they do the tokens are dramatically more expensive than open weights tokens at the same cyber intelligence level; CISO orgs are cost centers and these costs are often prohibitive - Ultimately we'll need open weights models on premise on air gapped networks and on small devices where they can detect breaches and attacker implants - We'll also need open weights in situations where defenders can't send secret data to the closed weights model providers (who have security problems of their own!) As long as these conditions persist, we should be pushing for diffusion of frontier dual use open weights cyber capabilities, because they net benefit defenders.

1/ Once you see that verifying AI’s work is the bottleneck, and that checking runs on tacit knowledge, you can predict the hiring shift before seeing the data. We did, back in February. Anthropic’s junior share of technical hires: 27% to 9%.

6/ These dynamics are visible inside @Anthropic than in aggregate labor data. Why? Diffusion. The lab is the extreme case of adoption. The broader economy still has to reorganize around AI, as with other general purpose technologies. That lag is our window to act!

Equilibrium final exam question, though: if working with agentic AI is a form of management (it is), hence requires higher skill (it does), but the distribution of experience/skill in the economy is fixed in the SR, what happens to the *distribution* of hires across firms?

2/ Output is priced in tokens. Verification is priced in human hours. Anthropic estimates that reading a week of Claude output would take 424 hours per person! A week has 168. That’s the verification constraint. And reading is the easy part!

3/ Junior employment grew more slowly where Claude handled more complex tasks. Our model’s Missing Junior Loop: firms need top expertise while automating the work that produces it. They’re hiring from a stock they’re doing less to replenish.

6/ These dynamics are visible inside @Anthropic than in aggregate labor data. Why? Diffusion. The lab is the extreme case of adoption. The broader economy still has to reorganize around AI, as with other general purpose technologies. That lag is our window to act!

5/ Fix two is the superpower. The models that ate junior tasks also collapsed the cost of learning and experimentation. Explore more fields. Test more ideas. Discover what you’re unusually good at, then use AI to do more with it. Superpowers, if you deliberately use them.

3/ Junior employment grew more slowly where Claude handled more complex tasks. Our model’s Missing Junior Loop: firms need top expertise while automating the work that produces it. They’re hiring from a stock they’re doing less to replenish.

2/ Output is priced in tokens. Verification is priced in human hours. Anthropic estimates that reading a week of Claude output would take 424 hours per person! A week has 168. That’s the verification constraint. And reading is the easy part!