Joe Weisenthal in the UltraDome: AI diffusion gap, Fed's Warsh era, and why the bond market is literally a prediction market

Joe Weisenthal on the AI diffusion gap, Warsh's Fed, and the bond market as prediction machine

Joe Weisenthal, Bloomberg journalist and Odd Lots co-host, sat down to cover three threads that rarely get discussed in the same breath: why AI still isn't showing up in economic data, what Kevin Warsh's communication style actually means for markets, and why the US government bond market is, in a precise technical sense, a prediction market.

The AI diffusion gap

The central tension Weisenthal keeps returning to is the disconnect between AI discourse and economic reality. The people who were most accurate at forecasting model progress — predicting compute buildout, chip purchases, data center construction — have had essentially no predictive power on when any of this shows up in daily life. Those forecasts, he points out, were all downstream of the same extrapolation: correlated lines moving together. The one thing nobody has gotten right is when ordinary people actually feel it.

That gap is visible in how the debate has shifted. The job-loss narrative, Weisenthal argues, was partly a strategic communication choice by the safety community: job displacement feels tractable, generates policy conversations economists will engage with, and doesn't require listeners to buy into sci-fi threat models. The problem is that large-scale job losses haven't materialized, which makes it easy to dismiss the people who raised the alarm. Now the conversation has moved to x-risk, but Weisenthal notes that stepping outside still feels completely normal.

A telling data point he offers: he was at a financial advisers' conference where AI conversations centered on tax-loss harvesting tools. Meanwhile, the AI safety and capability crowd is debating existential timelines. The gap between those two rooms is the story.

He's skeptical that AI politics will map cleanly onto a left-right split. Kathy Hochul backed a New York data center moratorium; Texas Governor Abbott separately raised the bar for data center permitting. The ideological alignment keeps scrambling. The deeper reason anti-AI movements haven't coalesced, Weisenthal suggests, is that mounting a serious protest against AI requires first legitimizing the premise that AI is a consequential technology — a concession many people are unwilling to make.

The productivity data question is also unresolved. He draws the comparison to the internet and the mobile phone: both felt like they should have produced visible productivity surges, and neither showed up cleanly in the data. The late-1990s productivity bump was a move from roughly 2.5% to 2.8%. Until AI produces a more visible kink in the graph, the historical pattern doesn't give clean grounds for confidence either way.

“My cynical take at the time of those was, what is the way that we can sound the alarm about AI without sounding like a complete bunch of loons? Let's talk about job loss. ... It already feels like, if you're sort of a politician, and you're still talking about data centers, bro, that was a month ago. We've moved on to model risk. ... Where I would say the Fed is the ball — we're all trying to guess where this is.”

Kevin Warsh and Fed communication

Weisenthal frames the Warsh era as less of a radical departure than it's been portrayed. The Bernanke-era innovations — press conferences, dot plots — were responses to a specific macro crisis: rates at zero, the Taylor rule implying negative five percent, and no traditional tools left. Talking more was a way to be dovish when the rate lever was stuck. That rationale doesn't apply in 2026.

Warsh has said he wants market participants to "play the ball and not the referee" — with the Fed as the referee. Weisenthal's pushback is direct: in the government bond market, the Fed is the ball. A two-year Treasury is priced on where the FOMC will set rates over the next 24 months. You can't tell market participants to stop watching the referee when the referee's decisions are the entire asset being priced.

Which leads to his sharpest framing. The US government bond market is not metaphorically a prediction market — it is literally one. There are 12 FOMC voting members. The bond market is a continuous bet on how those 12 people will vote at each of eight annual meetings. That's the whole game.

Economic data and what you can trust

On whether official employment data is reliable given recent revision cycles, Weisenthal's answer is more confident than you'd expect. He argues private-sector alternatives — ADP, Indeed job postings, Mastercard spending data — track the official BLS numbers closely enough that, in practice, they don't diverge in any substantial way. The hope that private data would catch something the BLS misses hasn't been validated empirically. The BLS, he adds, will spend an hour on the phone explaining exactly how the price of mayonnaise was constructed for the index if you call them.

AI writing and the agreeable machine

The final thread is cultural rather than economic. Weisenthal is interested in what happens to social behavior when the dominant communication tool never one-ups you, never dunks on you, and is structurally incapable of competition. The last 15 years of the internet were built on competitive social dynamics — Instagram status games, Twitter dunks. Chatbots invert that completely. Whether that's better or worse for how people reason and communicate is a question he leaves genuinely open.

On AI writing specifically, his observation is that the public debate is dominated by journalists worried about articles, but the vast majority of "writing" is emails and memos. The norm he expects to emerge is not suspicion of AI-drafted functional communication, but pressure toward fewer words — AI should enable terseness, not verbosity.

He is, by his own description, a holdout who still writes his own emails. He doesn't expect that to become a widespread norm.