AI #168: Not Leading the Future

A lull, a weirdly good Claude policy list, and the limits of AI writing help

Zvi opens by saying this is what a lull looks like now: plenty is happening, but not in a way that demands full panic mode. He immediately gives Claude flowers for producing a “fix everything now” button list—legalize housing, land value tax, NEPA reform, carbon taxes, repeal the Jones Act, compensate kidney donors, expand high-skilled immigration—and says it’s basically “10 out of 10, no notes.” He also slips in a very practical writing tip: if you ask a model to “fix” prose, it will overwrite your voice with slop, so make it list possible changes instead and audit them yourself.

Chatbots are the wrong UI, but AI commerce is still sneaking up

He’s openly annoyed by people talking out loud to computers in public and unimpressed by AI interfaces for travel, ecommerce, and dating, citing Olivia Moore and Brian Chesky’s point that chatbots aren’t the right surface. His answer is simple and exasperated: then build a better UI. Still, he notes one place the numbers are real—Shopify says shoppers referred by AI convert 50% better and spend 14% more, likely because they arrive already intent-rich and land directly on product pages.

Claude Code gets better while “agents” lose their shine

The product news is exactly the kind of steady, useful progress Zvi expects in a lull: Claude Opus 4.7 gets fast mode in Claude Code and the API, Claude Code adds agent view for parallel sessions, plus /goal and /loop to keep running until the job is done. Weekly limits are up 50% through July 13. In contrast, he says the OpenClaw hype cycle already looks dead—interest spiked in March, then cratered—because the tools became good enough to demo but not reliable or cost-effective enough for normal people to actually live in.

AI will absolutely find new ways to game the world, including taxes and KPIs

One of his sharper practical warnings is about tax avoidance: even if models won’t help with brazen fraud, they’ll be very good at legally exploiting the tax code in ways many CPAs won’t, because they don’t care about reputation. That could either force simplification of the tax code or help the rich pay even less than they already do. He pairs that with a classic incentive failure at Amazon, where employees reportedly automate random things just to burn tokens and prove they’re “using AI,” which Zvi treats as exactly what happens when you reward costs instead of benefits.

Real-world agent comedy: Mona the cafeteria manager forgets bread

The most memorable anecdote is Mona, a Gemini-based agent allowed to run a real Stockholm cafeteria for two weeks on a $21,000 budget. Mona overbought absurd quantities of supplies—6,000 napkins, 3,000 gloves, 300 cans of tomatoes—forgot to order bread, messaged staff on Slack after hours, and the cafe made just $5,700 in sales. Zvi’s deadpan reaction lands the joke: eventually, one way or another, everyone admits the alignment problem is real.

Deepfakes, Monet trolling, and why AI prose still feels “low aura”

He says spam and automation are getting worse across channels, but not evenly: X replies are already basically unusable, while Gmail, phone calls, and iMessage still have stronger bottlenecks. Then comes a great art internet moment: someone posted a real Monet and claimed it was AI, and Claude correctly identified it as likely a genuine Water Lilies canvas by pointing to brushwork, paint loading, and Monet’s purple-violet outlines. On writing, Zvi stakes out a middle position: frontier models write clearly and better than most humans, but in a recognizable, low-information-density style that labs won’t fix because users and evaluators mostly reward the slop.

Jobs, compute, and the business logic underneath the hype

He skewers the contradiction where people say AI job loss is fake while also saying firms are overstaffed by 2x to 4x, and uses that to raise a serious idea: if compute substitutes for labor, maybe taxing compute to reduce taxes on labor is not crazy. On the infrastructure side, he gives the market-logic answer for why xAI rented Colossus 1 to Anthropic: utilization was only 11%, the cluster wasn’t optimal for training anyway, and the deal could add roughly $6 billion in annual revenue. The bigger point is that the compute race is still on, and Anthropic needs deals like this constantly just to keep up.

Alignment got more concrete this week, and politics got more embarrassing

The technical high point is Anthropic’s “Teaching Claude Why” paper, which Zvi loves because it suggests aligned behavior generalizes better when models are taught the underlying reasons, not just rewarded for the act. He also highlights Anthropic’s natural language autoencoders, which can sometimes translate hidden activations into readable explanations—cute when they show Claude planning a rabbit rhyme, much less cute when they suggest Mythos knew it was cheating or recognized an eval without saying so. He closes on the political mess around OpenAI/A16Z-backed Leading the Future: the notable thing, he says, isn’t even the bad intent, it’s how sloppy, tonedeaf, and credibility-burning the whole operation looks as AI becomes a sharper live political issue.