Reading Room

Outside your usual reading: this is a worked example of how tax design collides with wealth that can move, which touches your own thinking about concentrated stock and where to live.

Last week California certified a Billionaire Wealth Tax for the November ballot. It is a 5 percent one-time levy on the state's billionaires, paid out over five years. The goal is about $20 billion a year for health care, food aid, and schools after federal cuts. Greg Miller and Lars Doucet, co-founders of the Center for Land Economics, argue the tax targets the wrong pool of money, because land cannot leave the state.

Their first point is that the tax base is already shrinking. Six billionaires had already moved their tax residency out of state before the measure's January 1, 2026 cutoff: Larry Page, Sergey Brin, Peter Thiel, Don Hankey, Travis Kalanick, and Steven Spielberg. Together they are worth roughly $540 billion. Mark Zuckerberg, at about $220 billion, followed in early 2026 and will almost certainly fight the retroactive reach in court. Add a roughly $200 billion overestimate that other economists have flagged in the proposal's own model, and nearly half the assumed base is gone.

That creates a trap. To still raise $20 billion from what is left, the authors say the rate would have to climb from 1 percent toward 1.6 or 1.9 percent, depending on whether Zuckerberg prevails in court. A higher rate only gives the billionaires who stayed one more reason to follow their neighbors out.

Their alternative is land. Summing California's land values parcel by parcel, they estimate the total at about $8.14 trillion. That is roughly eight times the billionaire wealth the state can still realistically tax. Los Angeles County's land alone is worth more than the entire billionaire base the wealth tax is chasing.

The comparison they want you to remember is the rate. A land value tax of just 0.25 percent would raise the same $20 billion a year, on a base eight times larger that grows with California's economy instead of fleeing it. A rate of about 1 percent would cover California's entire $87 billion health and human services budget. A land value tax is an ordinary property tax with one change: it falls only on the value of the land, not the buildings on it.

The authors also point to how California assesses property. They estimate it now assesses property at somewhere between 44 and 60 percent of what it is actually worth.

Read this as an advocacy piece from a group that wants land taxes. The land estimate and the flight numbers are their own. The structural point still travels: a tax on a mobile base pushes the base away, and a tax on an immobile one does not.

On your desk today as a counterweight to the AI income hype you save, and as a test for any claim you make about leverage. Only an excerpt of the piece came through, so this covers its opening, not the full argument.

The essay is by Joe Procopio in Inc., dated July 10, 2026, titled "The AI Productivity Argument Is Over." Its subhead sets the frame: you can 10x or 100x or 1000x all you want, and that is not the productivity boost you are looking for. He says he is going to hammer the final nail in the coffin of the AI productivity argument.

His case starts with timing. He says the myth began to unravel over a year ago, when reports surfaced that once actual measurement of AI productivity was done, the results were far from spectacular. His main exhibit is the oft-quoted MIT study, which he quotes as saying just 5 percent of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L (profit and loss) impact.

He then speaks to readers who feel the opposite. You may have never been more productive using AI, and he says he believes that person and is rooting for them. But he argues that coding 30 apps in a weekend, or sending 1,000 personalized sales emails in 10 seconds, is not why everyone got fired. The excerpt cuts off in the middle of that sentence, so his explanation of what did drive the layoffs is not in hand.

The usable point is the gap he names. A person's felt speed and a company's measured profit are different things. A big personal multiplier on one task says little about whether the business ends up with more money. For your own leverage math, that suggests asking what shows up in revenue, not how fast the output arrived.

Outside your usual reading: this is a working designer's view on what comes after the desktop, which is useful as AI changes how people use software. The source text cuts off partway through the talk, so this covers only the first part.

Scott Jenson has worked on user interfaces and user experience for many years at Apple, Google, and other companies. He now works for Mastodon and Home Assistant, trying to squeeze more good UX design into open source. At Akademy 2026, KDE's annual developer conference, he tried to convince the room to take the lead on desktop design and move past the age-old "windows, icons, menus, pointer" model, called WIMP.

He told stories about UX details he worked on at Apple, including the Mac's Finder and the way dragging files between windows works on the Mac. The details of those stories are not in the text we have. He said he raised the dragging issue two years ago on Mastodon, and a KDE developer responded that it was a good idea to separate the mouse down and mouse up actions when dragging data between windows.

His main claim is that desktop UX stabilized. There were plenty of new ideas and improvements in the 1980s and 1990s, and then not much changed in 20 years. Much of the UX for Linux systems was originally copied from Windows and the Mac, and he said KDE even says it is for Windows users. He said there is nothing wrong with that, but felt the Linux community had been leaning on Microsoft and Apple to do all of the UX research and testing, and, in his words, make the mistakes. He also said Apple has let its desktop UX and desktop hardware languish for years because it shifted its focus to the iPhone and iPad.

LWN adds a caveat: GNOME has done some user research and KDE's guidelines came from contributor research, but these are exceptions, and no one is funding or conducting regular UX design research and development for the Linux desktop overall.

This is on your desk because you already run many agent sessions in parallel, and this is a look at how Anthropic's new Projects handles that. Only an excerpt of the newsletter came through.

Anthropic's redesigned Projects, in beta since September 17, lets one Claude Code conversation split a body of work into parallel threads that keep running after you close your laptop. Anthropic's pitch is that you brief a project the way you would brief a Chief of Staff. Each thread is a full Claude Code session on its own branch, and every thread shares the project's repositories, instructions and memory. You can check in and steer from your phone.

Reception is mixed, and the split is between people inside Anthropic and people outside it. Frank Lee, a member of Anthropic's technical staff, wrote that the new coordinator "unlocked a totally different mode of working" for him. He said it felt like "having a chief of staff dedicated to each thing you're working on." Dickson Tsai, who works on Claude Code, says it moved him off the command line and into Claude Desktop full time.

Outside Anthropic the tone is cooler. The Register headlined its launch story as a way to "work and pay in parallel." A Windows Forum analysis warned that parallel threads still produce merge conflicts and pull requests that do not fit together, leaving human review as the bottleneck. In Anthropic's own compiler experiment, 16 agents working in parallel hit the same bug and kept overwriting each other's fixes until the work was split differently.

Access is expanding to more Pro and Max accounts. The newsletter says the same model is coming to Claude chat and Cowork next, and that the habits you set in your first projects carry into every thread after them.

The full playbook sits behind the newsletter, and we only have part of its list of contents. It says it covers the firm model (how the coordinator, threads, memory, instructions, Library, Overview and routines map to the roles in a working firm), the economics on Pro and Max (what drains your plan, the auto-resume trap, and the settings to change before your first batch), 15 copy-ready templates from the engagement letter and kickoff brief to meeting synthesis, memory audits and closeout, and 8 engagements with thread plans. Those are client and market research, a build from a spec, a multi-repo migration, a goal-driven optimization loop, recurring ops, investor technical diligence, a founder's chief-of-staff inbox, and an overnight idea incubator.

We cannot vouch for anything beyond that list. We have not seen the setup steps, the templates or any plan-drain numbers. The compiler experiment is the warning worth keeping: parallel agents can collide on one bug when the work is not split cleanly.

Outside your usual reading: 351.tax tracks funds launched when investors contributed securities in kind (handing over shares, not cash) under Section 351 of the tax code. It counts 116 funds, and 53 of them cite a filing that names Section 351 outright. The rest are included because the economics match: shares contributed at launch and the original cost basis carried over. Seed size and cost basis come from fund filings, but the cost basis for funds without a filed report is an estimate. Your note ties it to a WSJ piece on tax strategy for concentrated stock winners.