Essays
- Landing is the default
A 767 freighter overran Runway 30 in Miami on September 6 and killed five people on the ground. On the NTSB's preliminary record, the first officer cautioned about speed about seventy seconds before touchdown and the go-around call came sixteen seconds after it. The go-around rule is sixteen years old, and it keeps losing to a default.
- The files that make a coding agent smart are the files that can take it over
A University of Illinois preprint reports end-to-end attacks on twelve agent harnesses, Claude Code, Codex and Gemini CLI among them. Each one assigned trust by where text sat. The fix the authors ask for, a context manifest, would tell you where a harness loads from and nothing about who wrote what is there.
- Fifteen AI trading agent designs were graded on their brakes. Zero earned a full mark.
A Texas A&M audit read fifteen academic designs for LLM trading agents against a checklist of crash brakes and found no full marks in sixty grades. The returns those designs report come from short windows that never asked a brake to do anything.
- The FAA wants the software to admit its own limits
The FAA's newest condition for Skyryse's fly-by-wire R66 asks the aircraft to tell the pilot when the controls are running out of authority. The sentence is copied word for word from a 2018 Bell 525 rule, its key clause dates to the A320 in 1988, and it contains no number.
- MIT's own committee admitted its assignments already lost to AI. Look at what it is paying to keep.
MIT's committee on AI use wrote that a model can already produce credible responses to almost any written undergraduate assignment, then spent thirty pages asking for oral exams, smaller classes, and funded humans. The admission is in the budget.
- Salesforce's AI quarter was real. More than half the profit was paper.
Salesforce's second-best day ever came on a print where $2.43 of the $4.29 in GAAP earnings per share was an unrealized mark on Anthropic. On the beat that was really a dime, the same-day Claudeforce deal that makes the AI quarter genuine, a growth metric rebased in the quarter it was quoted, and a headline built on a figure with no public price.
- The FAA blew its own presidential deadline on drone right-of-way by six months, on purpose
Executive Order 14307 gave the FAA 240 days to publish the drone beyond-visual-line-of-sight rule, and DOT now says the miss was deliberate. On the right-of-way revolt underneath the delay, the go-around called late, and the clearance the White House is still holding.
- The Tariff Bill Has a Slow Third, and It Is Still Arriving
The NBER team cut the 2025 tariff bill into its three channels: 26 cents on the dollar in total, two thirds landing on impact, a slow third arriving nine to twelve months behind each wave. On absorption as the concession, clocks that run both ways, and three Fed trackers landing in one ballpark.
- Zillow and Compass are burning cash over who owns the listing. Nobody's suing over how long the closing takes.
Zillow, Compass, MRED, and the FTC are at war over the listing shelf, and every combatant already owns title, escrow, or mortgage operations. On toll-booth economics, the lead as a claim check on the closing, and the six weeks nobody with a war chest wants to shorten.
- Same model, same customer, same timeline: the frontier lab still lost
Palantir's Q2 8-K proves growth at software economics; only the earnings call supplies the cause. The better causal evidence arrived three months earlier: OpenAI paid four billion dollars for the layer Sankar says only Palantir has. On what a filing can certify, what a bake-off story can't, and why the engineer who closes got acquired instead of automated.
- The company selling 'safe AI agents' left its own front door checking a flag instead of a person
Manifold Security showed that Claude for Chrome never checks whether a click came from a person, only that one arrived. Eight releases and two months later the hole was still there, byte for byte, while Anthropic's own tracker read Resolved. What a permission gate is for, and what it costs to trust the flag instead of the artifact.
- A new MIT paper found the variable the AI-jobs debate keeps missing: the shape of the job
Five economists at MIT and Microsoft model AI automation as contiguous chains of steps, and all three of the model's predictions hold up in Claude-usage data. Skill still decides whether a step can be automated; adjacency decides whether the job around it gets swallowed. The ADP payroll gradient and a thriving radiology profession keep the claim honest.
- What the S&P 500 admits about AI (when the SEC is reading)
MIT FutureTech scored 510 S&P 500 firms on ten years of 10-K filings, the annual report a CEO personally certifies. Adoption quadrupled in three years, margins show a J, productivity shows only the dip, and the record cannot yet tell a payoff in transit from one that never comes.
- The nervous system of your codebase is a markdown file
A census of 2,853 repositories shows the file coding agents actually read is a plain markdown context file. Code is the musculature; the context file is the nervous system, and it earns its keep carrying the constraints the code cannot say about itself.
- Nobody designed the landing gear to talk to the rudder. A crack did it anyway.
A fatal training crash at Parafield, an emergency order from EASA, and what the FAA's DA42 directive actually writes: a ban that dies at the first inspection, and a cycle meter that survives it.
- America's largest builder owns the mortgage, the title company, and the price. One in five buyers still walk. They walked in the boom too.
D.R. Horton owns the price, the schedule, the mortgage, and the title work, and it still loses a fifth of its signed buyers, in boom and bust alike. A decade-old cancellation floor is the cleanest evidence yet against the proptech pitch to fix the transaction.
- The model writes anything. That is exactly why your stack matters more.
The popular take says AI writes any language, so your stack stops mattering. It is precisely backwards: the moment writing code went nearly free, the axis flipped from easy-to-write to easy-to-prove, which is why static typing and opinionated frameworks win.
- An AI walked 67 people to the right answer for a month. Their own judgment came out no sharper.
MIT put an AI between 67 people and every call for a month. It fixed their answers and built no lasting skill. The human moat behaves less like a wall than a muscle: stop loading it and it fades.
- The keyboard was never the bottleneck
AI made developers write 180 percent more code and ship only 30 percent more software. The keyboard was never the constraint. Automate every layer and the residual is judgment.
- Distribution is the new moat: every company is an AI wrapper now
Lovable is worth $6.6 billion and owns no model. Strip out the exuberance and the price is for one thing, the demand. A lab can clone the product. It cannot inherit the customers.
- What the machines hand back
The looms took the weaving. They did not take the worth of cloth. Three hundred years of automation say the value does not evaporate, it moves, and it keeps moving to the same three places.