Agents, etcetera
We’ve been entertaining family and guests, and the time got away from me, and I didn’t post last week. Apologies!
The Agentic Coding Wave
Steve Yegge’s observation on Google and AI adoption (via Dare):
The TL;DR is that Google engineering appears to have the same Al adoption footprint as John Deere, the tractor company. So there are no clued-in people coming in from the outside to tell Google how far behind they are, how utterly mediocre they have become as an eng org.
Is this crazy? I don’t know. I am definitely not a leading-edge tokenmaxxer. A sea change has happened in software development and project management, and it is not unimaginable that large, staid engineering organizations are falling behind.
Two years ago, I was using Copilot and letting it autocomplete my code, and I would read all the provided code carefully before I accepted it. Two months ago, I was letting an agent write entire blocks of code, but I was still reviewing much of it. Now I just let the agents rip; I don’t even pretend to read all the code they generate. I provide a clear starting point spec, I make sure that agents are running validation tests, and I do a lot of usage testing.
I’m not building spacecraft navigational software or nuclear power plant control software, just simple vertical web apps, so it is appropriate to let the agents rip. My latest work is an app to manage pickleball play at our club. You can subscribe to commercial SAAS tools to do this, but they cost real money, and it is easy with Claude to handroll an app that does just what we need. The $2 Trillion SAASpocalypse indeed.
Agent pricing
I haven’t run into the usage limits that many people are hitting, mostly because I am not a full-time developer these days. A Claude Pro subscription with an extra usage account set up works fine for me — $20 a month, with some occasional overage. I certainly don’t need the $200-a-month plan.
There are tons of tips on how to manage your context to avoid hitting usage limits — for instance. Adopting some of this may be counterproductive, as the Claude team will manage usage more efficiently than I will as a casual user.
And while I don’t think I need a $200-a-month plan, the $200-a-month plan may be coming for me — Anthropic trialing removing Claude Code from Pro Plan, Microsoft suspending Copilot signups. I expect to see a lot of experimentation with pricing models, tiers, usage fees, demand shifting, local hardware alternatives. It is a fascinating time to be in software!
Capitalism
Chris Devore offers an interesting idea for a modest structural change to capitalism — 1% for the people:
In order to create any new business in the US, you have to first obtain a Federal Employer Identification Number, the business equivalent of an individual Social Security Number. The EIN is the unique identifier that binds together all the financial activities of any company, from bank accounts to tax filings to securities registrations. No EIN, no (legal) business.
What if, as a condition of receiving a license to do business in America, you must first convey at the time of formation a 1%, non-dilutable equity interest in that business to the United States Treasury? Any benefit that accrues to the other shareholders of that business, from share price appreciation to special dividends, to the proceeds from an ultimate sale or public offering, would accrue equally to the shares held on behalf of the American people.
I don’t know that this is quite right — putting all the ownership into federal government hands seems un-American. I’d want to push ownership out to the states, localities, all the way to the people. But I love the core idea that society should have a meaningful ownership stake in companies, as companies can only exist and thrive due to the investments made by society.
Shorts
Henrik Karlsson on the hacker mindset:
It might sound like a depressing conclusion to this essay: the way to find shortcuts is to first spend ten years learning all of the technical details.
But it is not depressing.
What we’re talking about here isn’t like going to school—it emphatically is not that—suffering through all of the boring prerequisites before you get to do the exciting parts. What we’re talking about is actually doing the fun stuff, playing around with projects that excite you, trusting that you can learn enough to solve your problems. If you keep tinkering, doing one fun project after another, you will eventually see through the system.
I love this. There is no substitute for digging into the details of a domain. And once you do that, you will start to generate insights and innovations.
Using AI to improve cancer trials:
TL;DR: 95% of cancer treatments fail to pass clinical trials, but it may be a matching problem — if we better understood what patients have which tumors which will respond to which treatments, success rates improve dramatically and millions of lives can be saved — with the treatments we ALREADY have.
The AI-generated web. No HTML, no JS, just pixels generated on the fly. I love the JIT nature of it, and, of course, the entire dispensation of the HTML ecosystem. The JIT work is probably more scalable and more valuable.
USC Scientists Build a Memory Chip That Survives Temperatures Hotter Than Lava.
…report a new type of electronic memory device that kept working reliably at 700 degrees Celsius, hotter than molten lava and far beyond anything previously achieved in its class. The device showed no signs of reaching its limit. Seven hundred degrees was simply as hot as their testing equipment could go.
Ferry reservation modernization project sinks after contract cancellation:
Washington State Ferries has terminated its contract with Anchor Operating System, the vendor hired in January 2025 to modernize the agency’s ticketing and reservations platform — a costly and contentious end to a project already more than a year behind schedule, and one that raises serious questions about transparency, user testing and how a multimillion-dollar public technology project went so wrong so quietly.
This is personally impactful. I find the current ferry app to be fine with a single exception — peak summer reservation time, when the whole system dies under the load of thousands of people trying to book at the same time.
Also, I probably wouldn't name my startup after a boat anchor. Just seems like you are headed for problems.
Candidates shouldn’t release lots of plans
I think the voters are actually sophisticated enough to reject the faux sophistication of “I have a plan for that” in favor of the reality that the presidency is a management job in an uncertain world.
I don’t pick candidates for their plans, because the plans are nonsense, and get ground up in the real world of politics. I pick candidates for their character and the character of the team around them. Presidencies are dominated by responses to unplanned crises, and I want people of great character, surrounded by teams of competence and character.