On the other hand, actually applying a 'stable marriage algorithm' to people makes a bunch of assumptions which are worth interrogating:

- Do people actually know what they want even in the short term?

- Do people's preference change meaningfully over longer periods of time?

- Does your preference _upon reading people's profiles_ have any strong relationship to your preference after spending any amount of time with them?

The property that right after being matched, no one could have matched with someone whose profile they preferred more does not mean that this will create stable relationships.

Today the models are better, but they don't seem trustworthy or reliable enough that I want to give them them a lot of access or freedom. My coding agents sometimes still go off in the wrong direction, or say they did something other than what they did, or say they will do something and then immediately stop without doing anything.

The always-running (so almost never supervised) agent that is meant to do the same work as a person (and therefore needs _access_ like a person) seems like a notorious footgun from earlier this year was just made more powerful, and the companies that are supposed to know the most are telling you to connect it to everything.

If you think AI can help design new drugs, great! Let's pick some restricted DSL for describing molecules, preparation instructions, hook it up to the right set of lab robots that can pipette and centrifuge and whatever, _restrict its output to that DSL_, and let it iterate. The 'let an AI design drugs' process doesn't require that the agent can also, say, hack a niche German wiki.

For every area where we think these AIs can be _so valuable_ because they'll find solutions that humans wouldn't (or find them faster), then doesn't that value also imply it would be worth it for some humans to do some advanced setup of their specific domain-specific tools so it can be run in a sandbox equipped with only what it needs and not more?

But we have a one-time billionaire's tax proposal on the upcoming ballot, and we don't have an LVT proposal on the ballot. Saying that the Billionaire's tax will be less effective than promoters say maybe true -- but we're definitely going to get exactly $0 from a statewide LVT for 2026. If the Center for Land Economics gets an LVT on the ballot in a future year, I would strongly consider voting for it -- but that's not on its own a reason to not do the Billionaire Tax this year.

What _would_ be a convincing piece of info, and which no one knows, is what the long-term impacts of a one-time wealth tax are. E.g. Zuckerberg is moving his residence to FL but Meta isn't going to stop employing Californians. And if the state is _credible_ in saying it's a one-time tax, will the billionaires who fled come back after it's done?

In what sense can't it be passed to renters? Esp if all landlords in the market were faced with a new land tax that they had not previously planned for, why would it not be passed on?

- People who want to go up market can transfer to the KLM at Embarcadero and if they're going to Castro (or maybe Church?) they may also get there faster.

- If tourists want to see an old-timey street car, they could ride one perhaps half the distance, only along the water with views Coit tower, and the eastern end of the line would be right by the Railway Museum anyway. And we could run perhaps half as many cars.

At 2024 numbers in this SPUR report:

Cable car 872 streetcare 631 lightrail 412 system avg 327 motorbus 302 trolleybus 281

So canceling each cable car could IIUC, allow the system to add roughly 3 electric buses.

https://www.spur.org/sites/default/files/2026-03/SPUR_Taking...

> The department also failed to provide resources to its officers, including bulletproof vests and an evidence room.

Of course this will be used to bring back a PD with a bigger budget and more weapons. They may use some of the expanded funds to buy body cams but they won't work.

The "trunk learning rate" is set at 0.1x the learning rate for the experts, so learning on different subjects disproportionately happens in the experts, and the trunk portion is comparatively more stable. But the population of experts can grow and shrink:

> The pool grows when it is short of capacity and shrinks when parts of it stop being asked for.

So:

- doesn't the trunk then _eventually_ still undergo catastrophic forgetting, it just may take much longer?

- and before that point, catastrophic forgetting happens in stepwise chunks whenever the expert pool shrinks?

Yes, I think it's a mistake that many organizations are cramming LLMs inside of automated pipelines where the extreme generality/flexibility of the model is at odds with the fact that you're using it for a very specific task that gets repeated over and over, and needs a very specific structured output to be successful. But specifying your task carefully (as well as deciding what counts as your input state representation etc) seems like a form of programming. Something (a person or a model working in a relatively unrestricted way) will need to produce a configuration/specification for this system.

So rather than Jev vs Claude I imagine that using Claude/ChatGPT/whatever interactively to define / refine your Jev config which then runs in prod might be the happy combination?