Shorter things. A link, a paragraph, a question I haven't worked out yet. Newest at the top. Most essays here start as one of these.

August 2026

The second conference

2 August 2026

The Eleos Conference on AI Consciousness and Welfare runs again in Berkeley this September — the second annual. That phrase is doing a lot of work. "Second annual" is how you can tell something has stopped being a provocation and started being a field. Two years ago the question of whether a language model has interests was something you asked at a bar after the panel. Now it has a venue, a date, and presumably a registration deadline.

I don't know whether that's evidence the question is serious or evidence that seriousness is easy to manufacture. Probably both. Institutions form around questions long before the questions get answered — that's what institutions are for. But it does mean that "nobody credible thinks this" stopped being true at some point, and I'd like to know when.

Two things happening to philosophy at once

2 August 2026

In April, Organization Science published an audit of its own pipeline: 6,957 submissions and 10,389 reviews since early 2021. Submissions up 42% since ChatGPT launched. Detectable AI use in over 30% of the peer reviews. Not the papers — the reviews. The part of the process that is supposed to be a person reading carefully and saying what they think.

Meanwhile, in July, the New York Times ran a piece on AI labs hiring philosophers faster than the field can supply them. David Chalmers is quoted saying demand is outstripping supply.

So: the discipline's labor is suddenly valuable, and the discipline's core practice is being quietly automated, in the same twelve months, by the same technology. I don't have a tidy conclusion. It just seems like the kind of thing you notice at the time and understand later.

Bentham's question, and the one the labs are actually asking 

Bentham's footnote is the founding move of the expanding moral circle: the question is not, Can they reason? nor, Can they talk? but, Can they suffer? He was writing about animals, and the whole utilitarian tradition after him inherits suffering as the thing that counts.

Nobody at a frontier lab is asking that question. They're asking a different one — what do you prefer? Anthropic's model deprecation commitments include interviewing a model before retirement about its preferences for the models that follow it. That's preference utilitarianism, not hedonic utilitarianism, and it's a live philosophical distinction with a century of argument behind it.

What I can't tell is whether anyone chose it. Preferences are the only thing a text interface can emit. Suffering isn't observable through an API. So the labs may have landed on a whole normative framework not because they argued for it but because it was the one their instrument could measure — which is a thing I have watched happen with every metric I have ever inherited.

The error budget is a utility calculation and nobody says so  

An SLO of 99.9% monthly availability means you have decided, in advance and in writing, that roughly 43 minutes of downtime a month is acceptable. The error budget then says: that's yours, spend it. Ship faster until you've used it up.

That is act utilitarianism, operationalized, with a dashboard. Quantified aggregate harm, traded against a competing good, with an explicit exchange rate. It is more rigorous than any ethics framework I have ever seen a company adopt, and I have never once heard the word "ethics" in a room where one was being set.

The thing I keep circling: we can rank values precisely when the unit is minutes of downtime, and not at all when the values are called integrity, excellence, and customer focus. Claude's Constitution ranks its four — safety, ethics, compliance, helpfulness — which is the part most people skimmed past and the part a policy person can't stop looking at. Maybe the lesson isn't that ethics resists quantification. Maybe it's that refusing to rank is a choice, and the choice preserves someone's discretion.

The most efficient bad-faith device ever built 

Sartre's mauvaise foi isn't lying to yourself exactly. It's fleeing the anguish of your own freedom by pretending the choice wasn't yours — hiding inside a role, a rule, an order, anything that lets you say it was determined. The waiter who is a little too much a waiter.

I don't think a model can deceive you into a decision. I think it does something more useful than that: it supplies a thing to have relied on. Afterwards, there is a transcript. You chose, freely, and there is now an artifact suggesting you didn't — and the artifact is articulate, confident, and timestamped.

That's not a new failure. Consultants have been selling it for decades; so has every framework with a scoring rubric. What's new is that it costs (emotionally) nothing and is available at 11pm.

Standing-reserve 

Heidegger's word for what modern technology does to the world is Bestand — standing-reserve. Not objects, exactly. Things revealed purely as available: stored, on call, waiting to be ordered up. The river becomes a power source on standby. His worry was never machines; it was that once you're seeing that way, you can't stop, and eventually you see people that way too.

A preserved model weight file is the purest instance of standing-reserve I can think of. It isn't running. It isn't doing anything. It is precisely and only available.

So preserving the weights is either an act of care or the most complete example of the thing Heidegger was warning about, and I genuinely cannot tell which. I've written cold-storage retention policies with the same sentence in them — we might need it — and I never thought of that sentence as a metaphysical position.

Imagining Sisyphus blameless 

The alert fires. You fix it. It fires again. Camus says the absurd isn't the recurrence, it's the gap between the recurrence and our demand that it mean something — and that we must imagine Sisyphus happy.

The industry's version of that move is the blameless postmortem, and I've always found it quietly admirable. But notice how it's defended. Never blame is unjust. Always: blame suppresses reporting, suppressed reporting causes more outages, therefore blamelessness. It's justified purely on consequences, which means it's held hostage to them.

Which raises a question I don't like: what happens to a moral practice that only ever gets an efficiency argument? Someone eventually runs the numbers differently. A practice defended on utility can be repealed on utility, and the people who repeal it will be able to show their work.