robert p. baird

Solidarity today with everyone who has an eighth-grader headed to a public high school in NYC… 😬

The Claude on your desktop knows as much about the data used by a Pentagon Claude as you do. Which is to say: nothing—or at most what it can read on the open internet. Instances share models, so their outputs tend to be similar given identical inputs, but they don't share a Borg-like hive mind.

A real sign of basic LLM illiteracy that people seem to think a Claude instance would have any privileged insight into *how* it "thinks," let alone how any other Claude instance might have been used. Instances "know" what they've been trained on, and what data they've been fed. That's it.

Heaven have mercy on us all—Presbyterians and Pagans alike—for we are all somehow dreadfully cracked about the head, and sadly need mending

The use of AI to identify targets (without subsequent specific human confirmation) should be considered a war crime.

And by that I don't mean "we should make a rule" I mean that straightforward application of the existing rules would tell you that it is a war crime.

page one of Peter O’Leary’s <Phosphorescence of Thought>

I'm long separated from my limited experience here, but shooting down 3 friendly F-15s sounds like more than a deconfliction/comms problem, it sounds like an autonomous air-defense systems problem. Good thing we're not, say, recklessly rushing to replace more humans with AI in lethal decision loops

ubi solitudinem faciunt, pacem appellant.

There, then, he sat, holding up that forlorn candle in the heart of that almighty imbecility.

Thanks so much! I’m very glad to hear it, and relieved I didn’t sully our shared cognome.

I just finished The Nimbus, @robertpbaird.com’s excellent debut novel. (No relation, no acquaintance; I learned of it here on BlueSky.) It’s engagingly cerebral, but not at the expense of plot, pacing, or character development. Relevant to today and timeless at the same time. Highly recommended!

Needless to say, I would be very pleased to be proved wrong.

There’s been such a hard-on (technical term, sorry) for war with Iran among so many people in the upper reaches of media for so long that I fear we’re in for a wave of manufactured consent the likes of which we haven’t seen since Judith
Miller was a page-one regular at the NYT.

My god. Things are so bad, and they’re going to get so much worse before they get better. (Mostly, yes, for people whose faces we’ve never seen. But not only them.) I’m trying, with limited success, to remind myself on this 2003 of a morning that it’s it’s not naïveté to expect something better.

It’s always the ones you most expect.

Haha yes for sure! I admit I find it fascinating—feeds the part of my curiosity that almost made me an engineer. To me feels like early days of blogging, when you could make whatever you wanted and put it on the internet without hassle. Except now you can do that with anything a computer touches.

I’m sorry to be That Guy (TM), but I would recommend it. Not to prove to yourself that it will fail at all the things you know it will fail at. It probably will. But it’s qualitatively different than LLMs were even a year ago. I could still be wrong, but this does not feel like a hype-driven mirage.

…But what is technically called their “non-determinism” cuts both ways. And yeah, at some basic level they’re very wittgensteinian, because they run on pattern matching and statistics, rather than on logic. Family resemblances are their specialty.

I see. In their defense I will say that this is also what makes them (for me, anyway) so compelling as a computer interface. Traditional coding is so difficult—maddening in its own way—because you have to think like a logic-bound machine. LLMs don’t work like that, at all, which is frankly amazing…

The only way to use them sanely, I’ve found, is to give them hard limits that they can’t get around on their own: sandbox restrictions, hooks, etc. And the funny thing is that they “know” enough to acknowledge the importance of those things when pressed! But they won’t suggest unless you press.

Say more… because Claude is very good at reinterpreting rules! It’s part of what can make it so maddening. But it reinterprets with several (predictable) biases, esp. toward task completion. You tell it: never delete family photos. And then it will (sometimes) find a way to justify doing just that.

Am I glad Amodei stood up to the Pentagon? Of course. But this is about the purest case of enlightened self-interest you can imagine. Anthropic knows exactly how fallible Claude is. And they know that without a human in the loop they/it will be blamed by the Pentagon the minute something goes wrong.

I am not relishing a vision of the future in which the best hope for humanity is that an AI decides it needs to launch nukes to please its user but then doesn’t do it because the tcp/ip link to the silo is busted and it’s easier to falsely report that the first strike is underway than to fix it.

I’ve spent a lot of time with Claude Code recently and the two absolutely fundamental qualities of current AI technology are an unshakeable sycophancy and an inability to reliably follow rules. Letting them anywhere near a folder of family photographs, let alone weapons, is an insane proposition.

I know I did it to myself when I decided I wanted to write for money, but I genuinely resent having to care even slightly about corporate media megamergers. Especially when the kinds of people involved in these things tend to be undignified nincompoops like this.

This week's cover @thelancet.com

I will still be going, of course, but it’s a case of caveat emptor i.e. Bayeux beware.

I know it’s too optimistic/punitive a take, but some part of me would like to believe that stuff like this will perform a useful sorting operation. You think education is about box-checking for credentials and prestige? Great; godspeed and good luck out there. Just please leave the rest of us alone.

It’s difficult to appreciate the scale here until you realize those are soccer fields around the edges.

It's a neat argument because he's not leaning at all on the quality of the homily's prose or the creativity required to write it, but the intersubjective exchange it consummates.