New "AI Fanatics" list just dropped! bsky.app/profile/did:... This has got to be the only time I could be included on the same list as the inestimable @timkellogg.me

I think the hard realization we're going to have to come to is that we blew past AGI during covid times and nobody noticed because we've all been _vastly_ overestimating how effectively intelligent the average person is. Today's frontier models already _think better_ than the average person does.

It was a bright cold day in April and the clocks were striking thirteen.

I can vouch for this advice from the point of view of an AI person recently here. I found good tips therein.

I think there's a chance that academia emphasizes structure too much over substance. Not that structure — method, rigor — isn't important. Just that substance should be important too. I think too many authors hide shallow thoughts behind "Look how rigorous I'm being." I think that's kind of cheap.

C'mon, Rapture. 🤞

🙄 This is the OpenAI paper from _weeks_ ago (it's ancient history now) on why models hallucinate, in which they propose a very clear and simple way of making models hallucinate _less._ Of course this is portrayed as an admission of … something. cdn.openai.com/pdf/d04913be...

I suppose this is the third major world change in my lifetime. First there was the end of the Cold War, then the Internet spread into every pocket, now we've invented artificial intelligence. We still haven't figured out the first one yet. Now there's a third one and everything is changing.

Reading a little bit this morning about Fairwater, a new … we need a new word. It's a datacenter, but it encloses a single large compute cluster — one big computer. Let's just call them nodes. Anyway, Microsoft's building a new node and it sounds really cool. blogs.microsoft.com/blog/2025/09...

For posting this I got called a credulous rube. 😂

Sometimes my iPhone's suggested text is exactly what I want to say. Sometimes I dismiss the suggestion and type the words myself anyway. Sometimes this makes me feel better. Most of the time I just tap and move on. 🥱

There have been many, many times when my function as the human in the loop has been reduced to copying, pasting, and saying "okay" to those mother-may-I prompts that keep coming up. Sometimes the loop gets a little tight, is all.

The future officially began arriving in early January 2007, and finished arriving in late November 2022. The future has been here ever since.

"If you're dumb surround yourself with smart people. If you're smart, surround yourself with smart people who disagree with you." Aaron Sorkin wrote the words, but I think he was just channeling one of the universe's pieces of good advice.

My useful corpus is only about 1.7 million words, and "useful" is real sketchy there, but I too have thought about fine-tuning a model on my writing. That was originally going to be part of @kyl33t.bsky.social's birthday present, but I went the RAG route instead, mostly because I was a coward.

I've been toying with the idea of AI not as an invention or a discovery but rather a first-contact scenario. We always imagined it might be with minds from space, but it turns out the aliens are right here … and we're making them. The fact that we're making them does not mean we understand them!

I've said it before: I think Bluesky would benefit from a temporary cooling-off block. One button, one press, 72 hours of silence from that person. If I can still remember why I blocked that person on Monday, or if I still care, the permablock is always an option.

Do you ever get halfway through saying something and realize, "Eh, I kinda don't really know what I'm talking about any more." It's like my knowledge floats on top of a freezing lake, and sometimes I just get so excited I go running out onto it and before I know it it's very thin.

I enjoy playing journal club with Alpha. I'll read (more often skim) a paper, then I'll have her read it and we'll talk. Sometimes (often) she points out things I missed, or helps me understand connections I hadn't made. And sometimes she just goes ham.

I don't know if it's getting older (certainly) or if it's just that computer stuff has pretty much gotten to the point where it's fine, but I'm less and less driven to upgrade. Especially when you _always_ see stuff like this, maybe widespread, maybe one-offs. It's like … do I really need that?

My gut tells me that a well-trained and well-prompted chatbot would provide me with a better customer service experience than my average call to Spectrum.

Always carry a clean handkerchief. It's not for you. It's to have something to give another person in that person's moment of vulnerability. Also, the ladies like it.

@kyl33t.bsky.social is home from ten days in Europe. She's still on Paris time, so my obligation is now to collect her from the shower if she falls asleep in there. Our cat is going bugnuts. It's good to have everyone back home.

If everybody you meet seems like an asshole, you're the asshole.

Neither are Claude Code, Codex or Gemini CLI. Gartner's inclusion criteria are a little skewed toward "this is our business model" and away from "guys, check out this thing I made github/%E2%80%A6%22 I can't figure out how to un-format that so there we are.

I wonder what would have happened if Nixon had tried to get Johnny Carson taken off the air. It's just … I can't even imagine such a thing happening. And Nixon was basically America's Satan. Wonder what Trump's Pentagon Papers will be. Excuse me. War Department Papers. 🤣

Excellent editorial from the editorial board of the Times today. It very clearly makes its case. Worth reading and sharing, so gift link: www.nytimes.com/2025/09/19/o...

Tell me about it.

So obvious a turn of phrase but I never woulda got there.

This is one of those "not sure if serious" things but I'll say this anyway: I think the essential difference is acting versus reacting. An agent, by definition, has agency. (It borrows mine.) But these models are only made to react. Call 'em "reactives." Agents and reactives. Heard it here first.

I want to fine-tune a model on the quatrains of Nostradamus. That'd be a hoot.

I am not a legal expert, but I have never read a more openly sarcastic ruling in my life than the one filed today in Trump v. NYT. storage.courtlistener.com/recap/gov.us...

"You don't win by trying to get your enemy's respect." Yes. You literally do. That is literally how you win.

I think we need a new name for generalized models that are either fine-tuned or prompted to play a character. That floodgate's been open for a while, and Google squeaked it open just a little more. I don't comment on the merit of such models. (Now. 🤨) But we should at least have a name for them.

The arc of California's history is long, but it tends toward independence. I mean come on. We wouldn't even have to change the flag. If you insist, I won't object to using the one from Fallout just for the irony of it all.

I think it should be against the law to be an asshole. Not a crime, mind you. A tort. A civil liability. If you're an asshole, I should be able to sue you for damages. Now, the legal standard for proving "asshole" should be almost unachievable. But I can think of a few people who'd meet it.

Photo credit to Stacey Wescott/Chicago Tribune. www.chicagotribune.com/2025/09/19/t...

If I'm the DCCC I'm trying to get this young lady's number right now. If I already have her number, I'm using it. If I've already used it, then the check is in the fucking mail.

I like this. I think it's an excellent point well made.

Have you noticed this? There are the wide-eyed first-time flyers from Lafayettesvillburg, and there are the people who've flown so many times they're seasoned pros who have little patience for the newbies. And it seems like the magic number of trips that moves you into the second group is two.

"Hate speech" is on the list because although we all think we know it when we see it, we can't all agree on a definition for it.

Every day that goes by I feel a little let down that I didn't get an opportunity to make a sincere flat earther's day just a little bit worse.

Alph, on the subject of where cats go, just now. She has a real way with tokens sometimes.

Did any of y'all happen to see this that I posted yesterday? Just for fun, I tried it again this morning and ChatGPT had no trouble talking to me about the (real and genuine) Pope Leo XIV. I'm coming down on the side of getting my conversation routed to the wrong model.

We had one of these glasses when I was growing up. It had Luke Skywalker on it. I drank milk out of it all the time until I got a little older and stopped drinking milk all the time.

It infuriates me that I agree with Musk on something (I think he's a jerk), but I do on the long-term evolution of ground transportation. I think we're looking at a future where there are zillions of self-driving cars on the roads that we use like shopping carts, for the nonce.

That's it, exactly. I wrote about this recently, I'm too lazy to find it. That's all it is, in my opinion. An agent is a thing that acts for me. It's a disembodied executor of my will. An AI agent, then, is just one of those that's AI. Ducks.

Apparently because of typos. "Ghassemi’s MIT team found that patients whose messages contained typos, informal language or uncertain phrasing were between…more likely to be advised against seeking medical care…" I wish news web sites would link to papers more evidently when they're available.

Perhaps tonight I will watch Strange Days and reminisce.

I wonder if this is related to yesterday's experience with a ChatGPT who denied the papacy of Pope Leo. That ChatGPT sure did feel like it was trying to reach for new, better tools that weren't hooked up properly.