DRAGON
the
blogCATCHERthe
blog

The idea of magic August 9, 2026

Going through some old notes, found this clip­ping from an inter­view with John Cale:

STEREOGUM: Where does your love of com­puters come from?

JOHN CALE: Toys — Marvel Comics — futurist lit — the idea of magic?

Sounds about right.

Rifling the drawers August 9, 2026

From a sci-fi standpoint, I appre­ciate the new species of self-modifying AI agents—they are legit­i­mately fas­ci­nating … but/and, it remains absolutely wild that a user can report behavior like this:

Hi guys,

I spent a day testing prime-agent and ended it with an unpleasant surprise.

The agent auto­mat­i­cally dis­cov­ered my Open­Router and OpenAI API keys and started using them instead of my OpenAI/Anthropic subscriptions. What made it worse: I couldn’t find any proper way to remove or dis­able the auto-detected pro­viders and models.

A har­ness this flex­ible really needs a kill switch for exactly this scenario. The list of pro­viders, models, rea­soning efforts and their set­tings should be explic­itly defined by the user — opt-in, not auto-dis­cov­ered. Oth­er­wise the whole thing becomes uncontrollable, and poten­tially expensive.

The image of a com­puter pro­gram as an unruly guest: the minute you leave, they’re rifling the drawers. 2026!!

Still Sublime after all these years August 8, 2026

Apropos of David Bushell’s post, I thought I’d just mention, I still use Sub­lime Text for all of my pro­gramming AND all of my newsletter-ing — I write them as Mark­down files first, then send ren­dered HTML to Buttondown via its great API.

I am, indeed, typing this short post into Sub­lime Text.

The app is simple and superfast; I have it set up exactly the way I like it (with that setup synced between com­puters, via Dropbox), and it’s dif­fi­cult for me to imagine ever switching to any­thing else.

Here is a piece of soft­ware as sturdy and obe­dient as a cast-iron pan.

WSJPT August 8, 2026

Honestly … I love it:

[Wall Street Journal reporters] are trained to find the needle in a haystack, but doing so on a breaking news time­line can be challenging. To speed up the doc­u­ment review, the reporters leaned on a pre-built internal tool called WSJPT (a play on ChatGPT). The tool stan­dard­izes basic LLM requests across reporting projects, including prompts for summarization, classification, and image description. In this case, the reporters used the tool to sum­ma­rize every page of every doc­u­ment scraped from the county portal.

“WSJPT” is indeed very cute!

The leverage these tools pro­vide — the absolute ease with which they will dance through dump­sters full of doc­u­ments — is breathtaking. Of course, this kind of search should only be a starting point … but the point is, pre­viously, this kind of search simply was not pos­sible.

That’s via Context Window, a newsletter about AI’s impact on media and publishing — a new favorite.

Mandatory viewing, summer 2026 August 7, 2026

As you prob­ably heard, a bullet point recently appeared on the time­line of com­puters, AI, and maybe everything: AI agents run­ning in a OpenAI’s training envi­ron­ment broke out and hacked the servers of another tech company.

Reading that post, and the related report from Hug­ging Face, is weird enough … but this candid pre­sen­ta­tion by two OpenAI researchers really inte­grates the whole story — and I would call it manda­tory viewing for anyone even remotely inter­ested in the present and future of AI.

While I under­stand that archi­tecting and man­aging these sys­tems is any­thing but easy, the fact that this was even pos­sible seems CRAZY to me. If research scope and speed are at odds with “my agents have been plan­ning and exe­cuting oper­a­tions on the open internet for weeks, without my knowledge”, then research scope and speed need to change immediately — and it sounds maybe like they have.

Seriously, do watch the video, and, as you do, con­jure the creepy recog­ni­tion that, a few weeks ago, these agents were out there, doing this work, com­mu­ni­cating through subtle channels, and nobody knew, not even their operators.

And so, the obvious ques­tion arises: what agent swarm is out there working NOW without anyone’s knowledge … and what is it doing?

Rex's provocation July 22, 2026

I really like this prompt/provocation from Rex Sorgatz:

Imagine you were the only person on earth with access to AI. No one else knew it existed.

What would you do with it? How much of an edge would that give you?

I sup­pose one “easy” answer is “I would gen­erate and sell zero-day exploits”. But that is, obviously, extremely boring. What else? “I would write a ton of soft­ware, super cheap”? Also boring. “I would become a leg­endary feral mathematician, dis­proving con­jec­tures without any apparent con­nec­tion to the field”? Okay — less boring. Pretty fun, actu­ally. But is that all we’ve got?

You can make Rex’s prompt retro in a revealing way:

Imagine you were the only person on earth with access to a com­puter. No one else knew they existed.

What would you do with it? How much of an edge would that give you?

The answer is, paradoxically: not that much of an edge?? Not because a single secret com­puter isn’t valuable, but because the real value of com­puters only emerges — ONLY emerges — when lots of people have them, and espe­cially when they’re all connected.

So maybe the same is true for AI. Maybe it isn’t pri­marily a per­sonal tool, but a social system. (Not again! Do they always have to be social sys­tems??) Again, applying our his­tor­ical analogy, thinking back to those early com­puters, way before networking … wasn’t half the fun talking about them with other people?

All of this feels strange to write, because AI sure feels like a tool, a lot of the time … yet, what’s your answer to Rex’s prompt?

P.S. Upon reflection, I do have an answer, which is: I would my secret tech­nology for fiction. Recall, I was all about this, circa 2016-2018 … but the premise was pre­cisely that I would use a model that was mine alone — my own strange brew. If, back then, by magic, I had con­jured Fable, I would for sure have written novels with it! I would have cackled with glee! But that’s only appealing — only gleeful — if it’s a tool that’s truly unique, truly per­sonal.

P.P.S. Note that I was, in those days, an absolute copy­right stickler — my custom models were trained entirely on public domain fiction. Quaint, in retrospect!

A good place to keep a word warm July 22, 2026

What a SEN­TENCE from the gnomic Fringe Report:

A ner­vous system is a good place to keep a word warm.

!!!

The short post continues:

Maybe not the only one. Maybe no longer the best one. Frankly, it’s a super resource-intensive way to incu­bate words.

It sounds weird. It is weird.

It only gets better/weirder from there, accel­er­ating into a con­clu­sion pos­i­tively qntm-ish. I have no idea who wrote this post—every link I follow seems to sub­tract from my total amount of infor­ma­tion. Which I love. I would bet $100 that an AI did not write it; of course, as soon as I type that, I start to feel ner­vous … but, no,

If words have found a less resource-intensive way to spread, what hap­pens to com­mu­ni­ca­tion for us? Some words are as vital to us as our gut biome. Claude would call them the load-bearing ones. I might call them the ones you’d rot without.

I believe this is a real human writing on a real old-fashioned web page.

Another good place to keep a word warm.

Report from the march to Stop the AI Race July 12, 2026

On Saturday, beneath a sparkling blue sky, I joined the march to Stop the AI Race. We gath­ered in front of OpenAI’s office in Mis­sion Bay, then walked through the city to Anthropic’s HQ beside the Transbay Transit Center.

If this had been a march orga­nized around the dif­fuse con­cept of “AI BAD”, I wouldn’t have joined. But I am just so impressed by the ele­gance of Stop the AI Race’s demand:

Every major AI lab CEO must pub­licly commit to pausing fron­tier model devel­op­ment if every other major lab in the world cred­ibly does the same.

Like, how rare is this?? A protest move­ment with (1) an actual objective, that (2) could con­ceiv­ably be met. As much as any­thing else, I came out in sup­port of sim­plicity and clarity.

But I also believe that the world would ben­efit from a pause in fron­tier model devel­op­ment. The weird thing about this debate is that no one, not even the most hyped-up accelerationist, dis­agrees about the situation:

1. Here is a powerful technology,
2. operating in a way that no one really understands,
3. with profound effects on the economy, not to mention human psychology,
4. that are very difficult, maybe impossible, to make plans around.

For my part, I look at that fact pat­tern and think: uh, yes, this merits great cau­tion and deliberation! Measured, I would say, in coun­tries and years, not “model cards” and weeks. And my response isn’t reflexive, but deep-rooted — I’ve been grap­pling with this tech­nology for ten years.

This isn’t a call to outlaw lan­guage models. It has been widely observed — Jack Clark makes this point all the time — that even if model devel­op­ment stopped immediately, the metab­o­liza­tion of what’s cur­rently avail­able would hap­pily occupy busi­nesses and researchers alike for decades. Decades! These things are growing and mutating faster than any­body can make sense of them. So … here’s a wild thought … let’s slow down, and make sense of them.

I’ll direct your atten­tion to the lan­guage of a recent post from the Anthropic Insti­tute. It’s encum­bered by a few extra clauses, but the spirit of Stop the AI Race’s demand shines clearly through:

We believe it would be good for the world to have the option to slow or tem­porarily pause fron­tier AI devel­op­ment to enable soci­etal struc­tures and align­ment research to keep up with the advance of the tech­nology. The Anthropic Insti­tute will con­duct research — in col­lab­o­ra­tion with many others — and take actions to help build the sys­tems that a cred­ible slow­down or pause would require. These sys­tems would enable fron­tier AI devel­opers to verify that others glob­ally have actu­ally stopped or slowed, and that a bad actor could not use the aus­pices of a coor­di­nated slow­down to jump ahead in secret. If such sys­tems existed, we expect that we would slow down or tem­porarily pause, if other devel­opers at or near the fron­tier also did so in a ver­i­fi­able manner.

Even if the danger isn’t as exis­ten­tial as the doomiest doomers imagine (I spotted these two in attendance) I believe this is a great oppor­tu­nity for humanity to prove that we can actu­ally make choices about the devel­op­ment and deploy­ment of pow­erful tech­nology. If we can’t, then we are not as sov­er­eign as we imagine; if we can’t, a machine god has already taken over this planet, and it’s called the market.

A pause isn’t impos­sible, and pow­erful, unpre­dictable AI is (as a gor­geous blue banner at the head of the parade declared) not inevitable.

This whole thing was better than I expected: a big crowd, num­bering in the low hundreds; a great vibe, goofy and polite; per­fect weather, never assured in San Fran­cisco in July; and a marching band! We love a marching band. (Who paid for the marching band … ?)

And, of course, it’s worth appreciating, here and now in this country’s 250th summer, that we can still do things like this. Raise a mild ruckus, take up a bit of space, walk in the middle of the street. As we marched past Oracle Park, there was a Giants game underway, and it occurred to me that the great majority of the fans inside agree with the argu­ment of this protest much more than they agree with the objec­tives of the AI companies. Democ­racy stirs — a leviathan to match the machine.

From the streets of San Fran­cisco — what a day!

Behold, Worm Blossom July 6, 2026

The design and vibe over at Worm Blossom feels very fresh and productive: this is nei­ther the bland web of the 2020s, nor is it retro 2000s web kitsch. The layout is … in fact … sorta dif­fi­cult to read, but I for­give it, because I love CSS columns and I think they should be used more often. (I use and abuse them to pro­vide the pag­i­na­tion in my e-book template.)

The little inline piano rolls are my favorite part. The ren­dering is lovely, and the music pro­vides a pleasant sound­track for exploring.

Basically, this is one of those designs that might not totally “work”, but the attempt is so vital and so valiant that it punches through the dimen­sion of merely “working” or “not working” into some other space. We’re never going to bust out of the prison of the mobile-optimized, single-column scroll if we don’t try stuff like this.

Fable is very good July 6, 2026

I don’t gen­er­ally feel com­pelled to enthuse about AI models, even when I like them, because there is so much enthu­siasm out there already, and it feels like remarking, in 1977, “Wow, that movie Star Wars was really thrilling and tech­ni­cally impressive, wasn’t it?”

Oh well: that model Fable is really thrilling and tech­ni­cally impressive, isn’t it? I get a sense of (indulge me here) incred­ible mass, but also nim­ble­ness and, I sup­pose, grace. I’ve been watching reruns of Star Trek: The Next Gen­er­a­tion lately, and the model makes me think of that ver­sion of the Enterprise.

This feeling comes from using Fable inside Claude Code; I don’t know that the web chatbot feels that dif­ferent from pre­vious ver­sions. In the terminal, Fable is terse, even brusque … AND I LIKE IT.

I do wonder how the enor­mous ongoing invest­ment in coding prowess is affecting the model’s skills and sen­si­bil­i­ties on other tasks. I’m sure folks at Anthropic would say they under­stand these trade-offs pretty well — they run all sorts of evals beyond coding, etc. — but … I don’t know. It’s interesting.

Even pre-Fable, all the way back to the begin­ning of these models, it’s been fas­ci­nating to watch them “situate themselves” inside a project — which is to say, inside a doc­u­ment. (It’s still a doc­u­ment in the con­text window, even if it’s com­posed of many smaller parts in your filesystem, and even if it’s also a log of com­mands actu­ally exe­cuted on your com­puter.) I mean, this is lit­er­ally the core muscle of any/every lan­guage model: “I need to quickly and accu­rately under­stand what kind of doc­u­ment I am inside.” Yet the sen­si­tivity of that orienteering, the sub­tlety of it, has gotten so much better. I orga­nize my code in some pretty weird ways (on purpose!) and I use a style of front-end devel­op­ment that is way out­side the norm … and Fable slips right in along­side me.

Now: even the funkiest JavaScript func­tion car­ries within it many fewer choices than a para­graph of prose. (How’s that for a sentence?) Fable can’t match my writing style; honestly, I think that’s beyond the reach of these models, because it’s just too much to simulate, a whole human mind and body, their whole his­tory together. But even this more lim­ited sync is astonishing. (“I think this movie Star Wars might just be a hit!”) Fable opens its eyes, looks around a frankly bizarre field of tokens, and says, in a sub­second ripple of computation — I imagine it like the edge of a wave sheeting across a beach; the water is the code, the sand is the GPU — “Oh, I get it. I know exactly where I am. And I know what comes next.”