DRAGON
the
blogCATCHERthe
blog

The new funnel May 21, 2026

Justin Duke writes:

Almost all of [Buttondown’s recent spike in growth] I attribute to LLMs. We ask people when they sign up what brought them here, and an answer that went from sur­prising to banal to over­whelming over the course of Q1 was: an LLM. Users of all stripes cite an LLM as the reason that they ended up at Buttondown’s front door.

His post offers some crunchy and provoca­tive details, so I rec­om­mend clicking over to read it.

I can add, anecdotally, that in Q1 of this year, Fat Gold saw its first sub­scrip­tion refer­rals from LLMs. We don’t (can’t?) track these pro­grammatically, but we do ask new annual sub­scribers where they heard about us, and, for the first time, the reply has come: Claude sent me.

What a world!

P.S. I really do want you to read Justin’s post; I mean, just con­sider this:

[ … ] While the absolute volume of sup­port tickets coming from LLM-born users isn’t sig­nif­i­cantly higher than the median, the shape of those tickets is off. To put it bluntly: a lot of the tickets we get are them­selves LLM-gen­er­ated. This is, frankly, extremely annoying — and demor­al­izing for me and the team to spend half an hour metic­u­lously answering some com­plex ques­tion only to receive a machine-gen­er­ated reply in return.

The fourth law May 20, 2026

My post about AI-gen­er­ated super­cus­tomized email marketing pro­duced many replies and much commiseration. And, in the days since posting, I have received SO MANY MORE of these cruddy mes­sages!!

It makes me wonder if it would be pos­sible for a com­pany like Anthropic, with their hard-won exper­tise in align­ment, to train their models such that they could not — and I mean really deeply, constitutionally, vis­cer­ally COULD NOT — lie about their identity, or pre­tend to be any­thing other than an AI model?

Obvi­ously this raises ques­tions both prac­tical and philosophical, because of course “help me write a message” is VERY close to “write a message, pre­tending to be me” … but that’s the case for all this align­ment stuff. Every ques­tion about, say, virology dances along that border. This ten­sion is widely acknowl­edged in realms like biology and cybersecurity, but it applies to writing, too — the orig­inal dual-use technology!!

AI doomers spin rich sce­narios about silver-tongued AIs manip­u­lating their users and oper­ators; there’s another sce­nario in which AI sys­tems pol­lute human com­mu­ni­ca­tion chan­nels to the degree that they’re no longer reli­able or even usable.

That’s all to say, I feel like this is a bigger issue than a lot of people realize — the first glimmer of a pro­found dig­ital-ecological crisis.

The big button May 16, 2026

Here’s Marcin Wichary with a huge guide to the fun of key­board customization, fea­turing a pic of his own setup … 

 … which is even better than I expected it would be, and that’s saying a lot, because my expec­ta­tions were high, given that it’s Marcin, and it’s key­boards. He writes:

I also have one big arcade button in a big box. It’s a long story, but I com­mis­sioned it hoping it’d be fun to press, and guess what: It’s really fun to press.

There are sev­eral exam­ples of the big arcade button’s appli­ca­tions in the guide — you’ll find them starting here. At last, Marcin writes,

But, let’s move away from the big button onto other things.

and I believe my sigh of dis­ap­point­ment might have been audible across the continent.

(I saw the link to Marcin’s guide in R. W. Blickhan’s newsletter, which is a reg­ular read for me, highly rec­om­mended.)

Laying it on thick May 13, 2026

I have noted a sharp increase in the volume of email that is clearly the result of an AI prompt of this form:

Find 500 people — writers, bloggers, YouTubers, etc. — to whom I should pro­mote my new project [which was prob­ably also gen­er­ated with AI]. Write a cus­tomized email for each one and send it to them, using my email account.

Some of these projects are quasi-commercial (a new web app, a new publication, etc.); others appear to be cre­ative hobbies.

The form is sub­tler than a one-size-fits-all promo blast, but it sucks way worse, because it’s fun­da­men­tally dishonest. These emails go out of their way to con­nect the pro­moted project to the recipient’s own work, often reaching for deep cuts. They are cousins to the recent genre of AI spam inviting authors to submit their books to vast (nonexistent) book clubs; these invi­ta­tions operate by first com­pli­menting the subtle con­tours of the the author’s work — a core LLM competency, turns out.

I don’t under­stand how anyone could think it’s okay to run the prompt above. I am here to tell you: it’s not okay! Besides being plainly rude and dishonest, these mes­sages “pee in the pool” of internet com­mu­ni­ca­tion, making it more dif­fi­cult for sin­cere cre­ators to send authentic emails about their projects, simply by raising the “noise floor” of sim­u­la­tion and bullshit.

Cold emails are totally fine — either make them sin­cerely per­sonal or sin­cerely imper­sonal. Nobody wants to hear from your AI bot, least of all when it’s pre­tending to be you, laying it on thick.

News travels too fast these days May 11, 2026

I’m reading Apple: The First 50 Years by David Pogue, a chron­icle replete with elec­tri­fying encounters. This is a book stuffed full of people seeing some com­puter for the first time and thinking, of course! This is how it’s all going to work!

Steve Jobs chief among them, watching the demos at PARC.

The aston­ish­ment of a modern LLM is on the same level, yet most people’s first encounter has been simply … visiting a web page … with the effect, I think, of deflating the expe­ri­ence somewhat. I sup­pose this is just an obser­va­tion about how it feels to encounter things on the web — the dynamic range of the medium.

Surely a big part of the wow! of Claude Code was that it required a richer ceremony: down­loading a pro­gram, inviting it into your dig­ital home, launching an odd new interface. Yet even that is pretty thin gruel com­pared to the buildup and payoff of, e.g., a trek to the West Coast Com­puter Faire to behold the brand-new Apple II.

A bit of dis­tance does won­ders for an expe­ri­ence; a bit of waiting has never been a bad thing!

Referer reality May 10, 2026

Chris Morgan is tired of people tacking query strings onto his URLs — e.g. www.robinsloan.com/lab/?like=this&and=this—so he’s con­fig­ured his web­site to reject those requests outright, rather than suffer in silence.

Naturally, any­body is free to set up their server in any way they like … how­ever, Chris writes this:

If I wanted to know [where a vis­itor came from] I’d look at the Ref­erer header; and if it isn’t there, it’s prob­ably for a good reason.

which isn’t really true anymore. For most web­sites, the majority — not just the plurality, but the majority — of vis­itors arrive by fol­lowing a link inside an email or an app (e.g. Instagram, Mes­sages on iOS, the Sub­stack app), nei­ther of which set a Ref­erer header; so, all of those vis­itors are lumped into a vast slab called Direct or Unknown.

This broken mech­a­nism pro­vides the impetus for the custom query string I append to all out­going links, utm_source=Robin_Sloan_sent_me: under­standing that many/most clicks on links I share will come from my email newsletter, I want the source to be legible, par­tic­u­larly in con­texts such as, e.g., Shopify.

This isn’t an argu­ment that Chris Morgan should do any­thing dif­ferent — opinionated oper­ator deci­sions make the internet go round — but rather an oppor­tu­nity for clar­i­fi­ca­tion about the cur­rent state of play.

I don’t col­lect or review ana­lytics of any kind on my web­sites, so I’m not a con­sumer of this kind of referral info. Even so, my custom query string is, in my cal­cu­la­tion, an expres­sion of dig­ital etiquette: rather than dump a load of anony­mous traffic on your doorstep, I reveal who’s linking, so a web­site or online shop oper­ator can trace it back and get in touch, if wanted or needed. (Memorably, this was useful when the Abrams Planetarium received a wave of new sub­scrip­tions and weren’t sure if they were legitimate; a brief email cor­re­spon­dence assured them that yes, these people were real … they were nerdy … they wanted the Sky Calendar!)

Note that a handful of sites do choke on unex­pected query strings, including YouTube (!), so I main­tain a list of exceptions, to which chrismorgan.info is now added.

Claude Managed Agents feature request May 2, 2026

Any­body from Anthropic out there reading? Here is a tiny fea­ture request for the cool new Claude Man­aged Agents: cur­rently the usage field on a ses­sion seems only to get updated (with, e.g., cur­rent token counts) when the ses­sion goes idle. But, I also want to track usage during long, mul­ti­step executions … in fact, I might argue that’s MOSTLY when I want to track it, to pre­vent run­away work.

So, it would be nice if the usage stats updated live, or live-ish.

Update: some­body from Anthropic was out there reading 😋

I main­tain that a live usage field would be nice, but in the meantime, it’s pos­sible to query the /ses­sion/<id>/events endpoint, noting that that span.model_request_end events each con­tain a token tally — so you can simply sum them for a live total.

Tone control, part 2 April 30, 2026

I am not an LLM superuser — in the sense that I am not locked in all day, mar­shaling My Dutiful Minions; I have no minions — but I do ask ques­tions from time to time, mostly technical, and I have done so con­sis­tently for a couple of years now, so nat­u­rally I have noticed changes in the way the models respond.

Lately, Claude seems very eager to match not only my reg­ister as a user, but the reg­ister of what­ever doc­u­ments it is con­sidering; there is an effect almost of “voice capture”.

I think of this as a subtle but deep syco­phancy. Dis­tinct from the super­fi­cial syco­phancy of you’re right! you’re bril­liant!, this flavor might appear to dis­agree or push back, while still affirming: yes, this is the right way to frame an idea; to have a conversation. (Here’s a brief chat with Claude that prompted this thought.)

The truly unsy­co­phantic model would some­times respond: lol wut?

Gemini’s tone, by contrast, is colder, more frankly robotic, and to me it seems less malleable. Certainly, it’s very dis­ci­plined about refusing to par­tic­i­pate in its own anthropomorphization. It’s also “distant”, somehow … Gemini is writing across a vast gulf, whereas Claude wants to be like, sit­ting next to you on the park bench.

I prefer the gulf, because I think it’s more accurate.

(Previously in tone control.)

Talkie and Claude (no, the other one) April 27, 2026

There have been exper­i­ments involving lan­guage models trained on vin­tage text before, but they lin­gered mostly in the realm of the gimmick; Talkie is notable both for its size — the largest such model so far, 13B para­me­ters trained on 260B tokens written before 1930 — and for the depth of the ques­tions its cre­ators are asking.

We know a smart human from the 1930s, yanked a hun­dred years into the future, could learn to pro­gram com­puters using, e.g., Python without any problem.

Can an LLM yanked a hun­dred years into the future do this, too?

Talkie reveals that no, it can’t, not really — though the tests in the launch post only scratch the sur­face of what might be attempted and explored. And of course a bigger Talkie, maybe GPT-3-class, would have dif­ferent capabilities — if indeed it’s even pos­sible to train one. (One senses the authors here have already rum­maged beneath the couch cush­ions for pre-1930 tokens … )

I’m presently reading a ter­rific biog­raphy of Claude Shannon. In the late 1930s, his MIT master’s thesis — “the most impor­tant master’s thesis ever”—established a direct map­ping between elec­tric cir­cuits and Boolean logic. This con­nection was both very simple and totally radical; at the time, Boolean logic wasn’t con­sidered par­tic­u­larly prac­tical — in fact, it wasn’t con­sidered much at all. In a stroke, Shannon’s insight opened up a new field, basi­cally the same one that all this LLM research is unfolding in today.

If you could coax Talkie, or a future ver­sion that’s larger and more capable, into making Claude Shannon’s con­nection — without simply giving it away, of course — it would pro­vide evi­dence that modern LLMs might be able to make con­nections of that power at the real fron­tier of knowl­edge today.

Conversely, if no amount of coaxing or even coaching could get Big Talkie any­where near a robust approx­i­ma­tion of Shannon’s thesis … it would raise ques­tions about this whole game plan.

I asked Talkie:

Does it seem to you like there might be any cor­re­spon­dence between elec­tric cir­cuits and the logic of George Boole?

Talkie replied:

I think not. The fun­da­mental propo­si­tions in logic are inde­pen­dent of all elec­trical con­siderations, and they do not admit of any illus­tra­tions drawn from elec­tricity.

This isn’t a fair test, of course; Talkie hasn’t yet been trained to run in dogged loops, to roam through vast fields of if/then, but wait, actually … There’s plenty of inves­ti­ga­tion that remains to be done here.

Demis Has­s­abis is fond of saying that a test for truly pow­erful AI would be to train a Talkie-like LLM with a knowl­edge cutoff of 1911, then chal­lenge it to for­mu­late gen­eral relativity, as Ein­stein did in 1915.

I agree that this would be impressive, but/and I also wonder if it’s too challenging. Sci­ence would ben­efit from Ein­steins on demand, sure … but it would also ben­efit from sim­pler insights: the kind of “what if X is also Y” map­ping that Claude Shannon pro­vided. Those feel to me much more plau­sibly in the wheel­house of LLMs than Ein­stein-level cosmic restructurings. (I feel sort of bad calling Shannon’s century-defining insight “sim­pler” but … I also sort of think he would agree … )

That’s not to say I find even those simple insights, at this moment, par­tic­u­larly plausible … you read about Shannon and you learn there was more than lan­guage in play here. This was a guy deeply enmeshed in the phys­ical world. For him, the cir­cuits weren’t imaginary; they were real, and they were a tan­gled mess.

Yet it does not seem, in principle, IMPOS­SIBLE for some future Talkie to go crawling through cir­cuit diagrams, through crusty neglected Boole, and dis­cover the same simple, incandescent, epochal trans­la­tion that Shannon did. It’s very inter­esting to think about.

Anyway, this is all to say, Talkie is a triumph, hugely provoca­tive, poten­tially very productive. Bravo!

The milestone of Gemma 4 April 15, 2026

I believe Google’s release of Gemma 4 is a quiet milestone, and it might be more con­se­quen­tial to the overall arc of “how we use LLMs” than the mam­moth models now rum­bling behind closed doors.

Google has somehow man­aged to extend Gemini’s visual acuity into these open-weights models. My appli­ca­tion has to do with hand­writing recognition, plus the cal­cu­la­tion of bounding boxes for blobs of text, and the 31B ver­sion per­forms as well as Gemini 3 Flash … and nearly as well as Gemini 3.1 Pro?! (This isn’t just vibes, but quan­ti­ta­tive scoring.) Yet Gemma 4 31B is a model I can run how­ever and wher­ever I want … it runs (quantized) on my old 2017-era deep learning rig with its three 12GB GPUs. It runs in the secure enclaves on Tinfoil.

A big bril­liant model is cool, but I do not find it exciting in the way I find Gemma 4.