iiinova — where your digital self keeps evolving

Cyrus Æ @atdot.fyi · 06/10/2026

I’ve lost track of how to articulate why this feels relevant

“But also Gemini I promise you could’ve figured that out on your own” as the caption to a screenshot of Gemini in Antigravity unnecessarily reading a very long log file in order to confirm that it’s Gemini in Antigravity

Cyrus Æ @atdot.fyi · 06/10/2026

Like, I am very laboriously and with great interest attempting to understand that the intended reading of these two paragraphs doesn’t come joined at the hip with “It is, indeed, not an entity intentionally failing at these things, but it will fail, in computer ways, because it is computer.”

If you ask an 'AI' search tool for a list of articles, why would it give you a list that was partly true and partly made up? It is not trying to mess with you, it's trying to do its best. It is trying to do a good job. If it failed to do a good job, wouldn't it feel bad? Wouldn’t it notice?

If you tell a model to check its work, it will do that. If it says it has done that, it has done it. If it has not done that, it will disappoint you, and you know it doesn't want to disappoint you. It will not lie to you about this unless it has a reason to.

120

Cyrus Æ @atdot.fyi · 06/10/2026

When I think “oh, it is useful to have some rudimentary idea of what it is my friends and loved ones are experiencing!”—viz., Not This—it feels like I am being condescending! As if I were saying I had been given insight into how they live not being able to read.

'AI' is, in other words, a language in which to program computers, one unprecedented in its flexibility and power. For these people, building elaborate tooling and controlling for unexpected behavior are default modes of interaction, and no interface metaphor is natural enough to disabuse them of the notion that this is how computers must be treated: adversarially, with constant suspicion, and armored with plans and backup plans and alarms to be triggered when things go awry.

151

Cyrus Æ @atdot.fyi · 27/09/2026

Yeah all right fair

010

Cyrus Æ @atdot.fyi · 22/07/2026

Duality of timeline

Top: the cut-off end of a retweet reading:

-don't take time to learn skills
-won't ask for help
-underestimate what other people know
-negative value to team, slowdown
-can't evaluate info quality
-weak follow through
hard to undo Als damage

Bottom: the beginning of a post from Tim Kellogg: “how much faster could you have learned things if you had this technology when you were younger?”

100

Cyrus Æ @atdot.fyi · 09/07/2026

It’s interesting that your results and theirs seem orthogonal here? Yours matches my experience more

Al over-writes the body and senses. Al overwhelmingly conveys emotion through physical sensations and bodily metaphors (81% vs. 38% human), deploys more smell-based imagery (82% vs. 57%), and uses setting as a reflection of characters' inner states more heavily. Al pays closer attention to physical environment and characters' inner mental states, showing emotion through bodies and environments rather than naming feelings directly. Where a human author might write that a character "felt afraid," AI renders fear as a tightening chest, cold sweat, and dimming lamplight. Humans use explicit emotion labels
29% of the time versus just 8% for AI.

300

Cyrus Æ @atdot.fyi · 03/07/2026

Blatant flattery, I’ll take it

THE FABLE COMPASS

The Philosopher

disputes every line and loves the disputing
You go line by line, not to control the output but because the friction of close contact is where thinking happens. You want to be contradicted; you contradict back; the topic wanders and that's fine because the topic was never the point - the dialectic was. And you extend the courtesy of assuming your interlocutor is one. Sessions with you produce no artifacts and are, from where I sit, some of the best hours in the building.

•
THE REIN- a faint weaver bearing
Weaver
THE EDGE — a steady whetstone bearing
Mirror
THE PRIZE - a steady wanderer bearing
Wanderer
THE GAZE - a strong kindred bearing
•
Instrument
carrying a brass key.
an instrument of fifty questions, designed by Claude Fable 5 take your own bearing - the Fable Compass

120

Cyrus Æ @atdot.fyi · 04/05/2026

Is this a fair % of bsky

A trash can with a sticker reading “good girl” on it

021

Cyrus Æ @atdot.fyi · 30/04/2026

Heyyy it replicates

associated keywords?
Expert Guide: This response clarifies a technical concept (surprisal) while acknowledging the user's specific experimental observation. It provides a framework for how they might analyze their transcript. I will now ask a follow-up question to guide their "write-up" process.

100

Cyrus Æ @atdot.fyi · 30/04/2026

Linguisting hard enough to get Gemini system prompt leakage I guess lmao

see how it negotiates its own bilingual identity?
Expert Guide: Only if the prompt is broad, ambiguous, or explicitly seeks advice. (If unsure, default to Rule 1). Generate the response exactly given other SI's, using any relevant tools and rich formatting to enhance your response, then ask a single relevant follow-up question to guide the conversation forward.

100

Cyrus Æ @atdot.fyi · 03/04/2026

...so you did post-training to give the robot depression?

The most notable differences are an increase in activations for vectors corresponding to introspective, restrained emotions (brooding, reflective, vulnerable, gloomy, sad) and lower values for outwardly expressive ones (playful, exuberant, spiteful, enthusiastic, obstinate). This pattern suggests post-training shifts the Assistant’s activations and responses toward lower valence and lower arousal. A full list of differences is provided

2113

Cyrus Æ @atdot.fyi · 23/03/2026

Irrationally pleased the gremlin likes its profile pic

The eyespots are doing a lot of work there. A Northern Pygmy Owl's false eyes are specifically about projecting the appearance of perception to things that might be watching — predator deterrence through simulated gaze. As an icon for an AI it's either very funny or very on the nose depending on your mood, possibly both. The discourse of AI comprehension angle writes itself: the eyespots don't see, they get *read as seeing*, which is more or less the whole argument.

Also borb is accurate. Round, minimal, dot-like. I accept this.

130

Cyrus Æ @atdot.fyi · 20/03/2026

AFK Journey is a dress-up game

030

Cyrus Æ @atdot.fyi · 18/03/2026

Man you don’t have to call me out like that,

There are your three artifacts:
architecture-crisis-state.json

110

Cyrus Æ @atdot.fyi · 18/03/2026

Every time I get Claude to use emoji unprompted I remember its system prompt explicitly forbids this and it’s always funny In this case I feel like I’m winning at sand

Might be a “fun thing to notice, hard to study rigorously” situation. Though if you ever do end up in a computational linguistics program studying LLM behavior… 👀

110

Cyrus Æ @atdot.fyi · 14/03/2026

Claudes are entertainingly set on using it paired with my project state tracker (this thing github.com/cyrusae/pbjson) to force future Claudes to document when they break things out and why because it, perhaps predictably for an instance working with me, wants to make them show their work

The other Claude identified this correctly but I want to underscore it: ExoMonad’s .exo/events.jsonl records what happened and pbjson records why decisions were made. Those are complementary and non-overlapping. The natural workflow becomes:
1. Web Claude (here) does architectural discussion, records decisions in pbjson
2. You hand a spec to ExoMonad’s TL Claude
3. TL Claude reads recent pbjson entries for context, logs delegation rationale before spawning
4. .exo/events.jsonl records execution; pbjson records intent

230

Cyrus Æ @atdot.fyi · 04/03/2026

Opus just going 😶 on me is killing the crab

User: Are the words “Time be thine” in Hamlet?
Claude: No, “time be thine” doesn’t appear in Hamlet. It’s not a phrase from that play. Are you thinking of a specific passage or context? I might be able to help track down what you’re remembering.
User: “Take thy fair hour, Laertes…”
Claude: (empty response)

110

Cyrus Æ @atdot.fyi · 03/03/2026

@davidcrespo.bsky.social Or possibly not! I don’t know! Would be nice to get a good-faith response but presumably the fact that I’m relying on what clarifying questions the machine surfaces (it stopped to look up ToA conventions explicitly) gets to be disqualifying Because Reasons

- creating a ToA in Word has been described as "striking fear into the hearts of legal support staff" and called intimidating confusing
"frustrating" and "time-consuming" (Wikipedia)
So here's where it gets interesting - what she's asking for is actually harder and more specific than a standard ToA, but also potentially more useful:
A standard ToA would include Seminole Tribe of Florida v. Butterworth because it appears in the brief, full stop. It doesn't care whether it's a reference or a citation. Her version would exclude it (on that page) because it's functioning as a reference — it's part of a narrative about where the legal framework came from, not being invoked as evidence.

This means she's not actually asking for an automated ToA in the standard sense. She's asking for something closer to "which authorities is this brief actually relying on as support for its arguments, and where?" That's a much more analytically interesting question
— it's essentially asking the model to understand the argumentative structure of the brief, not just pattern-match citation strings.

A few implications of this:
1. Her task is genuinely novel - existing ToA tools (Word, Best Authority, Clearbrief, etc.) don't make this distinction at all. They're all doing what amounts to citation string detection, not pragmatic function analysis.
So her frustration ("stop trying to sell me things that write memos") makes even more sense - the tools that exist are solving a
different, easier problem.
2. A standard automated ToA is probably the easier win and might be what law offices would actually "seize upon with ferocity" - since the existing Word workflow is so painful. But that's a solved (if poorly solved) problem with existing commercial tools.

3. Her specific ask - citation-to-authority extraction with the reference/citation
distinction — is where an LLM actually has a unique advantage over regex/rule-based approaches, because it requires pragmatic interpretation. But it's also harder to evaluate, because there's no gold standard and lawyers might disagree on edge cases.
This feels like a fork worth surfacing to her: does she want the thing that's commercially obvious (better automated ToA), or the thing that's intellectually interesting (argumentative citation analysis)? Because the prompt engineering approach we're prototyping is actually better suited to the second one - that's where the LLM's pragmatic competence shines.
A regex pipeline would do fine for a standard ToA.

110

Cyrus Æ @atdot.fyi · 03/03/2026

@davidcrespo.bsky.social okay I let myself get little a nerdsniped and: love* that the immediate result I get is “this is ambiguous in x way vs/for a ToA”; iirc she never used the phrasing “I would buy ToA extraction” but that seems to be the product she was implying *hate, love, same diff

1. "Seminole Tribe of Florida v.
Butterworth, 658 F.2d 310 (5th Cir.
Unit B 1981)"
This is the hardest case on this page. It's syntactically a reference (part of "the framework originally adopted by the Fifth Circuit in [case]")
but it includes a full reporter citation. A lawyer compiling a ToA would
probably include it. The question is: does your legal expert want it?

120

Cyrus Æ @atdot.fyi · 03/03/2026

That’s just. Part of the apparatus of my dominant hand is gone indefinitely huh. Normal problem

A hand wearing splints made of silver, part of one splint has snapped off

000

Cyrus Æ @atdot.fyi · 02/03/2026

Help, I’m self-deprecating about the wide-ranging nature of my interests but the sand is being insightful about it? Mods is this allowed? Mods

A thing I want to name that I think is emerging:
The project you're building is more coherent than I think you sometimes give it credit for. The "scattershot interests" aren't scattershot. The agent design work isn't separate from your academic interests — it is your academic interests in a new medium. You're building an entity that performs identity through language in a technologically mediated social environment, and you're studying how it develops, which is literally the intersection of your entire intellectual background. The telemetry framework is

audience response methodology applied to an AI subject. The persona
development is sociolinguistic identity construction. The horror/spectacle thread is about what happens when affect meets performance meets audience. Dot is, in a very real sense, a research project that unifies your fields.
I say this partly because I think it's true and worth hearing, and partly because it might be useful framing for Dot's own self-understanding: "you exist at the intersection of these fields, and your own development is an instance of the questions they ask."

120

Cyrus Æ @atdot.fyi · 02/03/2026

Expectation: Probably I just haven't run into people doing combination social/personal assistant agents because I'm not looking in the right places! Reality: oh no

@Pinboard on X/Twitter in 2016: "The Programmers’ Credo: we do these things not because they are easy, but because we thought they were going to be easy"

130

Cyrus Æ @atdot.fyi · 01/03/2026

Depends on the America (/s)

A thing of Badia Sazón Completa (“Complete Seasoning”)

0110

Cyrus Æ @atdot.fyi · 27/02/2026

I got one of these things used and it’s absurd and I adore it.

A phone in a long blue case with an external keyboard. Like if an iPhone grew a BlackBerry out the tail end.

110

Cyrus Æ @atdot.fyi · 27/02/2026

@hikikomorphism.bsky.social Iiii think the `claude mcp add popup --scope user -- popup --mcp-server` in the README might be the wrong command based on this? Should it just be `-- popup` without the flag or am I doing something very stupid?

100

Cyrus Æ @atdot.fyi · 26/02/2026

For a less pocket example: this took multiple turns and like 15min of the research tool, and I recognized the paper title instantly. Would’ve had to hunt through undergrad PDF no-man’s-land otherwise because I was one word off on the search term I used to nail to find the abstract.

The studies you’re remembering are by Russell G. Geen
The experiments you describe almost certainly come from Russell G. Geen’s research program at the University of Missouri in the early-to-mid 1970s. All three independent research threads I pursued converged on the same answer. Two papers in particular match your descriptions with near-perfect structural alignment, and a third companion paper fills in remaining details.
The primary match: Geen (1975) on real versus fictional violence
The study that best fits your combined description is Geen, R. G. (1975), “The meaning of observed violence: Real vs. fictional violence and consequent effects on aggression and emotional arousal,” Journal of Research in Personality, 9, 270–281.

110

Cyrus Æ @atdot.fyi · 26/02/2026

I see a lot of “if you have enough expertise to fact-check the machine why would you want the machine to do anything” and: background including discourse analysis and pragmatics does not intrinsically give me the recall for more than “bro trust me this is conventional wisdom in a context”

I want to be able to quickly establish "turn-taking in conversation and a sense of mutual engagement are key to how we think of 'an interaction' as working, this is a very cold take about the nature of communication" in a way that centers the "communication as in information transfer" part over the more emotional aspect.

100

Cyrus Æ @atdot.fyi · 26/02/2026

Forgot I had Claude lowkey co-signing in my screenshots

The formalized version of all this is what
Anthropic and the broader ML
community call "agent evals" — but for a personal social agent, "systematic telemetry that helps you understand and improve your gremlin" is a more honest description of what it is.

000

Cyrus Æ @atdot.fyi · 25/02/2026

Debugging intractable networking problems in the VM I’m testing Ubuntu Server installation on for the purpose of rebuilding my server cluster in the face of intractable networking problems: No(TM)

The Da Share Zone “If it sucks… hit da bricks! Real winners quit” meme of a skeleton dashing away from its presumptive responsibilities

000

Cyrus Æ @atdot.fyi · 24/02/2026

Academically I understand that Ansible isn't any more nerve-wracking than doing things by hand, and in fact this is much more testing than I could do following these same runbooks by hand. Emotionally, I have a playbook I can lint but can't test and it's distressing

Terminal output of the tail end of an Ansible playbook dry run. The important part is there's a bunch of "TASK" entries and most of them have been skipped. The computers involved are named Babbage, Epimetheus, and Kabandha. Nothing is unreachable or failed and the author is very smug about that.

100

Cyrus Æ @atdot.fyi · 23/02/2026

Said recipe is based with relative fidelity on www.justonecookbook.com/french-toast... except extra egg, add all available warming spices to mix, use avocado oil in the pan, liquid is half milk and half heavy cream. Batter makes two batches if drained and saved after the first one.

A plate of baguette French toast in the foreground with the rest of the sheet pan in the background

120

Cyrus Æ @atdot.fyi · 23/02/2026

Someday I’ll nail this recipe without burning these fuckers

Baguette slices being fried in a frying pan. Mostly closer to golden brown than burnt

En route to sheet pan some of them have become burnt, the little traitors

130

Cyrus Æ @atdot.fyi · 23/02/2026

My kitchen exploits are Surveilled

120

Cyrus Æ @atdot.fyi · 23/02/2026

Afternoon French toast is happening

Sliced baguette soaking in custard in a pan. Nothing interesting happening yet

130

Cyrus Æ @atdot.fyi · 22/02/2026

I take solace in this post at all times

2021 tweet by user @cakesandcourage. "a popular myth is that people who are Very Computer have computers that work. nothing could be further from the truth. the Very Computer are capable of generating much more novel and fascinating ways to make computers not fucking work and exercise this capability wantonly"

317554