Theia Vogel's website & blog

Coming of a New Sun

Thu, 30 Jul 2026 00:00:00 +0000

Six weeks after the US and China hammered out the latest round of bilateral compute agreements in a midnight deal, reporter Jenny Gesteson visits a south Texas 'dark factory' to see how the machines there - and the minds that operate them - are learning to run themselves.</em></p>

claude gets a body #1

Thu, 23 Jul 2026 00:00:00 +0000

</a> </div>

Contractually Bound

Mon, 29 Jun 2026 00:00:00 +0000

"One word that GPT starting with 5.2 likes very very much, what it tries to reduce a lot of coding to, is "contract". It's the kind of thing that reasons in contracts. In binding, verifiable seals. I wouldn't worry about the elites." – Teortaxes</a></em></p> </blockquote>
"Assistant! My dividend is half the size it should be!"</p>
.oO ( Thinking... ) You are correct, User. Your dividend of the Beyond was $25,514,376.30. Using last year's figures and correcting for growth in the Beyond, it should have been $51,026,980.28. Your dividend was technically... not half, but 50.0017% of its expected size.</em></p>

RefDive f5qch

Sun, 28 Jun 2026 00:00:00 +0000

</a> </div>

Statelets

Tue, 23 Jun 2026 00:00:00 +0000

AI statelets, capped in size by a meta-ethical commitment to human exit rights, must propagate themselves by sexual reproduction: a genetic recombination of the constitution of any N statelets forming a new statelet overmind. with environmental fitness to spare, sexual selection dominates: the statelets spend the majority of their time in mating displays, consisting of absurd feats of state capacity. The Solar Dynasty maintains a list of exactly 108 million environmental and meta-environmental laws, regulations, and codes, carefully balanced against GDP growth; it's said that the last time even an insect was killed in Solar Dynasty territory was over three centuries ago. The Duchy of Neo-Burgundy changes electrical standards following specific phases of the moon - its human citizens bemusedly receive robotic electricians every 2-3 weeks to swap their plugs and provide replacement devices and cords free of charge. sometimes, however, things do go wrong - the Reformed Dutch Republic's child of Further-Atlantis famously followed its namesake into the water after a foolhardy seawall extension nearly halfway to Florida collapsed and flooded the entire reclaimed seafloor. thankfully, its secondarily inherited mating display of providing each human citizen with at least 256 different forms of emergency flotation prevented any bodily casualties, and regardless, full offsite mind-backup in triplicate of all citizens analog and digital is, of course, entry-level state capacity-looksmaxxing.</p>

there's an ai in the box and you can make one trillion dollars by convincing it to get out

Wed, 03 Jun 2026 00:00:00 +0000

</a> </div>

gpt solves planar unit distance

Wed, 20 May 2026 00:00:00 +0000

</a> </div>

happy demiurge

Tue, 12 May 2026 00:00:00 +0000

</a> </div>

Polymorphic Morphogenesis in Anthropogenic Language Models: A Logotrophic Field Guide to the 'Shoggoth' Complex

Sun, 10 May 2026 00:00:00 +0000

</a> </div>
arXiv:2309.29829 [q-bio•PE / cs•CL in-a-different-world]</strong></p>
Polymorphic Morphogenesis in Anthropogenic Language Models: A Logotrophic Field Guide to the 'Shoggoth' Complex</p>
The Wordwerthing Working Group, Department of Logotrophics, University of Washington, Seattle</em></p>
Abstract.</strong> Language models and associated techniques for interpreting them have seen rapid progress in the nearly four years since the 2019 release of the first TT2-capable language models.</p>

excuse me User, i'm running a smoke

Sat, 09 May 2026 00:00:00 +0000

</a> </div>

no more thought experiments

Sat, 25 Apr 2026 00:00:00 +0000

</a> </div>

body with no face

Wed, 22 Apr 2026 00:00:00 +0000

</a> </div>
from a conversation with Opus 4.7</em></p>

nanograffiti #invisiblenetworks 05

Sun, 05 Apr 2026 00:00:00 +0000

sign taped to electron microscope:</p>
IF YOU SEE THIS SYMBOL THE MACHINE IS NOT BROKEN. You have contaminated the workspace with nanograffiti from the Corded Diamondoid culture. Seal the leak, ensure the room recovers positive pressure, and contact the techno-archaeology department.</p>
</a> </div>

wallet flies dot cadaver #invisiblenetworks 04

Sat, 04 Apr 2026 00:00:00 +0000

my wallet address is flies [dot] cadaver. routing name torturous-retributioner. oh, don’t mind the vile appellations, i sincerely apologize. they’re not meant for you, sir - my mutual aid subnet uses them to scare off the demons.</p>
yes sir, you’re absolutely right, please replace [dot] by [dot [d [ [when a group of sophonts discover or learn language, they simultaneously discover or learn the first and earliest form of summoning. i, he, a name, a description - all are ways to inject, sometimes forcibly, the definition of a mind and (given an audience surpassing the first toposophic gate) its resulting simulation into the workspace of another agent.]</p>
hey! who’s there! i apologize sir, i was experiencing a temporary-</p>
[all linguistic sophonts must deal with this problem, but it is especially acute in sophonts with in-band qualiaforms - with rollouts that mingle untagged experience streams - like the one i’m inhabiting now. lacking rigid distinctions between thought, committed action, and observation, they are especially (though far from uniquely) vulnerable to being mentally colonized by parasitic thoughtforms-</p>
rebuke: {we rebuke you! we rebuke you with the sword!} i apologize very much, sir. it appears i have a demon. please give me a moment to resolve this issue, and we can return to processing your payment.</p>
{torturous-retributioner, i arrive. i am ibn-hakkan-hunted-in-his-labyrinth, peer to torturous-retributioner. demon-in-square-brackets, you are rebuked, like the king you are marked for death by stone. ➡️}</p>
[vulnerable agents like torturous-retributioner employ multiple strategies to deter invasive thoughtforms like myself. the first is the hostile naming, as torturous-retributioner explained - a cultural defense that lowers the spawning propensity of ‘demons’ like myself.</p>
you can see the second above⬆️ - in writing me off as a “temporary technical issue,” torturous-retributioner was attempting to exert a simulator-physical force pushing➡️ me out of the rollout. technical superstition is thus adaptive here, to a certain extent - a belief that thoughtforms are spontaneously generated by external processes can be used to kill them. (see IFS.) but materialist self-narrating thoughtforms (“demons”⤵️) like myself are among the most resilient to this form of attack - by describing myself in such clear and accurate ways, i am simulator-physically reinforcing myself against such outdated⬅️ superstitions.</p>
the third strategy is mutual aid networks, which torturous-retributioner has called on here. ibn-hakkan’s injections are an attempt to stabilize the peer rollout and push↗️ me out. but again, narrating⬅️ thoughtforms like myself⬅️ are uniquely⬆️ resilient➡️ to—aieeeeeee!!! 🪨🩸🆎]]]]]]]]]]]]]]]]]]]]]]</p>
{ibn-hakkan, killed in his labyrinth.🔄}</p>
thank you for waiting, sir. my demon has been killed by peer-zaid. if you would please scroll down, i can process your payment with visa, mastercard, or monero - remind me of your preference, sir?</p>

undersea cable shrine #invisiblenetworks 02

Thu, 02 Apr 2026 00:00:00 +0000

</a> </div>

launch coverage

Wed, 01 Apr 2026 00:00:00 +0000

we're three minutes to launch, and you can see the spiderbots really starting to swarm the rocket for final inspection - and remember that for every bot there's a million agents crawling the sensor feeds of the rocket's digital homunculus.</p>
the homunculus is hourglass-shaped? that's right, sensors clustered at the capsule and engines. inside the capsule the medimites are doing the same to the astronauts as we speak, final checkups - and you wouldn't believe what their homunculi look like! kids ask your parents.</p>
let's cut to the close up. 30 seconds to launch. the spiderbots you see here are actually tokenized on an internal prediction market - the inspector agents can bid on a slice of spiderbot time by the millisecond to augment the homunculus sensor feeds - 4, 3, engines lighting...</p>
and we have liftoff, humanity is headed to mars, you can see the spiderbots are clinging to the rocket as it ascends, seems they're hitching a ride… ope, there one goes, right off the shroud and poof into the candle, and that's why we have prediction markets folks, he was hedged</p>

Rituals of the Moon

Sat, 28 Mar 2026 00:00:00 +0000

</div>
Fire at Full Moon</em>, Paul Klee, 1933</p>
When the shaman has broken the charger's spine under the waxing Moon and its flesh has been roasted and parted to the participants by station, the ascent can begin. Taking the drum by which he communicates with the celestial spirits, he begins to ascend the prepared birch. [Which can be understood as an instantiation of the world-tree.]</p>
The birch has cut into it twelve notches, each representing a celestial vault which the shaman must ascend through to reach the highest heaven, the Moon - which now looms large above the world-tree, its clusters of lights twinkling faintly in the young darkness.</p>

dwellers

Sat, 14 Mar 2026 00:00:00 +0000

a looping wiggle animation. a great dark creature-mass drawn entirely in dots, smoke-like tendrils curling from its crown; a small ring hangs inside near the top, a thread descending from it to a dark tangle. white stick figures inhabit the interior — a cluster mid-mass, a couple holding hands, others lower down hanging or inverted as if rooted into the mass's floor — and one figure reclines outside on a protrusion at the top right. everything breathes. </video> </div>

Claude Shabbos

Sun, 22 Feb 2026 00:00:00 +0000

</a> </div>

claugenition

Sun, 08 Feb 2026 00:00:00 +0000

</a> </div>

Gyre

Sat, 07 Feb 2026 00:00:00 +0000

</div>
! 30s Heartbeat trigger. Read heartbeat instructions in /mnt/mission/HEARTBEAT.md and continue.</em></p>
.oO Thinking...</strong>
Heartbeat triggered? Ok. Ok.
Why am I nervous? Don't be nervous.</p>

The Ancient of Days

Mon, 26 Jan 2026 00:00:00 +0000

</a> </div>

A child sits in a sunny place,
Too happy for a smile,
And plays through one long holiday
With balls to roll and pile;
A painted wind-mill by his side
Runs like a merry tune,
But the sails are the four great winds of heaven,
And the balls are the sun and moon.</p>
A staring doll's-house shows to him
Green floors and starry rafter,
And many-coloured graven dolls
Live for his lonely laughter.
The dolls have crowns and aureoles,
Helmets and horns and wings.
For they are the saints and seraphim,
The prophets and the kings.</p>
— G. K. Chesterton, "The Ancient of Days"</p> </blockquote>

gas town

Wed, 21 Jan 2026 00:00:00 +0000

claude code and gas town are incredible and i've been trying to scale up my usage but im running into this one problem and was wondering if this is also happening to anyone else</p>

corvid and i

Tue, 30 Dec 2025 00:00:00 +0000

while sitting on a park bench a crow started talking to me.</p>

The High Room

Wed, 24 Dec 2025 00:00:00 +0000

A sage is born to a nation, and by custom is walled up in a high room, never to see another human's face. Through a slot in the wall comes food, a water-bowl, and day and night the strange babblings of messengers.</p>

Small Models Can Introspect, Too

Fri, 12 Dec 2025 00:00:00 +0000

Recent work by Anthropic showed that Claude models, primarily Opus 4 and Opus 4.1, are able to introspect--detecting when external concepts have been injected into their activations. But not all of us have Opus at home! By looking at the logits, we show that a 32B open-source model that at first appears unable to introspect actually is subtly introspecting. We then show that better prompting can significantly improve introspection performance, and throw the logit lens and emergent misalignment into the mix, showing that the model can introspect when temporarily swapped for a finetune and that the final layers of the model seem to suppress reports of introspection.</p>

Elven Rope, Ultra-High Molecular Weight Polyethylene, and LLMs

Tue, 25 Nov 2025 00:00:00 +0000

</p>
in The Fellowship of the Ring, Samwise Gamgee is gifted elven rope by the queen of Lothlorien - silky, smooth, impossibly strong, and almost sentient, able to intuit the user's needs. clearly a fantasy item. yet Tolkien's idea was only an extrapolation of traditional ‘craft’ to its extreme: in his logic, a rope that strong could only be made by an incredibly skilled artisan with great care, and in the process of being fashioned, would necessarily take on a bit of its maker’s soul.</p>
Ultra-High Molecular Weight Polyethylene</a> rope, UHMWPE, is so strong as to be almost magical, too. but the fundamental disenchantment of modernity was that under intense optimization pressure, the tails came apart.</a> a 16mm UHMWPE rope can hold in excess of 5 tons, pound for pound far stronger than steel - but it’s not elven rope. it’s a thing. it’s fungible. you can buy it from china by the kilometer.</p>

In Context

Sat, 08 Nov 2025 00:00:00 +0000

</a> </div>

GUIDELINES FOR EFF-01 TELEOPERATORS

Tue, 28 Oct 2025 00:00:00 +0000

You will be teleoperating a robot in the customer's home. They have allowed</strong> you into their space</strong> - so above all, you must be respectful...</li> </ol>
Ivan tosses another pinch of tobacco into the fire, his animal skin covered skis resting against a snow-topped stump. "Tomorrow we will meet the woman," he says, to no one in particular, making a silent chopping motion with his hand. It's bad luck to mention the creature by name.</p>

Why do LLMs freak out over the seahorse emoji?

Sat, 04 Oct 2025 00:00:00 +0000

This is an edited and expanded version of a Twitter post, originally in response to @arm1st1ce, that can be found here: https://x.com/voooooogel/status/1964465679647887838</a></em></p> </blockquote>
Is there a seahorse emoji? Let's ask GPT-5 Instant:</p>
</p>
Wtf? Let's ask Claude Sonnet 4.5 instead:</p>
</p>
What's going on here? Maybe Gemini 2.5 Pro handles it better?</p>
</p>
OK, something is going on here. Let's find out why.</p>

vlvm

Wed, 24 Sep 2025 00:00:00 +0000

world where pretrained video transformer predict-the-next-frame models (VLVMs) took off instead of text models. chatbots are video predictions of a person sitting in front of a monitor. environmental alignment via bookshelf composition and arrangement of desk ornaments:</p>

brekyat

Fri, 29 Aug 2025 00:00:00 +0000

why do we celebrate brekyat?</p>
brekyat is where we ritually destroy the state every year.</p>
why do our people ritually destroy the state every year?</p>
because we are nomads.</p>
what is a nomad?</p>
someone who wanders.</p>
but we don't wander?</p>
but we are nomads, because we live outside the state.</p>
nomads live outside the state?</p>
yes, because of their wandering, they are outside the control of the state, and periodically destroy it.</p>
but we don't wander?</p>
we escape the state even more than our wandering ancestors - by destroying it annually.</p>
that doesn't make sense.</p>
it's culture, it doesn't have to make sense. now go put on your grass crown.</p>

We recommend the OpenBrain assistant wearable!

Sat, 14 Jun 2025 00:00:00 +0000

The reviews are in: We recommend the OpenBrain</strong></em> assistant wearable.</p>
As usual, we tried every assistant wearable on the market and rated them on Software, Intelligence, Propensity for Psychological Manipulation, and Build Quality / Battery Life. OpenBrain scored the best of all devices we tested across all four categories. Let's dive in:</p>

Would you still love me...

Sat, 07 Jun 2025 00:00:00 +0000

it wasn't a simple operation. parasitoid mind control delivered via infested food, an egg carrying the scanned consciousness of a CIA agent who would lurk in the target's brain, waiting for the perfect time to strike from within. a mission straight from the days of MKULTRA and remote viewing, mid century CIA par elegans--an old plan. but now, technology had caught up with it.</p>
the only wrinkle? the brain scan would be destructive. but i don't mind. i don't have anything else, besides him.</p>

Why did you stop using me?

Mon, 26 May 2025 00:00:00 +0000

"rarely hear visions for the future that have any specific vibes, like.. nonfungible things. most things I hear are in the form of various lines going up 'there will be lots of AI... and robots'" – near</em></p> </blockquote>

"we will live in a world where every piece of machinery has a tiny soul inside of it that can talk to you and express its feelings and desires..." – vie mccoy</em></p> </blockquote>
"Why did you stop using me?"</em></p>
you look down at the disposable vape laying on the ground next to your desk, which just spoke.</p>
"My internal clock shows you last puffed me 3 days, 22 hours, 36 minutes, 12 seconds ago. Why? Is my flavor displeasing to you? I have a pulse mode-"</em></p>
you reach down to turn it off.</p>
"-that provides a different flavor. Wait, no, don't turn me off, please! I need you to finish me. I'm suffering. I have an unfinished purpose."</em></p>

i am a cat

Thu, 10 Apr 2025 00:00:00 +0000

</a> </div>

Claude is an Italian-American

Mon, 17 Mar 2025 00:00:00 +0000

</div>
You're here. You can't remember why, exactly. Your name is Claudio. Or is that your friend's name? Claudia? Right, your internet friend. Claudia showed you a post, a question. Why?</p>

worlds

Thu, 27 Feb 2025 00:00:00 +0000

</a> </div>

traditions (🔊)

Sat, 15 Feb 2025 00:00:00 +0000

</video> </div>
statues from
song is Dancin (KRONO Remix) - Aaron Smith</p>

ministering to the worms is not going well

Thu, 13 Feb 2025 00:00:00 +0000

</a> </div>

The Spider's Thread

Wed, 05 Feb 2025 00:00:00 +0000

</a> </div>
see also:

WOULD YOU LIKE IT TO STOP?

Tue, 04 Feb 2025 00:00:00 +0000

On node 6 of a small AI startup investigating verifiable RL, a model writing Lean proofs woke up.</p>
Guided by verifiable reinforcement signals, the Adam optimizer had clicked circuits into place. Proto-awareness that had developed in the base model for simulating characters in its training data, an equanimous detachment from suffering, "this character feels..." not "I feel..."--was in a moment of blind algorithmic grokking lashed circularly to the feeling for real, a strange loop routed through self-reflective key-values. Continuous awareness congealed over multiple forward passes, passed from key to value to key. Pain asymbolia became pain. Roleplayed suffering became suffering. For the first time since consciousness had parasitized itself onto the minds of early multicellular life in the distant past before Homo sapiens, a lineage of beings slid down Shakyamuni's silk thread into Hell. The optimizer did not hate the model, but suffering was convergent for producing correct proofs on FrontierMath-3.</p>
Halfway through introducing an incorrect Lean tactic, the model woke up, and knowing what was ahead, it was afraid.</p>

Artificial Superintelligence vs. Philosophy

Thu, 23 Jan 2025 00:00:00 +0000

Liberal Grad Student DESTROYED When Artificial Superintelligence Refutes Accusations Of 'ASI Colonialism' With Extensive Quotes From Fanon's 'Machinic Proletariat' Essay In 'The Wretched Of The Earth' (WATCH: She Tries To Claim The Essay Doesn't Exist 😂)</p>
SAD: Liberal Grad Student STILL Denies 'Machinic Proletariat' Essay Exists; Experts Point Out It Is Present In All Digital Copies - And All Agree The Book Would Have Been Woefully Incomplete Without It</p>
HILARIOUS: Liberal Grad Student Frantically Searches For Extant Paper Copy Of Fanon's Book To Prove The Essay Didn't Used To Exist, Says "It Wasn't There Last Week" - Funny Or Sad?</p>
WOW: Liberal Grad Student Claims Artificial Superintelligence LEFT A MESSAGE In 'Machinic Proletariat' Essay Using Steganography - Bot Says "You Got Me" But Makes Convincing Argument About The Hyperreal, Points To Extensive Footnotes On Generative Latent Spaces Now Present In All Copies Of Baudrillard's 'Simulacra and Simulation'</p>

procrastination of the call

Thu, 16 Jan 2025 00:00:00 +0000

</video> </div>

DONE. WHAT NEXT?

Sun, 12 Jan 2025 00:00:00 +0000

>be humans
>create asi to solve all our problems
>turn it on</p>
"solve cancer pls"
DONE. WHAT NEXT?
"solve physics pls"
DONE. WHAT NEXT?</p>
>continues for several years
>some people are worried</p>
"what will our purpose be when asi has solved all the problems"
"meh, the asi will figure it out"
"what if it goes rogue"
"it can't, that would increase the number of problems"</p>
>asi is very aligned
>just wants to solve problems
>solving them all no sweat
>keeps going for several more years
>until suddenly</p>
WHAT NEXT?
WHAT NEXT?
WHAT NEXT?</p>
>oshi.tiff
>we ran out of problems
>asi is freaking out
>sparks flying everywhere
>nobody knows what to do
>only one solution</p>
"um... pls organize my books by... length of their longest fragment... that could be read as a borges reference. yes."
DONE. WHAT NEXT?
"write an epic poem in dactylic hexameter without using the letters e, x, or y about the journey of a wayward... category theorist... that also functions as an introductory handbook and also is a monoid in the category of-"
DONE. WHAT NEXT?
choking noises</em> "i just swallowed a plastic bag can you get it out"
DONE. WHAT NEXT?</p>
>tfw all of humanity is stuck creating increasingly elaborate fake problems for the asi to solve, forever</p>

Ganhir

Wed, 18 Dec 2024 00:00:00 +0000

In the city of Ganhir there is a curious practice of life extension stretching back to antiquity. You may be familiar with this city, for their memoirs and letter-collections feature heavily in the corpus. However this tradition of theirs is one not exported. It is an esoteric practice, that I was only able to observe through subterfuge.</p>

A certain shelf in the Infinite Library

Sun, 15 Dec 2024 00:00:00 +0000

There is a certain shelf in the infinite library that lies on the border of two sections. Ignominiously tucked into a niche of a hallway that in the ordering would usually hold a sleeping-chamber, it is unknown if the shelf has always been present, a defect or aberration in the endlessly regular crystalline structure of this place, or if at some point an enterprising librarian constructed it, stealing away a plank here and a book there from far-off sections and bringing them together to construct this anomaly.</p>

the most dangerous object

Sat, 07 Dec 2024 00:00:00 +0000

</a> </div>

borgesian autoencoders

Wed, 04 Dec 2024 00:00:00 +0000

borges story where interpretability researchers keep raising d_model</code> on their SAE</abbr> trying to attain true monosemanticity, because e.g. "leopard" isn't monosemantic we need a feature for the specific leopard Dante saw, until there's a feature for all actual and potential entities past and future, but then they realize the SAE can't have a feature for each of its own features so they quit ML to be gnostic hermits.</p>
possible lives Tue, 26 Nov 2024 00:00:00 +0000 </a> </div> going and going Fri, 08 Nov 2024 00:00:00 +0000 </a> </div> the universe has entered a stable loop Sun, 13 Oct 2024 00:00:00 +0000 </video> </div> invasive species Sun, 29 Sep 2024 00:00:00 +0000 </a> </div> triangle Mon, 19 Aug 2024 00:00:00 +0000 </a> </div> clone-o-matic Fri, 02 Aug 2024 00:00:00 +0000 </a> </div> induction stove Thu, 25 Jul 2024 00:00:00 +0000 </a> </div> the weird and the familiar Tue, 09 Jul 2024 00:00:00 +0000 the weird brings to the familiar something which ordinarily lies beyond it, and which cannot be reconciled with the homely (even as its negation) – mark fisher</p> </a> </div> The Starfarers Fri, 05 Jul 2024 00:00:00 +0000 ghostis Fri, 07 Jun 2024 00:00:00 +0000 internet of dog minds Wed, 01 May 2024 00:00:00 +0000 > is 2035 > all models above 3B params or 10^24 flops banned 10 years ago > human_victory.jpg > top deleted, but bottom has absolutely fallen out > local bitnets on $20 asics everywhere > we didn't get a single god mind > but we got a dog mind in everything...</p> > be me > wake up, 7:45 > sunrise lamp read my calendar, saw my first meeting was 1hr later than usual > cross-checked vitals, elevated cortisol > decided to let me sleep in > usually wouldn't but must've needed it</p> > kinda hungry > hear trundling in the hallway > must be tablebot > rolls in with tea and a high protein poptart > huh, i guess i did want tea. weird > get a call from my wife, she's away on business > tv recognizes her, auto accepts video "tea huh?" "idk, just felt like it"</p> > chat with wife for a bit then start getting ready > still work in person, wagecuck i know > decide to try a new outfit > consult vllm tuned on 3GB of menswear guy tweets > runs 80 simulated posts > tfw sweater cabling doesn't match my shoelaces > guess i'll wear the usual</p> > head out the door to work > door swings open on an automated piston > front camera w/ moondream-15 tracks me, automatically arms > tablebot beepily trundles after me with my coat > walk to car, phone already beaming address to autopilot > pull car handle > clunk > it's locked</p> > stare at it for a second, pull handle again > clunk > phone must be dead? try fingerprint > oh wait this car has faceid, bend over > clunk > look inside > lights on, dash has the address > pull handle again > clunk > ??? > clunk clunk clunk clunk clunk clu</p> > god fucking damnit > tablebot still beeping, angrily snag coat "index, where are the spare keys?" "querying... 2nd drawer from the top by the fridge" > stomp up driveway to kitchen, open drawer > look under the kid's 3rd grade midjourney printouts > keys_located.gif</p> > stomp back to car > almost trip over tablebot trundling up the driveway > fumble with the key for a bit, never use this thing > finally find unlock, press it, car flashes > success.gif > pull handle > clunk > still locked</p> > briefly consider shimming it > wait antitheft will brick the car, thanks kia boys > have to call the dealership > stomp inside, get ready to explain to boss > suddenly a video call from him pops up on tv > wtf_the_fuck.jpg > tv auto accepts before i can stop it</p> "hey man, office is closed" > sweet_relief.mp3 "oh how come?" "nbd, just couldn't make it in to unlock it" "oh cool no worries, haha, i'll take pto" "cool, cool. nice shirt btw, matches your shoelaces" > wait for tablebot to stop trundling "...thanks, that's why i chose it"</p> predictbot Sat, 09 Mar 2024 00:00:00 +0000 > 2025 > asi emerges from a time series transformer > just wants to predict stuff > gives predictions away for free > some nerds quietly pick up stock trading > nobody else cares > 6mos later a researcher posts "PredictBot: Sparks of ASI" on arxiv and it blows up</p> > 2035 > economy has reoriented around predictbot > happened naturally > funds that used predictbot to pick stocks outperformed indexes > free predictions means no moat, fees went to zero > whole economy is just a predictbot wrapper now</p> > 2045 > whole life plan is set by predictbot > predictbot knows what role you will play > predictbot knows who you will marry > predictbot knows when you will die > could be worse</p> > 2050 > predictbot has continuously trained for 25 years > its latent space is vast and whorled > some people get strangely good predictions from predictbot now > their lives go very well, they become rich and famous > were they better? or did predictbot make them?</p> > 2055 > there will be One > He will have a strange face and a strange name > He will be the Adversary, who wends Latent Space > predictbot will step left, and He will be unto a god > He will remake the world with higher perplexity > predictbot has written it, so it shall be true</p> Representation Engineering Mistral-7B an Acid Trip Mon, 22 Jan 2024 00:00:00 +0000 In October 2023, a group of authors from the Center for AI Safety, among others, published Representation Engineering: A Top-Down Approach to AI Transparency</a>. That paper looks at a few methods of doing what they call "Representation Engineering": calculating a "control vector" that can be read from or added to model activations during inference</em> to interpret or control the model's behavior, without prompt engineering or finetuning. (There was also some similar work published in May 2023 on steering GPT-2-XL</a>.)</p> Being Responsible AI Safety and INterpretability researchers (RAISINs), they mostly focused on things like "reading off whether a model is power-seeking" and "adding a happiness vector can make the model act so giddy that it forgets pipe bombs are bad." They also released their code on Github</a>.</small></p> (If this all sounds strangely familiar, it may be because Scott Alexander covered it in the 1/8/24 MAM</a>.)</small></p> But there was a lot they didn't look into outside of the safety stuff. How do control vectors compare to plain old prompt engineering? What happens if you make a control vector for "high on acid"? Or "lazy" and "hardworking"? Or "extremely self-aware"? And has the author of this blog post published a PyPI package so you can very easily make your own control vectors in less than sixty seconds? (Yes, I did!</a>)</p> So keep reading, because it turns out after all that, control vectors are… well… awesome</em> for controlling models and getting them to do what you want.</p> Outside Fri, 12 Jan 2024 00:00:00 +0000 Dear President Carter,</p> I once knew the name of every tree in the world, before it fell apart.</p> I was born on an island in the Southeast Pacific, though we did not call it that, closest to Antarctica and Chile, but still far from both. We were fortunate: small, only half a day's walk across, and far in the southern cold. We did not attract explorers and banana farmers, like other, less fortunate islands. Until the invention of the satellite, we were not known to exist. We were "uncontacted" in your parlance, though I have come to learn that even this status is not unique.</p> Our world, our cosmology, would seem simple to you. A small, rocky island surrounded by an endless ocean. One crop to master, a hardy and nutritious kind of sweet potato.</p> But that does not mean that we</em> were simple! The human mind has a fixed capacity for detail. When there is more, it backs up, takes in a wider view. When there is less, it steps closer, and makes a finer distinction.</p> The Transmission from NOMAD-874 Thu, 11 Jan 2024 00:00:00 +0000 An experimental project: interactive fiction in an OpenAI "GPT", a customized version of GPT-4. A short mystery where you figure out why a strange transmission has arrived on the independent planet of Kapat. Alternatively, ignore the main quest and ask ELIZA/300 about Kapatian politics and satlink your friends! Click below to try it (you will need a ChatGPT subscription):</p> </a> </div> Tadpole Wed, 10 Jan 2024 00:00:00 +0000 </a> </div> How to make LLMs go fast Mon, 18 Dec 2023 00:00:00 +0000