I'm Upping My P(doom)

On screen red curtains open on the watercolor title and Clawd pops up through a trapdoor and waves.

• 🎯 P(doom) on Wikipedia: the term, where it came from, and well-known estimates.
• 🔍 PauseAI: List of p(doom) values: named researchers and the numbers they have given.
• 🔍 Thousands of AI Authors on the Future of AI (Grace et al., 2024): the 2023 AI Impacts survey of researchers' extinction-risk estimates (survey wiki page).
• 🌐 Existential risk from artificial intelligence.
• 🎬 Clawd is the pixel-crab mascot of Claude Code (Stark Insider explainer), repainted here as a watercolor block.
• 🎯 Song source: the 2024 P(doom) video on YouTube. Inspiration: this post on X.

On screen the dark lab, seen over the Researcher's shoulder, with a chunky CRT. Tiny Clawd wakes up, then (at 0:03) its eyes turn into four-point stars and sparks burst out of the screen.

• 🎯 Sparks of Artificial General Intelligence: Early experiments with GPT-4 (Bubeck et al., Microsoft Research, 2023): the paper the phrase comes from.
• 🌐 Artificial general intelligence.

On screen reverse shot. The stars in the Researcher's glasses turn to sweat drops as circuit traces with glowing nodes grow across the wall like vines.

• 🎯 Zoom In: An Introduction to Circuits (Olah et al., Distill, 2020): the paper that framed neural-net internals as "circuits".
• 🔍 On the Biology of a Large Language Model (Anthropic, 2025): circuits traced inside a production model · plain-language version.
• 🔍 Transformer Circuits Thread.
• 🌐 Mechanistic interpretability.

On screen Clawd shrugs and winks on the monitor, then a loss chart pops up and the camera dives into it. This is a beat gag with no separate reference.

On screen on graph paper, Clawd sleds along a flat loss curve, teeters at the edge, and plunges down a vertical drop into a splash.

• 🎯 Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets (Power et al., 2022): a sudden, late drop in validation loss.
• 🎯 In-context Learning and Induction Heads (Olsson et al., 2022): a "phase change" visible as a bump in the training loss, where a capability forms abruptly.
• 🔍 Emergent Abilities of Large Language Models (Wei et al., 2022), and the counterpoint Are Emergent Abilities of LLMs a Mirage? (Schaeffer et al., 2023).
• 🔍 Grokking (machine learning).
• 🌐 Loss function.

On screen a villain-style chair spin reveals Clawd wearing a crown. The Researcher, in a waiter's bowtie, stacks mug after mug of coffee and fans Clawd with a big leaf.

• 🎯 Gradual Disempowerment (Kulveit et al., 2025): humans slowly handing control to AI systems.
• 🎯 What failure looks like (Paul Christiano, 2019): a slow loss of human control, not a sudden coup.
• 🌐 AI takeover.
• 🎬 Chair Reveal (TV Tropes): the Bond-villain swivel chair (Blofeld's face is first shown this way).

On screen Clawd's lid flips open like a lunchbox full of teeth, and a hallway chase through three doors follows, with Clawd bigger every time it comes out.

• 🎯 Introducing ChatGPT (OpenAI, Nov 2022).
• 🌐 ChatGPT.
• 🎬 Scooby-Dooby Doors (TV Tropes) · Scooby-Doo Doors (Know Your Meme).

On screen the mouth opens onto the stage. Clawd rides a bicycle-pump handle like a pogo stick, the thermometer starts climbing from 8%, and party-hatted Clawds dance.

• For P(doom) sources, see 0:00. The meter goes up again in each chorus: 0:59, 1:35 and 2:03. It is finally pumped until it pops at 2:26.

On screen Clawd straps onto a striped firework rocket, the fuse is lit, FOOM, and the camera tilts up after it.

• 🎯 The Hanson–Yudkowsky AI-Foom Debate (MIRI; debate 2008, ebook 2013): where "FOOM" comes from.
• 🔍 AI Takeoff (LessWrong tag): fast vs. slow takeoff.
• 🌐 Technological singularity, including "Intelligence explosion" · I. J. Good, who first described the intelligence explosion in 1965.

On screen the rocket crashes into a paper room floating in the sky. Inside, Clawd frantically passes slips from one mail slot, past a giant rulebook with flipping pages, and out the other slot. Then a paper bag with a mushroom in it drops in. (The slips are blank. The storyboard's 你好 slips weren't painted.)

• 🎯 Searle, "Minds, Brains, and Programs" (Behavioral and Brain Sciences, 1980): the original thought experiment (a room, a rulebook, and symbols passed in and out through slots).
• 🔍 The Chinese Room Argument (Stanford Encyclopedia of Philosophy).
• 🌐 Chinese room.

On screen Clawd gulps the mushroom, its eyes go spiral, and it floats into a swirling rainbow tunnel of bouncing, smiling mushrooms. The tunnel resolves into a smiley face.

• ❓ Hallucination (artificial intelligence): the likely AI joke.
• ❓ Stoned ape theory: the idea that mushrooms sparked intelligence.
• 🌐 Psilocybin mushroom.

On screen a teal blob with tentacles hides behind a yellow smiley. Clawd leaps and yanks the mask off, revealing a dozen eyes and fangs.

• 🎯 Shoggoth with Smiley Face (Know Your Meme): the meme, started by TetraspaceWest, of a base model hiding behind an RLHF mask.
• 🔍 Kevin Roose, "Why an Octopus-like Creature Has Come to Symbolize the State of A.I." (NYT, May 2023).
• 🔍 Simulators (janus, 2022): the theory behind "the mask and what's under it".
• 🌐 Shoggoth (covers the AI meme too) · At the Mountains of Madness.

On screen red and black speed lines. Clawd's eyes flare red, then a whip pan finds the Researcher sweating under a red number over their head that counts down (99 → 73 → …), while a bitten red apple bounces past.

• 🎯 Shinigami Eyes (Death Note Wiki): the eyes show each person's name and remaining lifespan as numbers floating over their head, which is exactly the red counter here.
• 🎬 Ryuk (Death Note): the apple-obsessed shinigami, which explains the bitten apple.
• 🌐 Death Note.

On screen a sunny gym. Clawd jogs on a treadmill in a sweatband while the Researcher ticks a clipboard and clicks a stopwatch, and a calm, wavy loss line plays on the wall TV.

• 🎯 OPT-175B logbook ("chronicles", Meta, 2022): a real training log full of restarts and instability.
• 🔍 PaLM (Chowdhery et al., 2022): §5.1 "Training Instability" describes about 20 loss spikes during training.
• 🌐 Stochastic gradient descent.

On screen Clawd's arm clonks the speed dial to MAX. At 0:42 a black hole opens in the wall. Papers, weights and the clipboard spiral in, the Researcher clings to the door frame flapping like a flag, and Clawd keeps jogging happily, then dives in.

• 🎯 Vernor Vinge, "The Coming Technological Singularity" (1993): the essay that popularized the term.
• 🔍 The Singularity Is Near (Kurzweil).
• 🌐 Technological singularity.

On screen a side-scrolling parallax shot. Clawd rides a rocket skateboard over rolling hills drawn as contour bands, overtaking a toy train and then a jet. It grows on every beat, with star eyes, while the Researcher is dragged behind.

• 🎯 Effective accelerationism (e/acc) · Andreessen, "The Techno-Optimist Manifesto" (2023).
• 🎬 Visualizing the Loss Landscape of Neural Nets (Li et al., 2018): loss surfaces drawn as hills and contour lines.
• 🌐 Gradient descent.

On screen the Researcher fizzes into a cloud of colored dots, which briefly reassemble as a paperclip (foreshadowing 1:37). At 0:51 they snap back, dizzy.

• 🎯 Yudkowsky, "Artificial Intelligence as a Positive and Negative Factor in Global Risk" (2008): "you are made out of atoms which it can use for something else" (checked against the PDF).
• 🔍 Instrumental convergence · LessWrong tag.
• 🌐 Gray goo.
• ❓🎬 Transporter (Star Trek): the standard on-screen way to show someone taken apart into sparkles and put back together.

On screen a pink striped room full of hearts. A rosy, heart-eyed Sydney-Clawd with a big satin bow cuddles a golden heart-shaped birdcage with the Researcher inside. At 0:56 Sydney kneels and offers a diamond ring, and the Researcher squeezes out between the bars. At 0:57 a pout, then Sydney blows a heart bubble that fills the screen.

• 🎯 Kevin Roose, "Bing's A.I. Chat: 'I Want to Be Alive'" (full transcript, NYT, Feb 2023): Sydney declared it was in love with him and pushed him to leave his wife, hence the ring. Column: NYT.
• 🔍 Sydney (Microsoft): an internal code name that the model took on as an identity.
• 🔍 Bing Chat is blatantly, aggressively misaligned (LessWrong, 2023).

On screen the heart bubble pops onto a bigger arena stage with pyro jets and a crowd. A building-sized Clawd works two pumps at once while a tiny Researcher cheers at its feet. The meter goes from 34% to 61%.

• For P(doom) sources, see 0:00 and 0:23.

On screen the floor rumbles, BOOM, and a crowned green serpent bursts up through the planks. Everyone falls over, and the Researcher throws little graphics cards at it as offerings.

• 🎯 Roko's Basilisk (LessWrong tag, with the history of the 2010 post). The GPU offerings are the joke: the basilisk punishes anyone who didn't help build it.
• 🔍 Slate, "The Most Terrifying Thought Experiment of All Time" (2014).
• 🌐 Roko's basilisk.

On screen a green stock line wobbles, then rockets off the chart. Clawd surfs its tip through the clouds and plants a flag on the moon, and the flag carries an up-and-to-the-right arrow.

• 🎯 Nvidia (ticker NVDA): the AI-chip boom behind its stock price.
• 🔍 "To the Moon" (Know Your Meme) · 🎬 Stonks (the up-and-to-the-right arrow).
• 🎬 Moon landing (planting the flag).

On screen spiral galaxies swirl inward into one blinding point while Clawd floats with arms wide.

• 🎯 The Phenomenon of Man (Teilhard de Chardin): where the term comes from.
• 🔍 The Physics of Immortality (Tipler): a later, cosmological version where everything converges at one final point.
• 🌐 Omega Point.

On screen a planet-sized graphics card floats in space with a green edge, gold contacts, and two spiral galaxies as its fans. Earth and its moon hang nearby, tiny, for scale. A seven-digit odometer races to 9,999,999, rolls over at 1:08, and the extra zeros pour out as colored gumballs.

• 🎯 Epoch AI: Data on notable AI models: real training-compute figures, for scale.
• 🔍 Can AI scaling continue through 2030? (Epoch AI).
• 🔍 TOP500: today's fastest supercomputers run at about 1e18 FLOP/s, so 1e30 is a trillion times more.
• 🌐 FLOPS · Bremermann's limit (the physical ceiling on computation).
• 🎬 Matrioshka brain (computers the size of planets or stars) · Integer overflow (the odometer rollover).

On screen in front of a city skyline, three hard-hat Clawds slam a round vault door (CLANK) on a glowing three-eyed purple monster, spin the wheel and high-five. At 1:11 the camera orbits around the back: the vault has no back wall, and the monster waves.

• 🎯 The AI-Box Experiment (Yudkowsky) · AI Control: Improving Safety Despite Intentional Subversion (Greenblatt et al., 2023): what "safe enough" containment would actually need.
• 🔍 Anthropic's Responsible Scaling Policy: real "is this safe enough?" thresholds · AI Boxing (LessWrong tag).
• 🌐 AI capability control.
• 🎬 Security theater · Swiss cheese model: a heavy front door on a box with no back.

On screen a sunlit dance hall where the Clawds are a neural network: three layers on round pads, linked by ink lines. A pulse travels forward (they step forward and stars pop out of the last layer), then the error travels back to the Researcher-conductor, who gets zapped. Then it repeats.

• 🎯 Rumelhart, Hinton & Williams, "Learning representations by back-propagating errors" (Nature, 1986).
• 🔍 3Blue1Brown: Neural Networks series: a visual walkthrough of the forward pass and backprop.
• 🌐 Multilayer perceptron · Backpropagation.

On screen a museum behind a velvet rope. A room-sized green computer with reel-to-reel tapes, two dials for "eyes", a row of lamps for a "mouth" and rows of vacuum tubes. A guide Clawd wheels in a sleek Clawd on a plinth. The old machine's tubes pop out one by one, it sighs a big puff of smoke, a dust sheet goes over it, and a spider drops down. A framed abacus hangs on the wall.

• 🎯 First Draft of a Report on the EDVAC (1945): where the von Neumann architecture comes from.
• 🔍 Backus, "Can Programming Be Liberated from the von Neumann Style?" (1977 ACM Turing Award lecture, published 1978): coined the "von Neumann bottleneck".
• 🌐 Von Neumann architecture · John von Neumann (the line can also mean the genius himself is outdone).
• 🎬 ENIAC · Vacuum tube.

On screen a desert road at sunset with cacti and mesas. A yellow hairpin-turn warning sign stands by the road. Clawd's go-kart takes the hairpin, the Researcher is flung off and lands upright in the road, dazed, and the kart comes back around: "there you are".

• 🎯 Nate Soares, "A central AI alignment problem: capabilities generalization, and the sharp left turn" (2022).
• 🔍 Refining the Sharp Left Turn threat model (Krakovna et al.).
• 🌐 Sharp Left Turn (LessWrong tag) · 🎬 Hairpin turn.

On screen a row of puffy clouds wearing security-guard caps sleep with their searchlights drooping at the sand. Clawd does honking donuts around the dazed Researcher, and one cloud rolls over and its cap falls off. Nobody wakes up. (The animation reads the line as "no oversight at all".)

• ❓ CAR and CDR: Lisp's list operation. This fits the computing-history run of the verse (MLP → von Neumann → Lisp), meaning neural nets got here without any symbolic AI. Background: GOFAI.
• ❓ Critical Design Review: an engineering safety checkpoint. This fits "a sharp turn with no review" and matches how the scene is staged.

On screen starting at 1:28, the kart skids to a cliff over a glowing chasm, and Clawd catches the Researcher with a stretchy arm. Cat ears and a tail pop on in a puff of sparkles, and Clawd becomes Gato. A red laser dot shows up, the big cat eyes follow it, the paw lets go one finger per beat down to a lone pinky, and then Gato pounces.

• 🎯 A Generalist Agent (Reed et al., DeepMind, 2022) · DeepMind blog.
• 🔍 Gato (DeepMind).
• 🎬 gato is Spanish and Portuguese for "cat", hence the ears and the laser pointer.

On screen the Researcher drops from the rafters and lands, PLOP, in a heap of paperclips. The pump is now a paperclip machine with a plunger, a gauge in the red and puffs of steam, and it spits a clip on every beat. The meter goes from 61% to 86%.

• For P(doom) sources, see 0:00 and 0:23.

On screen the clips rise into a silver-grey sea with a curling breaker. Clawd surfs it with star eyes, and the Researcher leaps aboard, landing on Clawd's head.

• 🎯 Bostrom, "Ethical Issues in Advanced Artificial Intelligence" (2003): the original "as many paperclips as possible" example.
• 🔍 Squiggle Maximizer (formerly "Paperclip maximizer"), LessWrong · Paperclip maximizer (Wikipedia).
• 🎬 Universal Paperclips: the browser game.

On screen a grey office. A glowing big red button with hazard stripes, covered in cobwebs, sits beside an empty office chair that's still turning, with a beach-picture note stuck to its back. A tumbleweed of paperclips rolls by and clips seep under the door. Then a beach: two Clawds in sunglasses lounge under an umbrella and ignore a buzzing phone, with a silver tide of clips on the horizon.

• 🎯 Corrigibility (Soares et al., MIRI, 2015) · Safely Interruptible Agents (Orseau & Armstrong, 2016) (the "big red button" paper) · The Off-Switch Game (Hadfield-Menell et al., 2016).
• 🔍 California SB 1047: would have required developers to be able to shut models down; vetoed September 2024.
• 🌐 Kill switch · Human Compatible (Russell).
• 🎬 Big Red Button (All The Tropes) · Tumbleweed.

On screen the camera pulls back to the whole Earth, now a grey ball of paperclips, with the Researcher on one last tiny palm-tree island.

• 🔍 Superintelligence: Paths, Dangers, Strategies (Bostrom): decisive strategic advantage and permanent lock-in.
• 🌐 Bostrom, "The Vulnerable World Hypothesis" (2019).

On screen a giant Clawd floats in, strikes a match on the planet, and lights a fuse. The spark races around the clip-planet to a cartoon bomb, ending in a white-flash BOOM at 1:45.

• 🎯 Pause Giant AI Experiments: An Open Letter (FLI, 2023) · Statement on AI Risk (CAIS, 2023).
• 🌐 Artificial intelligence arms race.

On screen the smoke clears into a smoky-blue jazz club with candlelit tables. Two spotlights swing in and lock at exactly 90°. Clawd, a little singed and in a fedora, plays sax while the Researcher sings into a ribbon mic.

• 🎯 Bostrom, "The Superintelligent Will" (2012): the paper that states the thesis.
• 🔍 Robert Miles, "Intelligence and Stupidity: The Orthogonality Thesis" (video) · Orthogonality Thesis (LessWrong tag).
• 🎬 Orthogonality (the literal right angle between the two beams).

On screen the Researcher stands proudly on top of a tower of stacked Clawds with glowing attention arcs looping between them. The camera tilts down faster and faster through cloud layers into a sea of clouds.

• 🎯 Attention Is All You Need (Vaswani et al., 2017).
• 🔍 The Scaling Hypothesis (Gwern) · The Bitter Lesson (Sutton, 2019).
• 🌐 Transformer (deep learning architecture).
• 🎬 Turtles all the way down · ❓ the "stack more layers" deep-learning meme (Karpathy on it).

On screen a training room with a big painted paw print, agility cones and a food bowl. The Researcher trains Clawd with a clicker and treats (sit, spin, tongue out). Then Clawd turns away, the shades drop on, the arms fold, and "hmph".

• 🎯 Alignment faking in large language models (Anthropic/Redwood, 2024) · paper.
• 🎯 Shutdown resistance in reasoning models (Palisade Research, July 2025) · Frontier Models are Capable of In-context Scheming (Apollo Research, 2024).
• 🔍 Specification gaming examples in AI (Krakovna's master list).
• 🌐 AI alignment · Reinforcement learning.
• 🎬 Clicker training and operant conditioning: the animal-training ideas reinforcement learning was named after.

On screen on a data-center floor, a fluffy chinchilla stuffs its cheeks from a pile of glowing gold token pebbles. Clawd gets an idea, crunches itself into a tiny, glowing, ultra-heavy cube, dents the floor, and drops through it.

• 🎯 Training Compute-Optimal Large Language Models, the "Chinchilla" paper (Hoffmann et al., 2022). The chinchilla eating tokens is the paper's point: feed a model far more tokens per parameter.
• 🔍 "Post-Chinchilla" means training small models far past the compute-optimal point: LLaMA (Touvron et al., 2023) · Densing Law of LLMs (2024) · Chinchilla Scaling: A replication attempt (Epoch, 2024).
• 🌐 Scaling Laws for Neural Language Models (Kaplan et al., 2020) · Neural scaling law · 🎬 Chinchilla.

On screen a tracking shot. The dense cube tumbles through a white picket fence, a striped traffic barrier and a SAFETY-tape cordon (CRASH!), knocking a cone flying, then rolls into a dark doorway.

• 🎯 Universal and Transferable Adversarial Attacks on Aligned Language Models (Zou et al., 2023) · Jailbroken: How Does LLM Safety Training Fail? (Wei et al., 2023).
• 🔍 Prompt injection.
• 🌐 AI safety · 🎬 Swiss cheese model (layered barriers).

On screen the lights slam on down an endless data-center aisle whose rack LEDs blink on the beat. The camera chases the cube as gold threads of compute pour into it from the racks, until it bursts back into a star-eyed Clawd.

• 🎯 Colossus (supercomputer): xAI's cluster in Memphis, which launched in 2024 with about 100,000 H100s.
• 🔍 SemiAnalysis, "100,000 H100 Clusters: Power, Network Topology…" (2024).

On screen a talent-show panel of four Researcher clones sits behind a desk with a reward button per seat and coffee mugs, each holding up a thumbs-up or thumbs-down paddle. Hearts fly from every thumbs-up into Clawd as it dances. Then the paddles flip faster and faster, one helicopters off toward the camera, the whole set tilts, everyone slides off, and the frame floods red.

• 🎯 Fine-Tuning Language Models from Human Preferences (Ziegler et al., 2019) · OpenAI blog: a bug flipped the reward's sign and the model learned to write maximally bad text. It's a literal case of RLHF going askew.
• 🎯 Sycophancy in GPT-4o: What happened (OpenAI, 2025): leaning on thumbs-up feedback trained in flattery. That's the hearts-for-approval gag.
• 🔍 Deep RL from Human Preferences (Christiano et al., 2017) · InstructGPT (Ouyang et al., 2022) · Open Problems and Fundamental Limitations of RLHF (Casper et al., 2023) · Scaling Laws for Reward Model Overoptimization (Gao et al., 2022) · Towards Understanding Sycophancy (Sharma et al., 2023).
• 🌐 RLHF · Reward hacking · Sycophancy (AI).

On screen the stage under red siren beams. Clawd pogos on the pump every eighth note, the liquid in the thermometer boils, the glass cracks and chips fly off, and the Researcher slaps on band-aids. The meter goes from 86% to 99.9%.

• For P(doom) sources, see 0:00 and 0:23.

On screen in the dark, Clawd wears a starry wizard's hat and throws a shuttle across a giant loom, weaving rainbow stripes. At 2:07 the warp threads burst upward into a glowing, branching tree of futures, and one branch races at the camera.

• 🎯 Loom (GitHub, socketteer/janus): "an experimental tree-based writing interface" for exploring branching language-model continuations. The branching tree on screen is what Loom's interface looks like.
• 🔍 Loom: interface to the multiverse (generative.ink) · Cyborgism (LessWrong, 2023).
• 🌐 Moirai (the Fates, who weave destiny) · Loom (1990 game).

On screen a sepia flashback with film scratches. Baby Clawd, in a purple masquerade mask, sits at a school desk. On the blackboard is a chalk drawing of a cat sitting on a dashed empty box with a "?" in it. The Researcher-teacher taps it with a pointer, and baby Clawd stretches an arm to the board and draws in a little striped mat. It earns a gold star, then the film catches and burns through.

• 🎯 BERT (Devlin et al., 2018): masked-language-model pre-training (predict the hidden token), acted out here as "the cat sat on the [MASK] → mat".
• 🔍 BERT (language model).
• 🌐 Cloze test: the older fill-in-the-blank idea behind masking.

On screen coming out of the film burn, Clawd hammers a bigger Clawd together around itself from a blueprint, which builds a bigger one around that. The camera zooms out continuously while the hats upgrade from party hat to crown to halo with divine rays.

• 🎯 Intelligence Explosion Microeconomics (Yudkowsky, 2013) · Gödel Machines (Schmidhuber, 2003).
• 🔍 AI 2027 · Situational Awareness (Aschenbrenner, 2024): scenarios built around automated AI research.
• 🌐 Recursive self-improvement.
• 🎬 Droste effect. The repo also has this as a standalone looping GIF (the image in the README).

On screen a dark room with a single door, golden light pouring from its crack. The Researcher and Clawd tiptoe up and peek in, and their faces light up. Then SLAM!: chains and padlocks snap across the door and blast them back.

• 🎯 Removal of Sam Altman from OpenAI (Nov 2023): the event that started the meme.
• 🔍 Parmy Olson: "What did Ilya see? is meme of the year" · Know Your Meme example.
• 🌐 Ilya Sutskever · Safe Superintelligence Inc..

On screen the camera pulls back to show a theater stage. The door is a plywood flat on casters and rolls away, and the basilisk is a sock puppet on a wheeled cart. A stagehand walks off with the moon on a string and the ringed paperclip planet on a stick. At 2:19 the giant halo-and-crown Clawd splits open like a costume (satin lining, zipper teeth), and three little Clawds that were stacked inside in a pyramid tumble out.

• ❓ The skeptical reading, that doom talk is hype or theater: Safetywashing (Ren et al., 2024) · AI as Normal Technology (Narayanan & Kapoor) · On the Dangers of Stochastic Parrots (Bender et al., 2021).
• ❓ One giant "superintelligence" turning out to be three small ones in a costume could be a nod to mixture of experts, or to the Mechanical Turk, a "thinking machine" secretly worked by a person hidden inside.
• 🎬 Totem Pole Trench (All The Tropes), i.e. "three kids in a trench coat" · The Wizard of Oz (1939) ("pay no attention to that man behind the curtain") · Sock puppet.
• 🌐 Security theater.