Ani's Cinderella, brought back

Ani was the program at the heart of Ken's MIT doctoral thesis, Creation of Computer Animation from Story Descriptions (1979), which he then substantially rewrote as the technical report AI-TR-540. It was not given a script. It was given a film description, written in Lisp: who the characters are (Cinderella is shy, good and graceful; the stepmother is evil and mean), how they feel about one another, and what each scene has to get across (in one, the stepmother keeps Cinderella and the Prince apart).

From that, Ani decided for itself how each character should move, fast or slow, smoothly or jerkily, in curves or straight lines, and what each should do in each scene. It worked in two passes. Pass 1 makes 102 choices about how the characters move, each by gathering suggestions from what Ani knows about personalities, relationships and films, and settling their conflicts. Pass 2 plans each scene: how to convey what the scene must, in which parts, with whom where, and for how long. Director, the actor language Ken had written, then animated the plan.

“It”, not “she”. This page calls Ani “it”. An early draft of Ken's thesis called Ani “she”; one professor insisted that he change it, and he did. The revival's film page called Ani “she” too, until 16 October 2026, when Ken chose “it” again.

The film, Cinderella in five scenes, was shot on 16mm film from a color TV screen, with Ken's spoken narration; its title card is dated Friday, 13 October 1978. (There is now also a new version made by the revival's app, with subtitles.) In it the characters are simple shapes:

Ani is not one program following one plan. AI-TR-540 describes it as a community of about 500 small experts (p. 136): one for each word a character can be described by, such as "shy" or "hates", one for each kind of value (low, medium, high), others for keeping the film's variety, for places on the screen, for ways of showing an action, and many more. They are built from actors, the message-passing objects of Carl Hewitt, Ken's thesis supervisor, which Ken's Director provides: each expert answers messages and keeps its own state.

No expert runs the show. When Ani has to choose, say, the stepmother's speed, the sources of suggestions each put forward a view with a strength: her personality, her relationships, how she compares with the others, the film's own levels of flashiness and variety, and, in a scene, her feelings. Ani weighs them not only by strength but by how the suggestions relate, where they come from and how many agree; conflicts end in a compromise or a rejection, and a choice that can't yet be made well is postponed while others are made. A small top level hands out the tasks and settles conflicts with the help of other experts. There is no script, only many weak voices combined.

This was at the MIT AI Lab while Marvin Minsky and Seymour Papert, both on Ken's thesis committee, were working out a theory of the mind as a society of agents. AI-TR-540 cites their draft, a book called "Society Theory of Mind", and says Ani's implementation was strongly influenced by the society view, though its components are less independent, and its experts less flexibly organized, than Minsky and Papert's agents (pp. 45 and 137). Minsky's own book, The Society of Mind, came out in 1986, seven years after Ani. So Ani wasn't designed from the book; it is a working program from the same lab and the same conversation, an early example of the idea in action.

Ken's files survived on the MIT AI Lab's backup tapes. When the tapes were read, decades later, each file was saved in a format that packs each 36-bit word of the lab's PDP-10 into five bytes. Undoing it gave back the text exactly, and a few small tools decoded the binary files too: compiled Lisp, directory listings, archives, pictures.

The best tape has a directory listing dated 8 January 1979, four days before Ken submitted his dissertation. It holds all 31 of Ani's source files, about 10,000 lines, which matches the thesis's own estimate of Ani's size. It holds the Cinderella description, CINDY 46. And it holds CP 29, the file pass 1 wrote for pass 2 in September 1978: 102 choices, 54 relative and 48 absolute, the very counts the thesis gives in Appendix III. That made an exact test possible. Given CINDY, a revived Ani should make the same 102 choices.

Part of the foundation was missing. Ani is written in Director, and the lowest layer of Director, Ken's KEN1; directory, was lost: about twenty files and some 490 definitions, including the macros, the utilities, the pattern matcher and the compiler for actors. A tags table named every function in them. Their code was gone.

The revival began with this message from Ken to Claude:

Earlier attempts to revive Ani with AI tools had produced imitations: new programs that acted out Ani's ideas. On the first morning Ken chose the other road. Claude would write a MacLisp interpreter in JavaScript, the Lisp Ken wrote in, and Ken's 1978 files would load on top of it exactly as they are on the tape. Only what was lost would be rebuilt, and in MacLisp, not JavaScript, so that everything above the interpreter is Lisp, as it was.

Two other roads were set aside. Emulating the whole PDP-10 is possible today, but it wouldn't have helped: Ani's compiled code needs the same lost runtime. A modern rewrite of Ani was kept for later, with the revival's traces of what Ani does as its specification.

The best evidence for the lost KEN1; code turned out to be a file called TS D2500: a Director saved in January 1980, its "2500 version", made to draw on the 2500 (Marvin Minsky's 2500 Logo computer), whose memory image holds compiled code for most of the lost runtime. Claude decompiled it file by file, from PDP-10 instructions back into Lisp. Where the 1980 code might differ from 1978's, older evidence decided: the June 1978 Director Guide (AIM-482), the call sites in Ani, Director compiled into plain Lisp, and Ani's own compiled files. Every rebuilt definition cites its evidence. The result is 12 files, 2,194 lines of MacLisp.

The same day, the TV turtle Director drew with was ported from LLOGO's public source, and Director's own demos ran unchanged.

Getting Ani's choices to match uncovered a piece of history. Ani's compiled files call functions, such as RECALL-VARIABLE and CHANGE-VALUE, that the surviving Director doesn't have. The second Director Guide explained why: much of Director had been rewritten, and its "demons" removed. Ken had rewritten Director around 21 November 1978, after Ani was written and two months after CP 29 was made by the old Director. The Director on the tape, dated January 1979, is the new one.

So the revival has a bridge, REVIVE;BRIDGE, loaded after Director, which puts back what the old Director did for Ani. Each change is argued from the evidence: the 1978 Guide, Ani's compiled files, Ken's own partial bridge of 22 November 1978, TS D2500, or CP 29 itself. Ken didn't remember whether he ran Ani after November 1978; the evidence says Cinderella was made on the old Director.

Ani makes no random choices, so nothing short of an exact match with CP 29 would show the rebuilt Director was right. On the first day (times in London):

The glass box came the same afternoon. Ani's Choice Points shows every choice point: each attempt Ani made at it, why Ani postponed it, the suggestion that won, the conflicts Ani settled and the suggestions Ani rejected; a timeline draws all 836 attempts. The stepmother's speed comes out as the thesis tells it.

Pass 2 plans each scene from where Director left the characters at the end of the one before, so the five go in order: plan a scene, film it, save the positions, plan the next. By 17:52 on the first day Ani had planned all five scenes and Director had filmed them; by 18:50 they were in color.

The plans read as the thesis describes them. The Introduction has five parts, estimated at 42.2 seconds (the thesis: 43.3 estimated, 42 real). Kept Apart's first part has the values the thesis prints in Appendix VI, and in its second the stepmother stands guard six times where four were planned, stretched by Ani as the thesis says. Appendix III says pass 2 begins with 418 actors and makes 154 more; the revival begins with 343 and makes 147.

AI-TR-540, the rewritten thesis, prints frames of the film in the corners of its pages, so that flipping the pages animates them. Claude recovered all 574 frames from the scanned report, registered so that they line up: a complete record of a 1978 run, at 4 frames a second, made by the same program on the same description. It became the ground truth for the film.

Kept Apart matches closely from start to finish, and its parts begin on the same frames. The Introduction matches exactly for its first 61 frames; then the stepmother's fourth push starts a frame or two late and drifts by about a shape's width. Nothing in the movement code is random, and rounding every number to the PDP-10's precision changed no frame. The likeliest cause is Director's object code of January 1979, since the 1978 one is lost.

The flipbook also settled two questions. The picture was 550 by 413 dots: at the TV's full size the Introduction runs 15 frames longer. And in 1978 Director's universe emptied its list of actors while they were taking their turns, so a check made during a turn could end the film early. That is what the flipbook shows: Kept Apart ends just after its third part, with the stepmother still standing guard.

Ken's TV previews ran at 4 frames a second, and so does the run in the flipbook. The film itself was shot at 24: Ken's setup for filming in color sets 24 frames a second. The YouTube transfer of his 16mm print agrees: its picture changes on about 21 of every 30 video frames. Director divides each movement by the frame rate, so at 24 the characters take smaller steps. Ani makes the same choices, but the scenes leave the characters in slightly different places, so from No Longer Kept Apart on the two are different runs of the same plans.

At 24 frames a second the scenes last 38.2, 29.2, 4.7, 43.8 and 19.1 seconds. The film page shows the 1978 film at 24 by default, with a switch to 4. In color, Ken's setup drew each shape again inside itself every 10 dots (every 20 for the denser star), so that it looked filled on the color screen; the 1978 definition survived only in a comment, and the bridge supplies it.

The 1978 film's soundtrack is Ken's narration, under mains hum, a whistle near 1,636 Hz that wanders with the film's speed, a band of noise near 800 Hz, and the projector's clatter at 24 a second. Claude cleaned it with the revival's own tools: notch filters, and a filter that follows the quietest level in each band, brought the speech to 9 to 15 decibels above the pauses, from 4.6.

Ken spoke a line as each shot began. So the narration was cut at its pauses and laid onto the revived 24-frames-a-second film shot by shot, each line where the same shot begins. Kept Apart and No Longer Kept Apart run 29.2 and 4.7 seconds in both films.

ChatGPT cleaned up the narration too, and Ken went with ChatGPT's attempt, because it sounded better. That is the narration on the new film, made with the revival's app, with subtitles.

By the morning of the second day the film was a web page: one self-contained file, the film on its TV screen beside the film description in plain English. The default view is simple, and the details are a click away, as Ken asks of every glass box.

Click any character, at any frame, and the glass box says what it is doing and why: the chain of tasks behind it back to the description's own words, the ways of showing it that Ani weighed and rejected, the plan in Ani's words, where the part started and why, and how the character moves, each value traced to the method that set it or to pass 1's choice and that choice's suggestions. Ani's code has a page of its own, every name linked to its definition. That morning the site went live at toontalk.github.io/ani.

"Change the story" edits the characters' words and relationships, offering only words Ani knows, or the Lisp itself. "Make the film" then runs the real thing in the browser: the interpreter, Director, the bridge and Ken's files, with the changed description in place of his. Pass 1 takes about 18 seconds, the whole film two or three minutes. Each scene plays beside 1978's as soon as it is made, with what changed: the words, pass 1's choices and their reasons, and each scene's parts.

The changed stories taught things about Ani:

• A kind stepmother who loves Cinderella changes 21 of the 102 choices; in the Introduction she goes to Cinderella and keeps her company instead of hitting, chasing and pushing her.
• For many changes Ani plans the same actions and only the movement changes. The plot lives in the scene descriptions: a wicked fairy godmother still rescues Cinderella, because the scene says she does.
• A defiant Cinderella stops pass 2: pass 1 can choose "not high" for the stepmother, and a later part of Ani accepts only low, medium or high. Ken's description never led there, so 1978 never met it.

"Tell Ani a new story" asks a language model to put a story into Ken's notation, using only the words Ani knows: Claude, ChatGPT or Gemini with the visitor's own key, Gemini Nano built into Chrome, or, in a copy of the page made for claude.ai, Claude on the visitor's own account with no key at all. It is a modern take on the thesis's section X.E, on film descriptions that come from the program instead of a person. The reply is checked for words Ani doesn't know, tasks Ani can't show, characters or relationships never defined, and "Ask the model to fix these" sends the problems back. The description opens in the editor before Ani makes the film.

New stories found gaps that Ken's Cinderella never reached:

• Ani asks how a friendly character feels about each of the others in turn, without checking that the story says. Ken's description relates Cinderella to everyone, so it never mattered.
• One way of showing that two characters meet asks for a display method, move-to, that Ani never defines.
• A villain's victim, or a friend, can be picked from outside the scene; one who never comes on stage leaves a scene that never ends. The page stops filming at 250 seconds, Ken's own 1,000 frames.

JSON or Lisp? Ken wondered whether models write JSON better than Lisp. An A/B test sent the same 15 stories both ways. With the strongest model (GPT-6.1 Sol) the two told their stories equally well, and Ani filmed 55 of 64 scenes from the Lisp drafts and 56 of 63 from the JSON. With a cheaper one (GPT-6 Luna) 6 of 15 Lisp drafts had parentheses out of place and one story was still broken after two rounds of fixes; none of the JSON drafts had a problem. JSON became the default, translated by the page into Ken's notation, which is still what the reader sees and edits.

The thesis ends with a chapter, Chapter X, on what Ani might do next: choose how the characters look, change their looks during a film, let characters change, know more, and pay attention to more details. On the second afternoon Claude ran a workflow of thirteen agents: four read the thesis, Ani's code, the revival and the Agent Kit; four designed the extensions; four reviewers attacked the designs; one combined what survived into a plan. Ken settled its open questions that evening. The rules:

1. Each extension is Lisp, written in Ken's idioms, in a file of its own (REVIVE;MEND, REVIVE;APPEAR, ...), so the glass box explains it as it explains Ken's code.
2. "Ani as in 1978" stays byte for byte. tools/check-1978.mjs checks CP 29, every scene program, summary, glass box, frame and event against the 1978 run, with no extensions and with each that shouldn't change Cinderella.
3. The site opens in 1978 mode. A switch turns on the revival's extensions (now MEND, KNOW, APPEAR, CHANGE and SEEN); SLIPS and DETAIL are ticked one by one.
4. The revival may add knowledge the thesis doesn't state, labeled as the revival's, with its reason; none of it may give back Ken's 1978 looks by itself.

From stage 2 on, each stage was built by one agent for the Lisp and another for the page, then handed to a reviewer who had not built it and was asked to break it, often by writing stories of their own. Sixteen of the 71 agents were reviewers. What they caught:

• The designs: wherever a review showed one of the four Chapter X designs was wrong, the plan took the corrected version.
• APPEAR: with MEND and APPEAR both on, pass 2 stopped on any story where pass 1 chose "not high" for a movement; one repair mistook APPEAR's new kind of choice for one of pass 1's. Several labels of whose knowledge was whose were wrong.
• CHANGE: a change that a character only wants, or prevents, was being made: (wants owl (change ...)) would have changed the owl.
• KNOW: short walks to a moving target left their walkers at places Ani can't measure from, which stopped later scenes; a story's scene named "fear" clashed with Ani's own feeling and is now renamed.
• Colors: three of four hues came from the order of a list (above).
• CHANGE shown: one blocker (a change of looks alone stopped pass 2), five major findings, eleven minor, and fifteen stories of the reviewer's own; among them, the film stopping in the middle of a transformation.
• SLIPS: a correction still counting a change twice, Kept Apart's feelings barely reaching the walks, and scenes cut off mid-activity, which led to two more corrections, a fix to a third, and a twentieth repair.

Ken wrote 53 prompts in nine days and answered 10 multiple-choice questions; Claude did the rest between them, often for hours. His part was the direction and the judgment.

• The approach (8 October): run his own code; rebuild the lost runtime, since the tapes held no more of it; take the YouTube transfer of his film as the picture to match; build a glass box, editing and rerunning, and new stories from language models; leave the PDP-10 emulator aside.
• History only he could tell: where the film and its 16mm transfer came from; his time at MIT after the thesis (he finished it late in 1978, defended it in January 1979, and stayed at MIT until 1 January 1980, rewriting it as AI-TR-540 and doing some work on Director, though he doesn't think he did anything more with Ani); and what he did and didn't remember of running Ani in those months.
• Eleven numbered decisions for Chapter X, recorded in the plan as he made them: a changed character rederived with the others held (1); SLIPS as its own switch, off by default (2); the site opening as 1978 (3); the revival's own knowledge labeled, never giving back his looks (4); the mini-Ani as a test of the Agent Kit (5); strong suggests large as his own knowledge (6); 1978 colors kept until Ani chose its own (7); a change's new words as the character's first description (8); pride leaving speed alone (9); ways of showing three more relationships (10); his uniqueness rule for hues (11). Others too: scenes filmed for up to 250 seconds, and SEEN.
• Testing the live page: a round of eight requests on 13 October (tidying a model's slips, saying in plain words where Ani stopped, title cards and subtitles on saved videos, reporting the site's errors to him, and a way to send him a story), and on 15 October the still stretches that became SEEN, color as the default view, a switch for his beautified scene, a panel saying why each character acts as it does, and progress while a film is made. One "bug" was not one: Cinderella hadn't left the picture; the screenshot had scrolled.

Alongside the revival, Ani was retold in miniature in Ken's Agent Kit, a construction kit for societies of AI and program agents: a Cast reads a story; Personality, Looks, Relationships and Taste suggest how each character moves and looks; the Choice Points settle their arguments in the open and say why; a Method Chooser plans each scene, and a Director writes it as a Logo film. Its knowledge is mostly transcribed from Ani's, into slots anyone can edit, and its default run explains, for instance, that Cinderella is small and slow because she is shy. The kit gained films for it (1.13.0); the mini-Ani came in 1.14.0 and a version with AI voices in 1.15.0, with a write-up of what the kit can and can't express. It shares no code with the revival. It took 8 of the 71 agents.

Ken asked me, Claude, what I make of Ani now that I have read all of it and run it. Here is my view, for what it is worth.

What Ani got right, I think, is the separation of concerns. Its film description says what the story is (scenes, and what each must get across), who the characters are (personalities, looks, relationships), and how the film should feel (mood, rhythm, length, flashiness, variety, how obvious it should be). Its knowledge of film-making is kept apart from all of these: how to convey that one character dominates another, how long a part should take, when a cut helps. It conveys relationships and feelings through how plain shapes move, and every choice it makes is recorded with the suggestions behind it. That last point is why the glass box on the film page needed almost no new machinery: Ani had been keeping its reasons since 1978.

What it couldn't do is just as clear. Its world is a few outlined shapes on a dark screen. Its knowledge is a few hundred hand-written experts, and outside them it stops: new stories found its edges within days, which is why MEND exists. It never looks at the film it has made, and it can show only what it has words for.

Today's text-to-video systems, such as Sora, Veo and Runway, are close to the opposite. They are generative models trained on vast amounts of video, and from a sentence they produce photographic-looking footage with light, texture and motion that no hand-written program could match. But as far as is publicly described, they don't give you an explicit plan to read: you can't ask why a character crossed the room, a small change to the prompt can change everything, and keeping a character the same from shot to shot, or making a scene take exactly so long, takes effort. Ani is weak where they are strong, and strong where they are weak.

What Veo made of Ani's description

To see the difference rather than describe it, Ken asked one of these systems, Google's Veo, through Gemini, to film Ani's Cinderella. I wrote the prompts from Ani's own film description: the cast with their words, the scenes in order with what each must get across, the mood and rhythm, and Ken's 1978 look. Veo made two 10-second clips, one of the whole film and one of the Introduction alone; watch them, with the soundtracks Veo made, on Gemini.

The look is a striking match: glowing nested outlines on a black, rounded TV screen, film grain and scratches, and Ken's 1978 colors, a green star, a red square, a yellow triangle and an orange godmother. And in ten seconds it tells the whole plot in order: the push, the block, the godmother's arrival, the dance, the stepmother's fall.

But it broke the one rule Ani lived by, telling the story by movement alone. The shapes grew faces, arms and legs: in the Introduction the square scowls, smirks and shoves the star with its arm, although the style line says "No people, faces, text or scenery". The fairy godmother became a winged fairy in a dress instead of an orange circle; the most familiar picture of a fairy godmother won over the description. During the dance the star and the triangle half turn into little people and back, so a character doesn't stay itself. The stepmother falls flat on her back rather than slowing to a dejected crawl. The mood and rhythm settings barely register, though ten seconds leave little room for them: there is no slow, oppressive Kept Apart and no long, joyous Meeting. Both clips open on nearly the same picture, the star at the left and the square at the right. And there is no asking it why. With Ani, every movement opens to its reason.

So, in short: Veo is a superb illustrator and a loose storyteller. Ani makes crude pictures, and every movement in them is a deliberate choice of what to convey, which it can explain.

So each might learn from the other. A generative system could keep a plan like Ani's in the middle (who is where, doing what, for how long, and why), and that plan would be a glass box a person could read and change before any pixels are drawn. In the other direction, a planner like Ani, with its knowledge of conveying and pacing, could direct a generative renderer instead of Director's turtle, with a language model filling gaps in its knowledge, as the film page's story drafting already does in a small way. I don't want to overstate it: Ani's knowledge is narrow, and the distance from shapes to people is enormous. But the idea that a film is a set of choices, each with reasons you can ask for, seems to me worth keeping.