Gravit

Curta-metragem · 1 min 18 s · com som

O Próximo Token

Nove capítulos, do treinamento — com o erro que volta pela rede e ajusta os pesos — à primeira palavra da resposta, em 78 segundos — com os efeitos sonoros gerados pela própria IA. Logo abaixo, a animação explica cada passo em detalhe.

Como o filme foi feito
  1. Nove clipes de 8 segundos, gerados pelo Veo 3.1 (Google) em 1080p. Os efeitos sonoros são gerados pelo próprio modelo, junto com a imagem.
  2. Cada clipe começa exatamente no último quadro do anterior. É isso que faz os nove parecerem um único plano contínuo.
  3. O texto não é gerado pela IA — modelos de vídeo escrevem letras ilegíveis. As legendas são desenhadas por cima, em três idiomas.
  4. A imagem é uma metáfora: um transformer de verdade não tem partículas nem luz. Mas a ordem das coisas é a real — treinamento, tokens, vetores, atenção e um token de cada vez.

Os prompts de cada capítulo

Em inglês, do jeito que foram usados no Veo 3.1. Funcionam no Google Flow: gere o capítulo 1 e use o último quadro de cada capítulo como primeiro quadro do seguinte.

01 A chuva de experiência
A slow, continuous vertical descent through an endless dark indigo void. At first there is near-total darkness and only a few faint sparks. Then a gentle rain begins: countless tiny luminous shards fall past the camera — fragments of experience rendered as slivers of light, some holding blurred glimpses of landscapes and skies, some shaped like ripples of sound waves, some like streaks of motion or flickers of human expression, always abstract. The shards thicken into flowing streams. As they fall, faint threads of light begin to connect related fragments to one another. In the final seconds, far below, the first vast glowing network comes into view, like a sprawling city of light seen from high above. Camera: steady, slow downward crane movement looking slightly down, shallow depth of field, shards passing close to the lens as soft bokeh.
Audio — Ambient: silence slowly filling with a deep sub-bass drone. SFX: soft crystalline tinkles and a glassy patter as the shards fall, like rain on glass; tiny sparkling pings whenever a thread of light connects two fragments; a distant low rumble as the network appears far below. No music, no voices.
02 Adivinhar, comparar, corrigir
The camera holds above the vast glowing network as the gentle rain of light fragments continues. One luminous fragment drops into the network and races forward through its layers of light, and at the far side the network produces a guess: a small glowing shape. Right beside it, the true next fragment arrives. They do not match — the gap between the guess and the truth flares as a thin, bright red-amber pulse of error. That error pulse then travels backwards through the very same layers, retracing the path; wherever it passes, the connections it touches shift slightly — some brighten and thicken, others dim and thin — like thousands of tiny dials being turned. The loop repeats, faster and faster: guess, compare, the error flows back, the connections adjust; with each loop the guesses land closer to the truth and the error pulses grow smaller and fainter. By the end the rain and the network return to exactly the same calm view as at the start.
Audio — Ambient: the deep drone and the glassy rain continue. SFX: a quick rising whoosh as the fragment races forward; a soft discordant buzz when the guess misses; a reversed, descending whoosh as the error pulse travels backwards; tiny clicks and ticks as the connections adjust, speeding up with each loop; the buzz softening into a clean chime as the guesses land true. No music, no voices.
03 Um modelo do mundo
The descent continues down into the vast glowing network below — an immense, layered neural landscape of luminous nodes and fine filaments stretching to the horizon, layer beneath layer. Streams of light fragments pour into it and flow through the layers. Where patterns repeat, pathways flare brighter and thicken; unused filaments dim and dissolve into darkness. Gradually the chaos organizes itself: clusters crystallize into elegant geometric structures — lattices, spirals and nested shells interlocking — a compressed, coherent map of how the world fits together, never showing literal objects. The change is gradual and organic, like frost spreading across glass or a galaxy settling into shape. By the end the network is calm, dense and coherent, glowing steadily, and the camera eases into a slow pull-back that reveals it as a quiet luminous universe.
Audio — Ambient: an evolving low drone and a soft rushing flow like wind through glass. SFX: electric crackles and bright zaps as pathways flare, a soft fizzing as unused filaments dissolve, deep resonant crystalline chimes as each geometric structure locks into place, settling into a calm hum. No music, no voices.
04 Chega uma pergunta
Silence. The trained world model floats in deep indigo space: a calm, immense luminous structure of cyan, violet and amber filaments, softly breathing with light. From the darkness beyond its edge, a single narrow beam of warm white light arrives and pierces the outer boundary with a soft ripple. As it crosses, the beam breaks into a short procession of about seven distinct glowing objects — small three-dimensional crystalline tokens, each with its own unique internal pattern — that travel together into the model with a rhythmic cadence, the last one lifting slightly, like the rising intonation of a question. The camera tracks alongside the procession and closes in on one token. No letters or symbols anywhere.
Audio — Ambient: near silence, only the faint hum of the model. SFX: a soft airy whoosh as the beam arrives, a deep watery ripple as it pierces the boundary, a bright glassy ring as it splits into tokens, then a gentle rhythmic sequence of tones as the tokens travel, the last one rising in pitch like a question. No music, no voices.
05 As palavras ganham significado
Close tracking shot of a single glowing crystalline token drifting through dark indigo space. It unfolds and expands into an intricate multidimensional form: a luminous structure of hundreds of fine rays and nested translucent planes, each ray with a slightly different length, colour and angle — a crystal made of light that encodes meaning. The neighbouring tokens in the row unfold in the same way, each into a different shape. A soft wave of light travels along the row from the first token to the last, subtly tinting each one differently, giving every token a sense of its place in the order. The camera slowly orbits the complete constellation of unfolded tokens, revealing their relationships in depth, and at the very end the whole constellation holds perfectly still for an instant.
Audio — Ambient: a low drone. SFX: crystalline blooming sounds, like ice crystals growing, as each token unfolds; a soft rolling whoosh-pulse as the wave of light passes along the row; a sudden suspended hush at the final frozen instant. No music, no voices.
06 Consulta, chave e valor
Macro view inside the constellation of glowing tokens. From every token emerge three distinct, elegant streams of light: a thin searching beam of cyan light that reaches outward like a probe; a shimmering amber pattern on its surface that shows what it can be matched with; and a soft violet stream carrying luminous content. The cyan probes sweep across all the other tokens and test themselves against their amber patterns. Where a probe and a pattern align, the link snaps into brightness and violet light flows along it back toward the searching token; where they do not match, the link stays faint and fades away. Dozens of these comparisons happen at once, precise and mathematical. The camera glides between the tokens at microscopic scale, following one cyan probe to its strongest match. At the end, the bright links multiply and split apart into several parallel translucent layers.
Audio — Ambient: a warm low hum. SFX: soft sonar-like scanning pings as the cyan probes sweep; precise glassy clicks and a bright lock-in tone when a probe matches a pattern; a warm liquid flowing sound as violet light travels along a link; faint fizzles as mismatched links fade. No music, no voices.
07 Várias formas de olhar
The same row of glowing tokens is examined simultaneously in several parallel translucent layers of light, stacked one behind another in depth like sheets of luminous mist — not screens, not panels. In each layer the tokens are joined by a different pattern of glowing arcs: in one, short arcs link neighbouring tokens; in another, long arcs connect distant tokens; in another, a strong bundle of arcs converges on a single token; in another, arcs reach outward to the vast world-model structures in the background. Each layer has its own subtle tint. The camera moves slowly along the stack, then the layers slide together and merge, their patterns weaving into one coherent, richer luminous structure that glows brighter than any single layer. Precise and mathematical, not magical.
Audio — Ambient: an airy shimmer. SFX: several layered harmonic tones, one per layer, each with a different timbre; soft glassy sliding sounds as the layers move together; the tones converge into one rich resonant tone with a deep soft whoomp as the layers merge. No music, no voices.
08 Um token de cada vez
The glowing enriched structure at the centre releases a compact, bright beam of warm white light that travels forward into darkness. Around it, the vast learned world model appears as a dim, immense web of filaments; only the regions near the beam light up while everything else stays dark. At the tip of the beam, five to seven thin branching paths of light fan out like a fork in a road, each with a different brightness, flickering as their strengths shift. One branch grows clearly brighter than the others; a small glowing crystal token forms at its tip while the dimmer branches fade away. Then new branches fan out from the new token, one wins again and another token forms — three or four times, in a steady rhythm — building a short, luminous chain of tokens connected by light. The camera tracks steadily alongside the growing chain.
Audio — Ambient: a gently swelling drone. SFX: a soft flickering shimmer as the branches fan out; a clear bell-like chime and a soft pulse each time a new token crystallizes, repeating in rhythm; faint fizzles as the losing branches fade. No music, no voices.
09 A jornada inteira
The growing chain of tokens becomes a calm, coherent stream of light and flows out of the world model into open indigo space. The camera pulls far back in one smooth, continuous move to reveal the entire journey in a single frame: a gentle rain of light fragments falling at the top, the vast luminous world model in the middle, and the calm stream of the answer flowing out of it — all connected in one continuous system. The motion slows and settles into a serene, balanced composition.
Audio — Ambient: the drone swells gently. SFX: a smooth flowing whoosh as the stream leaves the model; a soft wind-like rush during the long pull-back; everything resolves into a warm sustained tone that fades into silence. No music, no voices.
+ Estilo comum, adicionado a todos
Style: cinematic scientific visualization of the inside of an artificial intelligence — photorealistic yet poetic, immense scale, deep indigo-black space, luminous particles and filaments in cyan, amber and violet, volumetric light, subtle depth of field, slow deliberate camera movement, physically plausible motion, elegant and sophisticated. One consistent visual universe, the same materials and colour palette throughout.
− Prompt negativo
text, letters, words, numbers, typography, captions, subtitles, watermark, logo, signature, user interface, screens, monitors, holographic interface panels, charts, robots, androids, humanoid figures, people, human faces, human brain, anatomical brain, cartoon, anime, low resolution, blurry, jittery camera, flicker