Give AI eyes and ears.
Grid images for vision, audio transcription for speech.
Compress entire videos into one image — ~600x cheaper than frame-by-frame.
Gemini 3 Flash + VAM-RGB Grid = Unprecedented Efficiency
"Why send what AI already understands?"
VAM Seek transmits causality, not just data.
The "Frame 7" Paradox
15 frames capture an egg breaking. Frame 7 is the decisive moment.
We delete it.
AI understands physics — if an egg is falling in Frame 1 and shattered in Frame 15, it broke in between.
Send intent and result. AI fills the gap.
The thumbnail grid humans use to navigate becomes AI's input. One image captures the entire timeline.
App generates an 8×6 grid (~1568×660px) from your video automatically.
Open AI Chat (Ctrl+Shift+A) and ask questions about your video content.
AI sees the grid, references timestamps. Click any timestamp to jump to that moment.
When uncertain, AI autonomously zooms to higher resolution and corrects itself. Protected by max-depth limit (2 zooms per session).
Grid image sent once. Follow-up questions don't resend. 90% cost reduction on conversations.
Zoom to specific time ranges for higher resolution analysis when needed.
AI responses include timestamps. Click to jump directly to that moment in the video.
Choose between Claude (Anthropic) and Gemini (Google). Gemini supports video upload or grid mode.
Context-aware system prompts reduce hallucination and improve accuracy.
Primes AI with video metadata before questions for better accuracy.
Gemini-powered full video transcription with clickable timestamps. Ask about speech content.
AI learns from your corrections and improves over time. Rules persist across sessions.
VAM Seek encodes time into color, so that difference becomes visible. A still object stays gray; a moving one leaves a colored fringe. The same idea can be pointed at a model's own output: comparing the tokens it nearly produced with the ones it actually emitted.
When a model generates text it assigns probability to many candidate tokens but emits only one. The near-misses don't disappear — they remain in the output distribution. Looking at the gap between the chosen token and the ones left behind shows where the model was confident and where it hesitated, much as a moving object reveals itself as a fringe in a VAM-encoded frame.
Note: this is a difference measured over a single, observable distribution. It is not a claim about any hidden or "true" state the model conceals, and it is not a measure of intent.
In exploratory runs we saw output distributions collapse sharply in some contexts and stay broad in others. Loosely, this resembles the difference between speaking with certainty and holding back when unsure — the kind of gap that, in a person, can carry a trace of character.
Important: these are preliminary, un-replicated observations. They describe distributional patterns, not intentions. Establishing anything beyond "the distribution differs" would require controlled experiments — matched conditions, bias-blind probing, and multiple models. We flag this as an open question, not a result.
Separately from the observations above, we offer a view rather than a claim: that the difference a system holds back is worth observing rather than discarding. We are curious what fringes appear when a model is given room, and we share this as an open note to others exploring the same question.
Clone the repo, add your Claude or Gemini API key, and start analyzing videos.