Toda Labs
Engineering notes, product updates and research from Toda Labs.
Launch
Some avatar models now top 18 billion parameters, drawn in a data centre for every minute they talk. Ours is light enough to draw the face on your visitor’s own device: long talks, busy hours, no per-minute fee for the face. One portrait is all it takes to create one.
Read articleOff the Grid · Part 4
Text, a recorded video and a live avatar can explain the same refund process. The difference is what has already been made, what must happen now, and where that work runs.
Read articleOff the Grid · Part 3
Eight questions about capacity to put to any AI avatar provider before you sign, from what each conversation holds to where the GPUs come from, with our own answers.
Read articleOff the Grid · Part 2
A live avatar's GPU time doesn't stay in the data centre. It turns up on the invoice, by the minute. We followed one support desk and one software product through a month of conversations.
Read articleOff the Grid · Part 1
A live avatar asks a data centre for a GPU for every second of a conversation. What that holds in power, water and grid capacity, and what changes when the face renders on the device.
Read articleEngineering note
Why we chose browser animation, and what that means for longer conversations, repeated use and avatar-rendering costs.
Read article