One model, cast into its own traces.
why one chip is so much faster
Drag to grow the model
smalllarger still
The model is the computer, in one picture.
01 · The box
AI in schools has meant a box you open, prompt, and wait on.
02 · The world
Runway's Solaris generates the interface itself, frame by frame, as you act on it.
Runway, August 31, 2026 →03 · The die
Taalas casts one model into its own silicon. Up to 17,000 tokens a second, by the company's figure. AMD acquired it in August.
Taalas →04 · The learner
A learning environment that draws the next step ahead of the student, in their own context, before the thought has cooled.
Read the signal“Tutorials no longer replay the same sequence for everyone. Instead, they render the next step in your own context, adapt as you make progress and recover naturally when you go off script.” Runway, Introducing Solaris, August 31, 2026
Acquired by AMD, August 2026
“Interactions stop feeling interactive somewhere around half a second of delay.” Runway, on why real time was the hard part
Clips: Runway, Introducing Solaris.
No screen existed here before you touched it.
Drag to grow the model
The model is the computer, in one picture.
Drag through the minute
Today 01Stop the task. Next 01The environment notices where the student is.
Touch the water
Under half a second. The answer arrives with your touch.
Touch a ripple
These are the questions to ask before the demo, not after it.
What is drawn, and what is cited
How it works. Schematic. Runway describes the second loop as treating “user input as conditioning for the next frame,” with a language model deciding what should happen and the world model drawing it.
Why one chip is so much faster. Schematic, not a wiring diagram. The round-trip counts are illustrative, chosen to show that the number climbs on one side and stays at zero on the other. The 17,000 tokens a second is Taalas’s published figure for the HC1 running Llama 3.1 8B. Taalas trades flexibility for the speed: the chip runs the one model it was built around.
The half second. The right end of the arc lands on Taalas’s published figure for the HC1 running Llama 3.1 8B. The left end, 40 tokens a second, is illustrative, not a measurement of any product. The wait is 120 tokens divided by the speed you chose, and 120 tokens is an illustrative length for a short explanation. The half-second line is Runway’s own threshold from the Solaris announcement. The koi moving on is a dramatization of that threshold, not a measurement of student attention.
One minute of the same student. Illustrative. One minute of one student, drawn to show the shape of the two paths, not a measured study. The five exits are the five steps drawn above, not a research finding.
The questions first. The lines under Accuracy and Text quote Runway. Accessibility and student privacy carry no quote, because none exists.
Figures are the companies’ own published claims and the independent measurements linked above. Karst has not tested either system. Videos belong to Runway and Taalas, are shown from their own servers for commentary, and Karst has no relationship with either company.