Similar to how some labs have done free stealth model releases. Trillions of coding tokens ought to improve performance of the downstream model. However, improving the base model architecture remains the bottleneck since you will not reach RSI with just that (assuming you ever can). And compute, lots of compute…