Actually I disagree with this. Inference costs are plummeting due to architectural improvements. Tool harnesses are getting a whole lot better in general. There’s tons of unpicked fruit there. But my argument is that pure parameter scaling is leading to diminishing returns in absolute performance.