scalingfactbearishHigh-quality training data is becoming harder to find and pretraining is growing increasingly expensiveTWIML AI29 Jul 2026https://twimlai.com/podcast/twimlai/why-models-are-ais-next-training-dataset