Is a $100B data bet the only way past GPT's skill ceiling

andrew ho departed openai after eight months to found a data company focused on high-value, missing skills like bioinformatics and lab evaluation, citing that current models like gpt-5.6 sol fail at these tasks 70% of the time. ho contends that simply scaling compute won't yield true generalization or reliability, forcing the industry to invest over $100b in specialized data acquisition. he is joined by skepticism from cambridge and google deepmind researchers who observe that new models are becoming more specialized rather than versatile, with capabilities in language and logic stagnating while coding improves. ho also criticized valuations of frontier labs as unsustainable given the rising costs of competing with efficient open-source models like qwen. this pivot implies that future ai advances depend less on parameter counts and more on curating expensive, domain-specific signal.