TechnologyFormer OpenAI research VP says simple failures reveal AI models' real limits
Jerry Tworek argues that gains in math and coding benchmarks do not solve weak performance on messy, context-heavy work and larger system decisions.
TLDR
Jerry Tworek, a former OpenAI research vice president and now co-founder and CEO of Core Automation, says AI models’ biggest limitation is not their peak performance but their failure on tasks that can appear simple. Speaking at The Information’s AI Agenda Live event, he contrasted large gains on well-defined math and coding problems with continued weakness on messy, context-heavy work and larger system-level decisions. His argument is that smarter models alone may not produce reliable systems unless their architectures and learning methods also improve.
Former OpenAI research VP says simple failures reveal AI models' real limits
Jerry Tworek argues that gains in math and coding benchmarks do not solve weak performance on messy, context-heavy work and larger system decisions.
TLDR
Jerry Tworek, a former OpenAI research vice president and now co-founder and CEO of Core Automation, says AI models’ biggest limitation is not their peak performance but their failure on tasks that can appear simple. Speaking at The Information’s AI Agenda Live event, he contrasted large gains on well-defined math and coding problems with continued weakness on messy, context-heavy work and larger system-level decisions. His argument is that smarter models alone may not produce reliable systems unless their architectures and learning methods also improve.
