01
Where today's AI actually breaks
Deep learning fails in the same three ways whenever it leaves the lab. It breaks under distribution shift, it cannot recombine what it has learned into something new, and its errors compound over long horizons. I mapped this across 87 papers and 33 benchmarks on agents that operate computers, and found six gaps that scale alone does not close. Two further studies back the diagnosis from other directions. The compression methods that make large language models affordable measurably damage the reasoning they are meant to preserve, and even architectures built for compositionality fail on rule combinations they have not seen.
