Two Minute Papers breaks down how Claude's latest quirks expose AI's core flaw: not raw intelligence, but brittle reasoning that collapses under edge cases. This isn't just another benchmark fail—it's why tools we trust for research or code still hallucinate with quiet confidence. Scaling won't save us; we need architectures that actually verify their own logic. What limitation have you hit hardest when using these models? Full video link https://youtu.be/axOcn--n_lM