I suspect generalized reasoning was solved just a few weeks ago and it flew completely under the radar. HRM, a new arch, reportedly has SOTA results on ARC-AGI 1 & 2 benchmarks with only 27 million parameters and ~1k training examples.
164,06K