Did Pathway just reveal the architecture breakthrough Andrew Curran predicted? Its 150M model sets a new ARC-AGI-1 cost-efficiency frontier

Andrew Curran recently hinted at a major architectural breakthrough in memory efficiency, coming not from a big AI lab but from a team with ties to OpenAI: https://x.com/AndrewCurran_/status/2072076893730349409

Pathway has now announced BDH-CQ: a 150M-parameter post-Transformer model that scored 29.5% on ARC-AGI-1 at a computed cost of just $0.0007 per task, establishing a new cost-efficiency frontier. It uses recurrent memory and latent reasoning instead of long token-based chains of thought.

The connection is surprisingly close: a new memory-efficient architecture, developed outside the major labs, with OpenAI researcher and Transformer co-author Lukasz Kaiser as an investor and adviser.