What Comes After Transformers?

Pathway claims a 150M-parameter post-transformer model matches larger LLMs on ARC-AGI at far lower inference cost.

• Post-transformer architecture: Pathway's BDH ("Dragon Hatchling") model, under 1 billion params (150M used in benchmarks), formally links transformer attention to Hebbian/synaptic-plasticity learning in the brain via four core equations.
• Latent reasoning beats chain-of-thought on cost: On ARC-AGI-1, Pathway's 150M-parameter model hit ~30% accuracy at under one-tenth of a cent per task, above DeepSeek and near Llama-level cost/accuracy at launch.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.