Blog
June 18, 2026 • ARC-M Team
Finite-State Models for Continual Learning
The companion to 'The Data Wall.' We argue that the architecture that learns from the world's operational stream needs two properties transformers structurally lack: a bounded internal state and an online update rule. We connect three live research threads — structured state-space models, learning at test time, and the plasticity/forgetting problem — into a single bet: finite-sized state models that keep learning after deployment without catastrophically forgetting.
Read moreJune 10, 2026 • ARC-M Team
The Data Wall
The transformer's superpower was eating the internet — and it has now essentially eaten it. We walk through the scaling-law evidence that the binding constraint has shifted from compute to data, why the usual escape hatches (repeat, synthesize, scrape harder) only buy time, and why this is an architectural problem rather than a supply problem.
Read more