9 min read · October 21, 2026
For years, foundation models were optimized to produce pleasant conversational prose inside browser windows. The emergence of closed-loop self-scaffolding systems like Ornith-1.5 reveals a different trajectory: lean parameter footprints tuned directly for terminal harnesses, sparse tool use, and verifiable assertions.
Core Takeaway: When post-training couples task proposal, verification harnesses, and solution rollouts through closed-loop group relative policy optimization, models shed their conversational bloat. Lean 35B Mixture-of-Experts architectures activating only 3B parameters can match monolithic weights on complex software engineering benchmarks simply by honoring execution feedback over rhetorical flair.
8 min read · October 16, 2026
Routing every production evaluation through a heavyweight autoregressive model burns cluster budgets on redundant monologues. Here is how typed System One evaluators and confidence-routed cascades restore discipline to the inference critical path.
Core Takeaway: Most production evaluations do not require an autoregressive chain-of-thought monologue. By pairing a single-pass System One evaluator like Jev ($0.042 per million tokens read) with a confidence-routed cascade, architects can resolve clear decisions at 10x to 300x lower latency and reserve heavyweight System Two reasoning strictly for ambiguous edge cases.
7 min read · October 13, 2026
The era of throwing $100M compute clusters at monolithic dense models has hit an economic brick wall. Here is how modern Mixture-of-Experts and speculative decoding actually work under the hood.
Core Takeaway: Frontier performance is no longer won by brute-force parameter expansion, but by decoupling parameter knowledge (671B) from FLOP cost per token (37B active) through elegant routing mathematics.
5 min read · October 10, 2026
Why models boasting two-million-token context windows behave very differently in high-concurrency production than they do in isolated single-user demonstrations.
Core Takeaway: Context length is not primarily a compute problem; it is a real-estate problem on high-bandwidth memory. Understanding KV-cache arithmetic prevents costly production outages.