Thinking Machines ships Inkling, its first open-weight model [Xtra!]A sparse 41B-active MoE aims to match dense giants at a fraction of inference cost.
Model routing gets harder as the model zoo grows [Xtra!]Naive benchmark-based routing breaks down once cost and latency enter the equation.
Vercel opens up its AI Gateway leaderboard data [Xtra!]Real usage-weighted model comparisons, not just vendor-published benchmarks.
xAI open-sources Grok Build coding agent [Xtra!]Another frontier lab ships its coding-agent scaffolding as open source.
Perplexity details secure sandboxing for its agents [Xtra!]Isolation design matters more as agents get direct code/tool execution rights.
Harness engineering claims early recursive self-improvement [Xtra!]Optimizing the scaffolding around a model, not the model itself, is becoming its own discipline.