Working notes from our research areas — written for practitioners shipping AI systems, not for leaderboards.
Self-evolving harnesses and agentic research that let AI systems improve their own tools, workflows, and reasoning loops.
Jul 22, 2026 · 7 min
The scaffolding around a model is now the fastest-moving part of the system. What happens when the scaffolding starts editing itself?
Read Harnesses That Rewrite ThemselvesJul 10, 2026 · 6 min
An agent that proposes, runs, and grades its own ablations changes the unit economics of research — if you get the verifier right.
Read Agentic Research Loops: Letting Systems Run Their Own ExperimentsJun 28, 2026 · 5 min
Routing is a control problem. The harness should decide which model does which step — and what that step is worth.
Read The Cost-Aware Agent LoopFurther reading
World models and synthetic data generation that connect intelligence to sensors, robotics, and the physical world.
Jul 18, 2026 · 8 min
A learned world model is useful long before it is accurate. The trick is knowing which errors you are allowed to tolerate.
Read World Models as Simulators, Not OraclesJul 2, 2026 · 6 min
Generated data is cheap. Generated data that improves a real deployment is not. Here is the filtering pipeline we use.
Read Synthetic Data That Survives Contact With RealityJun 15, 2026 · 5 min
Embodied systems do not have a token budget problem. They have a perception bandwidth problem, and it is architectural.
Read Sensors Are the Real Context WindowFurther reading
Indic models and multimodal research built for local languages, cultures, and real-world use cases.
Jul 20, 2026 · 7 min
Serving a billion speakers is not a translation problem. It is a tokenization, data, and evaluation problem — in that order.
Read Indic Models Beyond TranslationJul 5, 2026 · 6 min
Most real enterprise documents are not text. They are scans, forms, tables, handwriting, and photographs of screens.
Read Multimodal Research for the Messy MiddleJun 20, 2026 · 5 min
A 4B model that knows your domain beats a frontier model that does not — at a hundredth of the cost.
Read Small Models, Specific JobsCost reduction of inference through quantization, distillation, and performance-tuned serving stacks.
Jul 25, 2026 · 6 min
In agent workloads, most tokens are recomputations of tokens you already paid for. Caching is no longer a systems detail.
Read The KV Cache Is a P&L LineJul 12, 2026 · 7 min
Aggregate benchmarks hide quantization damage. The losses concentrate in exactly the capabilities you deployed the model for.
Read Quantization Without RegretJun 30, 2026 · 6 min
Two operators, identical GPUs, identical model — and an order of magnitude difference in cost per token.
Read The Serving Layer Is Where the 10× HidesFurther reading