Layer researchIntelligence and learning
Vendor landscape
As of 2026-08-19. Vendor capability claims are vendor-published unless third-party sourced. Re-verify quarterly.
- Vendor / product
- NVIDIA Data Flywheel Blueprint
- Category
- Reference flywheel
- Maturity (Aug 2026)
- Published Jun 2025
- Notes
- The only fully documented enterprise flywheel with numbers; promotion stage is manual by design [vendor]
- Vendor / product
- Databricks (Agent Bricks, Judge Builder; Quotient AI acquisition)
- Category
- Platform flywheel
- Maturity (Aug 2026)
- Acquisition Mar 2026
- Notes
- Judge building aimed at domain experts without code [vendor]
- Vendor / product
- Snowflake Cortex Agents; Microsoft Fabric; Google agent quality tooling
- Category
- Platform
- Maturity (Aug 2026)
- GA waves 2025-2026
- Notes
- Optimizer-evaluator decoupling is the published architectural rule [vendor]
- Vendor / product
- OpenAI reinforcement fine-tuning; on-policy distillation methods
- Category
- Model adaptation
- Maturity (Aug 2026)
- RFT GA May 2025
- Notes
- Requires a programmable grader; wrong tool for style [vendor]
- Vendor / product
- GEPA, ACE (research methods)
- Category
- Prompt and context optimization
- Maturity (Aug 2026)
- ICLR 2026
- Notes
- The evidence that the cheap lever is rarely exhausted
Agent-washing watch: "self-improving" agents whose improvement path is unlogged; optimization services that will not disclose the grader; benchmark claims without harness isolation.
Source: research/R06-intelligence-and-learning/vendors.md in the evidence repository behind this site.