TTT-Discover optimizes GPU kernels 2x faster than human experts — by training during inference

TTT-Discover is a breakthrough AI optimization platform that enables enterprises to solve complex, high-value problems by allowing models to train during inference. Developed by researchers from Stanford, Nvidia, and Together AI, the technology challenges traditional 'frozen' AI models by implementing test-time training that updates model weights in real-time as they attempt to solve specific problems. The system uses entropic objectives and PUCT search algorithms to aggressively hunt for optimal solutions rather than average ones. With proven results including optimizing GPU kernels to run 2x faster than human-written code, TTT-Discover targets 'million-dollar problems' where incremental improvements yield substantial financial returns. The platform works with open-source models and can be deployed within enterprise VPCs, making it suitable for sensitive proprietary optimization challenges.

TTT-Discover positions as a specialized AI optimization platform for enterprises facing high-value, verifiable optimization challenges where traditional methods have plateaued. It serves as an automated R&D lab for specific problem domains rather than a general-purpose AI tool. The platform targets 'static, high-value assets'—low-frequency but high-impact decisions where spending hundreds to thousands of dollars on optimization yields immediate ROI through compute savings, operational efficiencies, or breakthrough discoveries. Key differentiators include its ability to work with open models within secure environments, its focus on continuous reward signals rather than binary outcomes, and its fundamental shift from generalist AI to problem-specific adaptation.

• Test-time training during inference that updates model weights in real-time
• Entropic objective function that exponentially weights high-reward outcomes, encouraging exploration of 'eureka' solutions
• PUCT tree-search algorithm for systematic exploration of solution paths
• Integration with existing reinforcement learning infrastructure (GPUs, rollout workers, optimizers)
• Support for open-source models like gpt-oss-120b, enabling private deployment
• Orchestration via Tinker API for distributed training/inference management
• Continuous reward signal optimization (runtime, error rate, cost metrics)
• Domain-agnostic architecture applicable to systems engineering, algorithm design, biology, and mathematics

Primary competition includes: 1) Human expert consultants and optimization specialists who manually solve complex problems, 2) Traditional reinforcement learning platforms that optimize for average performance rather than outlier solutions, 3) Conventional AI optimization tools using frozen models without test-time adaptation, 4) Specialized optimization software for specific domains (e.g., logistics routing tools). TTT-Discover differentiates through its test-time training paradigm, entropic objective function, and ability to achieve breakthroughs on problems that have resisted conventional approaches. The platform's use of open models and private deployment also contrasts with closed AI optimization services that require data sharing with third parties.

B2B enterprise model with multiple revenue streams: 1) Per-problem pricing based on compute resources used (approximately $500 per discovery run as demonstrated in research), 2) Annual subscription for unlimited optimization runs within compute limits, 3) Value-based pricing capturing a percentage of customer savings (e.g., 10-20% of first-year compute cost reductions), 4) Professional services for implementation, integration, and custom problem specification. Additional revenue from managed services offering turnkey optimization labs and training/consultation on identifying 'million-dollar problems.' Initial target customers are enterprises with existing RL infrastructure to minimize adoption friction, expanding to broader market as tooling matures.