AI Briefing
KOSign in

TVM and Together Compiler explore divergent paths for agentic kernel optimization and verification

·2026.11.10 09:20

Key point

TVM and Together Compiler demonstrate that stable IRs and verification tools are necessary to constrain search costs and ensure correctness when AI agents optimize inference kernels.

Details

  • TVM and Together Compiler argue that while agents can handle low-level details, a minimal, stable ML compiler foundation is required to provide shared abstractions and reusable verification tools.
  • TVM formalized its IR in Lean to enable cheap verification of memory access and synchronization, treating kernel optimization as a search process where tokens and tool calls incur costs.
  • Together Compiler found that unrestricted source code access allows agents to reach a higher performance ceiling than traditional IRs, though agents spent approximately 20% of their time reading engine files to understand code paths.
  • Together developed a custom cost model to rank parallelism configurations analytically before expensive agentic optimization, reducing the search space.
  • OpenAI used Codex to remove heuristics like automatic FMA from the Triton compiler to create a smoother search landscape, while TVM's TIRx avoids heuristic-based optimizations to allow agents to iterate new capabilities.
  • Providing human-engineered tools for research and logging impeded agent throughput, with agents spending up to 43% of tool calls on understanding systems rather than optimizing.
inferencekernel-optimizationai-agentsml-compilerstvmtogether-compiler
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditAnthropic releases Fable 5.2 and Mythos 5.2 with 30% speed boost and 20% price cutOpenAI BlogOpenAI Registers Business Ownership in ChatGPT and Launches Agent Features in BetaLangChain BlogLangChain Argues Goal Engineering Will Replace Prompt Engineering for Autonomous AgentsRedditUnit 42 Reports Chinese Attackers Used DeepSeek and GLM Agents to Breach 10 OrganizationsPyTorchKRllama.cpp v1.0.0, First Major Release in Three and a Half YearsHacker Newsggml v1.0.0 Released with FP8 Support and Backend Optimizations