Is DOOM a Tensor? | LIVE165
Anthony Shaw explains what tensors are and why they matter for how ML models run, then connects that understanding to writing better prompts and benchmarking when using GitHub Copilot to optimize code.
Overview
This Microsoft Build 2026 session uses a playful “can DOOM run on a tensor?” thread to explain how machine learning models work under the hood, and why that mental model helps when you’re trying to get better results from GitHub Copilot.
What the session covers
Tensors and how models execute
- The session frames the idea that every model you use ultimately runs on tensors.
- It starts with an example model: Harrier Text Embedding Model.
ONNX and “running logic” in model graphs
- Discussion of ONNX being Turing complete and capable of running complex logic.
Where DOOM fits in (and why it’s a useful analogy)
- Reference that DOOM has run on Windows since 1995.
- An audience poll concludes that a CPU emulator approach is the correct framing.
- Discussion touches on Excel interpreting machine code as another example of “unexpected runtimes.”
Representing a CPU and memory as tensors
- Explanation of a setup where a RISC CPU and RAM are represented as tensors.
Scaling constraints
- Notes scaling-out challenges tied to DOOM being single-threaded.
Coaching AI agents: what changes when you understand the runtime
- Best practices for coaching AI agents, with emphasis on benchmarking so you can tell whether changes actually improved outcomes.
Speakers
- Anthony Shaw
- Burke Holland
Event context
- Microsoft Build 2026 (Broadcast Stage)
- Session: LIVE165