Frontier Lab Readiness MOC
This is the umbrella map for the subjects adjacent to Frontier LLM Architectures MOC. The architecture sphere is the trunk. These spheres are the branches that make you useful in a real frontier-lab environment.
Start Here
Read in this order:
- Frontier LLM Architectures MOC
- ML Systems for Frontier Models MOC
- Evaluation and Benchmarking MOC
- Mechanistic Interpretability MOC
- Agents and Tool Use MOC
- Multimodal Foundation Models MOC
- AI Safety and Security MOC
- Research Engineering Practices
Why These Matter
| Sphere | Why frontier labs care |
|---|---|
| ML Systems for Frontier Models MOC | Models are constrained by hardware, memory, distributed training, inference latency, and cost. |
| Mechanistic Interpretability MOC | Labs need to understand what models represent and how behavior emerges or fails. |
| Evaluation and Benchmarking MOC | Bad evals create fake progress; strong evals drive real research. |
| Agents and Tool Use MOC | Models increasingly act through tools, code, browsers, and long-horizon environments. |
| Multimodal Foundation Models MOC | Frontier models are no longer text-only; vision, audio, video, and action matter. |
| AI Safety and Security MOC | Capability without reliability, misuse resistance, and monitoring is not deployable. |
| Research Engineering Practices | Clean experiments, logging, ablations, and reproducibility are how research becomes knowledge. |
The Frontier-Lab Skill Shape
You want a T-shape:
- Deep specialty: one area where you can contribute real experiments.
- Broad literacy: enough adjacent knowledge to know when your result is actually data, systems, eval, or post-training.
Good specialty choices:
- Architecture and training.
- Inference systems.
- Mechanistic interpretability.
- Agent evaluations.
- Multimodal learning.
- Safety/security evals.
The Minimum Serious Portfolio
Build at least four artifacts:
- A tiny transformer implementation with KV cache and profiling.
- A scaling or systems benchmark with clean plots.
- A paper reproduction or critical replication attempt.
- A research memo series using Research Memo Template for LLM Papers.
Stronger portfolio:
- An MoE toy model.
- A long-context evaluation suite.
- A mechanistic interpretability replication.
- A tool-use agent benchmark.
- A multimodal prototype.
- A safety/security red-team evaluation.
Weekly Study Rhythm
- 2 days: read papers and notes.
- 2 days: implement or reproduce.
- 1 day: evaluate, plot, inspect failures.
- 1 day: write a memo.
- 1 day: review and choose next experiment.
North Star
Aim to become the person who can say:
I understand the mechanism, I can implement a toy version, I know the systems bottleneck, I can evaluate it honestly, and I can explain what result would change my mind.
That is the research posture.