Skills
What frontier labs hire for
339 recorded research postings · Partial update · October 8, 2026
This site is organized by the skills labs ask for in their own job postings. Each skill shows how often labs ask for it and everything here that teaches it: courses, book chapters, articles and notes.
Measure models
Evals, statistics and experiment design: whether a number means what it seems to.
-
Evals and benchmarks
Design an evaluation, run it, and say what the number means: graders, benchmarks, autoraters and error bars.
In 65% of research roles · 74 pieces here
-
Running experiments
Turn a vague behavior into a hypothesis, a pipeline and a result: ablations, baselines and research taste.
In 20% of research roles · 7 pieces here
-
Statistics and experimental design
Intervals, paired comparisons, sample sizes and controls: whether a difference is real.
In 17% of research roles · 9 pieces here
Train models
How language models work and how they are trained: internals, post-training, RL, interpretability.
-
How language models work
Tokens, attention, transformers and what a forward pass computes.
In 67% of research roles · 3 pieces here
-
Post-training and fine-tuning
SFT, preference optimization, RLHF and the data that drives them.
In 50% of research roles · 2 pieces here
-
Reinforcement learning
Policies, rewards and the algorithms that train reasoning and agentic models.
In 45% of research roles · 1 piece here
-
Multimodal models
Vision, audio, video and models that combine them.
In 32% of research roles · Not taught here
-
Pretraining and scaling laws
Data mixtures, compute budgets and the curves that predict loss before the big run.
In 22% of research roles · Course coming soon
-
RL environments and verifiers
Build the tasks, graders and verifiers a model is trained against, and keep it from gaming them.
In 11% of research roles · 1 piece here
-
Alignment research
Oversight, honesty and making a model's goals match ours.
In 10% of research roles · Not taught here
-
Interpretability
Look inside a model: probes, features, sparse autoencoders and circuits.
In 6% of research roles · Course coming soon
Build systems around models
Agents, retrieval, coding agents and security: systems that act, and how to make them hold up.
-
Agents and tool use
The loop, typed tools, memory and the boundaries that make an agent reliable.
In 44% of research roles · 79 pieces here
-
Safety, red-teaming and misuse
Prompt injection, jailbreaks, adversarial testing and the boundaries that hold.
In 10% of research roles · 6 pieces here
-
Retrieval and RAG
Embeddings, search and finding which stage of a pipeline failed.
In 5% of research roles · 7 pieces here
-
Working with coding agents
Claude Code, Codex and Cursor on real work, with acceptance the agent cannot redefine.
In 5% of research roles · 69 pieces here
Make it fast and big
Inference, PyTorch, JAX, distributed training and kernels: latency, throughput and cost.
-
PyTorch
Tensors, autograd and training loops in the framework most labs train with.
In 42% of research roles · Course coming soon
-
Inference and serving
Latency, throughput, batching, caching and cost per token.
In 37% of research roles · 15 pieces here
-
JAX
Functional transforms, jit and sharding on TPUs and GPUs.
In 27% of research roles · Course coming soon
-
Distributed training
Data, tensor and pipeline parallelism: one model trained across many accelerators.
In 24% of research roles · Course coming soon
-
GPU kernels
CUDA and Triton: writing, profiling and speeding up the operations a model spends its time in.
In 9% of research roles · Course coming soon
Foundations
Python, the math and production engineering every lab role assumes.
-
Python
The language almost every research and applied posting asks for.
In 48% of research roles · 5 pieces here
-
Production systems
Shipping code that survives: retries, state, monitoring and recovery.
In 33% of research roles · 10 pieces here
-
Math for ML
Probability, linear algebra and calculus, used rather than recited.
In 17% of research roles · 3 pieces here
Research record
Papers, results and the evidence of research a lab reads.
General engineering
Customer-facing work, systems languages and cloud infrastructure.
-
Working with customers
Scoping, integrating and handing over AI systems with the people who will run them.
In 17% of research roles · Partly taught here
-
C++, Rust or Go
Systems languages for performance-critical infrastructure.
In 14% of research roles · Not taught here
-
Cloud and Kubernetes
Containers, orchestration and cloud platforms.
In 13% of research roles · Not taught here
