Advanced · 07
Scaling laws, evals & safety
Why bigger got better, how we measure frontier models, and the open problems in keeping AI safe.
7 min read · Tap underlined words for a definition
Swipe up to begin
Much of the last decade of AI progress came from one discovery: performance improves predictably as you scale up compute, data and model size. That predictability shapes billion-dollar decisions, and raises serious questions.
Scaling laws
Kaplan et al. (2020) found loss follows smooth power laws in , data and compute.
Chinchilla (2022) showed many models were under-trained: for a fixed compute budget, scale parameters and together (roughly 20 tokens per parameter).
What was the key Chinchilla finding?
Evaluating frontier models
test knowledge (MMLU), maths (GSM8K, MATH), coding (HumanEval, SWE-bench) and more.
Problems: benchmarks saturate, and test questions can leak into (contamination). So labs increasingly use held-out, human-rated and agentic evaluations.
Alignment & safety
Alignment: making systems pursue what we actually intend. and sycophancy are everyday examples of misalignment.
Robustness: resisting jailbreaks and .
Interpretability: understanding what's happening inside the network, e.g. finding human-understandable features inside models.
Responsible deployment
Labs publish model cards and system cards, run red-teaming, and test for dangerous capabilities before release.
Policies like responsible scaling commitments tie stronger safeguards to more capable models.
What is "benchmark contamination"?
Before you drift off
- Loss falls predictably with scale (power laws).
- Compute-optimal training scales data with parameters.
- Benchmarks saturate and leak, so evaluation is an active field.
- Alignment, robustness and interpretability are core open problems.
Terms in this lesson
Go deeper, free and open
- Scaling Laws for Neural Language Models · Kaplan et al., 2020
- Training Compute-Optimal LLMs (Chinchilla) · Hoffmann et al., 2022
- Concrete Problems in AI Safety · Amodei et al., 2016
- Transformer Circuits Thread (interpretability research) · Anthropic interpretability team
That's the lesson
Scaling laws, evals & safety · 7 min
Each lesson you finish lights a star.
Spotted a mistake? Suggest an edit
Previous: Reinforcement learning, deeper