# Scaling laws, evals & safety · AI in Bed

Advanced level · 7 min · https://www.aiinbed.com/learn/advanced/scaling-and-safety

*Why bigger got better, how we measure frontier models, and the open problems in keeping AI safe.*

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.

*Figure: On a log–log plot, loss falls in a remarkably straight line as training compute grows (illustrative).*

### Scaling laws

Kaplan et al. (2020) found loss follows smooth **power laws** in parameters, data and compute.

Chinchilla (2022) showed many models were under-trained: for a fixed compute budget, scale parameters and tokens together (roughly 20 tokens per parameter).

**Quick check:** What was the key Chinchilla finding?

A. Bigger models always win, regardless of data
B. For a fixed budget, many models were too big and trained on too little data
C. Data doesn't matter
D. Scaling has stopped working

Answer: B. Chinchilla showed compute-optimal training scales data alongside parameters; a smaller model on more tokens beat larger ones.

### Evaluating frontier models

Benchmarks test knowledge (MMLU), maths (GSM8K, MATH), coding (HumanEval, SWE-bench) and more.

Problems: benchmarks **saturate**, and test questions can leak into training data (**contamination**). So labs increasingly use held-out, human-rated and agentic evaluations.

### Alignment & safety

**Alignment**: making systems pursue what we actually intend. Reward hacking and sycophancy are everyday examples of misalignment.

**Robustness**: resisting jailbreaks and prompt injection.

**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.

> Pillow thought: "Concrete Problems in AI Safety" (2016) is a short, readable paper that frames many of these issues with everyday examples like a cleaning robot. A good one to fall asleep thinking about.

**Quick check:** What is "benchmark contamination"?

A. A virus in the benchmark files
B. Test questions leaking into training data, inflating scores
C. Running the benchmark on a slow computer
D. Using too many benchmarks

Answer: B. If a model has seen the test questions during training, its score no longer measures real ability.

**Key takeaways**

- 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.

**Go deeper (free sources)**

- [Scaling Laws for Neural Language Models](https://arxiv.org/abs/2001.08361) by Kaplan et al., 2020
- [Training Compute-Optimal LLMs (Chinchilla)](https://arxiv.org/abs/2203.15556) by Hoffmann et al., 2022
- [Concrete Problems in AI Safety](https://arxiv.org/abs/1606.06565) by Amodei et al., 2016
- [Transformer Circuits Thread (interpretability research)](https://transformer-circuits.pub/) by Anthropic interpretability team
