Scaling Laws

Layer 1 · Intuition

Scaling Laws

Why loss curves look boringly predictable, and how a handful of power-law exponents rewrote how the entire industry spends its compute budget.

6 min read40 XP

Listen to this lesson

Podcast-style walkthrough · ~6 min

No recording yet · generated on first play

Test loss falls as a straight line on log-log axes as parameters, data, or compute grow.

The single most consequential empirical result in modern deep learning is not an architecture — it's a curve. Plot a transformer's test loss against the amount of compute spent training it, put both axes on a log scale, and you get an almost perfectly straight line, stretching across seven orders of magnitude. That line is a power law, and once you trust it, you can predict the loss of a model you haven't built yet.

Three knobs, one law each

  • N — parameters. More weights, more capacity to represent patterns.
  • D — training tokens. More data, less overfitting to any one example.
  • C — compute. Roughly floating point operations for one training run, so N and D trade off against a fixed compute budget.

Holding one of N or D fixed and growing the other, loss falls smoothly and predictably — until the other variable becomes the bottleneck and the curve flattens. The interesting question, and the one that split the field into two camps, is: for a *fixed* compute budget, what's the best split between N and D?

Kaplan et al. (2020, OpenAI)

  • Bigger models are dramatically more sample-efficient
  • Prescription: scale parameters fast, data slower
  • GPT-3 (175B params, ~300B tokens) followed this recipe
  • Optimizer/LR schedule details later shown to bias the fit

Hoffmann et al. (2022, DeepMind, 'Chinchilla')

  • Re-ran the experiment with fairly tuned LR schedules per run
  • Prescription: scale parameters and data ~equally
  • Chinchilla (70B params, 1.4T tokens) beat Gopher (280B, 300B)
  • Became the industry-standard 'compute-optimal' rule of thumb
Same question, two labs, two answers three years apart — the second one won.

Emergent abilities: real phase transition or measurement artifact?

Some capabilities — multi-step arithmetic, certain reasoning benchmarks — appear to switch on abruptly at a scale threshold rather than improving gradually, and this was branded 'emergence'. A well-known 2023 critique argued that much of this apparent discontinuity is an artifact of using discontinuous metrics (exact-match accuracy) on an underlying quantity (per-token log-likelihood) that actually improves smoothly. Swap to a continuous metric and many 'emergent' jumps turn back into smooth curves — but not all of them; some genuine threshold effects survive the swap.

  1. Train tiny models

    cheap, many configs

  2. Fit power law

    loss vs N, D, C

  3. Extrapolate

    predict loss at target scale

  4. Pick N, D split

    compute-optimal or inference-aware

  5. Spend the big budget

    with much less risk

Scaling laws turn 'how big should we train?' from a guess into a fitting problem.

Check your understanding

4 questions · answer all to submit

  1. 1.What does it mean, concretely, that loss vs. compute forms a straight line on log-log axes?

  2. 2.What was the central practical disagreement between the Kaplan and Chinchilla scaling law papers?

  3. 3.Why would a lab deliberately train a model well past its 'Chinchilla-optimal' token count?

  4. 4.What is the core claim of the critique that some 'emergent abilities' are measurement artifacts?