Log-Scale Thinking: When Linear Lies

Plot GDP of every country on a linear scale and the US, China, and Japan dominate; everyone else looks like zero. Plot on a log scale and you suddenly see that Norway is near the Dominican Republic by income per capita. Log scales make multiplicative differences visible.

When log scale helps

• Exponential growth (virus spread, compound interest)
• Power-law distributions (wealth, view counts)
• Measurements spanning many orders of magnitude (star masses, earthquake energy)
• Human perception (sound is heard logarithmically; decibels are a log scale)
• Model training loss curves (always plot on log-y)

Model scaling: a perfect log-scale story

Scaling laws for neural networks are almost always shown on log-log plots. The relationship between compute and loss looks like a straight line on log scale, but would be a dramatic curve on linear scale. This is why every paper on training curves uses logarithmic axes.

import matplotlib.pyplot as plt
import numpy as np

training_steps = np.array([100, 1000, 10000, 100000])
loss = np.array([4.8, 3.2, 2.1, 1.4])

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 4))

ax1.plot(training_steps, loss, marker='o')
ax1.set_title('Linear scale (hides the pattern)')
ax1.set_xlabel('Steps')

ax2.plot(training_steps, loss, marker='o')
ax2.set_xscale('log')
ax2.set_title('Log scale (reveals the pattern)')
ax2.set_xlabel('Steps (log)')

plt.tight_layout()
plt.show()

Two practical tricks

• Use log(1 + x) instead of log(x) when data contains zeros
• For percentages or ratios, use log-odds: log(p / (1-p))
• In pandas: df['log_followers'] = np.log1p(df['followers'])

The big idea: the world runs on multiplication more than addition. Log scales are how you see growth, scale, and skew clearly. Get comfortable reading them, and data starts making much more sense.