This site is where that work goes. Three kinds of posts to expect:

1. Working analyses. Empirical pieces with the code attached, starting with capability-evaluation questions and the economics of AI deployment.
2. Technical notes. Things I had to understand properly to make progress: evaluation methodology, threat modelling, the machinery behind modern LLMs.
3. Field notes. Honest write-ups of the transition itself, what a quantitative social scientist has to learn (and unlearn) to contribute to technical safety work.

Everything substantive gets cross-posted to LessWrong, where the safety research conversation actually happens. This site is the canonical home.

If you’re working on capability evaluations, evaluation assurance, or the economics of AI and want to compare notes — email me.

Here’s a minimal example: simulating the classic econometric cautionary tale of spurious correlation between two independent random walks.

import numpy as np
import matplotlib.pyplot as plt

rng = np.random.default_rng(4)
n = 250
x = np.cumsum(rng.normal(size=n))
y = np.cumsum(rng.normal(size=n))

fig, ax = plt.subplots(figsize=(8, 4))
ax.plot(x, color="#197A56", lw=1.6, label="Series X")
ax.plot(y, color="#566070", lw=1.6, label="Series Y")
ax.set_xlabel("t")
ax.set_ylabel("Value")
ax.legend(frameon=False)
ax.spines[["top", "right"]].set_visible(False)
plt.tight_layout()
plt.show()

And the punchline, computed live:

corr = float(np.corrcoef(x, y)[0, 1])
print(f"Pearson correlation between the two independent walks: {corr:.2f}")

Pearson correlation between the two independent walks: -0.90

Two series with no relationship whatsoever, and a correlation near −0.9 that would excite any careless analyst. Non-stationarity does that. It’s a two-line simulation, but it’s also the reason unit-root tests exist, and a decent metaphor for why evaluation results need the same statistical care as any other empirical claim.

The workflow, end to end:

1. Explore in a Jupyter notebook inside VS Code.
2. When the analysis stabilises, move it into a .qmd file like this one.
3. quarto render executes the code and builds the page; results are cached (Quarto’s freeze) so the site rebuilds fast.
4. Push to GitHub, an Action publishes the site automatically.

Code, results, and write-up live in one file, in version control, in public.