Nathaniel Forde

Thanks for the reminder. He was hand down the best thing I read in that school.

He's was a founder in something, something though... so, ya know, why split hairs.

I buy it.

Will never forgive them for that!

Once identified, these conditions suggest how we should design for robust collective decision making patterns in both organisations and agentic-architectures.

Hope to see you there!

This is not a talk about AI agents, but it indirectly probes the conditions required for useful truth-apt aggregates of agent opinion.

Python Ireland March 2026 Meetup, Wed, Mar 11, 2026, 6:30 PM | Meetup **Talk: Skill Estimation, Agents and the Condorcet Jury theorem.** The Condorcet Jury Theorem promises that groups are smarter than individuals — add more voices, get clos

Really looking forward the Python March-Meet-up in Dublin. I'll be speaking about Skill Estimation with IRT models in @pymc.io

We'l gauge skill and aggregate votes over agents to stress test the Condorcet Jury Theorem.

👉 Meetup Link: lnkd.in/d4bgGyAk

👉 Link to initial work: lnkd.in/dQ9TfFsa

Same group that killed the library bar iirc...

The Left Case for Great Books | The Point Magazine A great-books model at the undergraduate level is, in fact, so consonant with Freire’s radical critique that it represents a far better path forward for a left-wing vision of education than virtually ...

"a great-books model at the undergraduate level is, in fact, so consonant with Freire’s radical critique that it represents a far better path forward for a left-wing vision of education than virtually anything else currently on offer in the US" thepointmag.com/examined-lif...

Ah, don't knock the philosophy degree. Yglesias is not our best.

Bambi, PyMC...

Similar idea in the game God of War. Kratos carries around the chatty head of Mimir

This was fascinating! Thanks for sharing.

I don't have any answers, but i tried to work through some of these ideas with the condorcet jury theorem as the happy limit case:

Surprisingly easy to break the ideal conditons unfortunately: bsky.app/profile/nath...

A sensitivity plot of majority accuracy achieved by 30 votes when group think is at play. Even extreme and implausible "treatment" effects only slowly push towards accuracy.

The piece draws on James C. Scott's "Seeing Like a State" and @add-hawk.bsky.social 's steller work on value capture.

So, it mixes 18th century mathematics, contemporary anthropology and some philosophy, but if you've been thinking about organisation design or democracy you might find it useful.

The modeling suggests the problem is structural and hard to escape once established. I stress test these effects of groupthink and provide a sensitivity analysis to gauge how it might be disrupted.

There are no easy answers here.

The core question: what role does individual skill have when organisations optimize for "alignment" through culture fit, standardized training, shared AI tools?

How quickly does our collective wisdom plateau, and is this problem compounded by shared tooling and homogenised information diets?

The Legibility Trap

New blog post: The Legibility Trap - Why Organizations Fail in Lockstep

I stress test the Condorcet Jury theorem with Bayesian Item Response models and @pymc.io

Highlighting the role and importance of diversity in collective decision making

nathanielf.github.io/posts/post-w...

Book cover of Mike Davis's Prisoners of the American dream.

Maybe this one

Thanks owed to @juanitomedinart.bsky.social and others at @pymc-labs.bsky.social for the review and maintaining suc a great open source pacakge.

We also included control-function correction for price endogeneity:

• First-stage price model with instruments
• Correlated errors in utility
• Endogeneity-aware demand & WTP estimates

This is still uncommon in open-source choice tooling, but required for serious product preference models

What’s included in the implementation:

- Hierarchical random coefficients
- Individual- or group-level variation
- Centered and non-centered parameterizations
- Fully Bayesian inference in PyMC
- Wilkinson style formula interface

🚀 What shipped: Random-coefficients logit
This completes the discrete-choice stack in pymc-marketing:

MNL → Nested Logit → Mixed Logit

Built for real-world preference modeling. Why mixed logit matters 👇
It’s the workhorse model for:
• Preference heterogeneity
• Realistic substitution patterns

📘 Mixed Logit (Random Coefficients Logit) is now documented & available in @pymc-labs.bsky.social 's pymc-marketing
www.pymc-marketing.io/en/latest/no...

We’ve added a full mixed logit implementation to the consumer choice module.

🧵⬇️

The core insight: the journey IS the model.

Understanding how commitments transform over time lets us steer better, not just hope harder.

Whether you're tracking clinical outcomes or just wondering why plans go sideways—link in first tweet ⬆️

Happy New Year all! 🎆

This tutorial implements Aalen's dynamic path model in PyMC with:
✓ B-splines for smooth time-varying coefficients
✓ Poisson Bridge for proper generative likelihood
✓ Softplus link to maintain positive additive hazards
✓ Structural decomposition into direct/indirect paths

This pattern shows up everywhere:
- Clinical trials where treatments work brilliantly for 100 days, then trigger side-effects.
- Policy interventions that appear weak because protective effects get canceled by toxic indirect pathways.
- Personal goals that transform as you pursue them.

A series of evolving directed acyclic graphs showing the mediation relation X -> M, M -> X, X-> H with varying strengths across time,

The Problem:
We're measuring wrong. We treat interventions as binary switches: worked or didn't.

But causation isn't binary—it's a PROCESS.

Effects evolve through phases:
→ Strong direct effects early
→ Indirect pathways activate later
→ Hidden mechanisms undermine progress

New tutorial: Bayesian Dynamic Path Analysis in @pymc.io

Why do New Year's resolutions fail so reliably? The statistical answer involves time-varying causal effects and masked mediation.

Full tutorial (code + math + Odyssey metaphor): nathanielf.github.io/posts/post-w...

🧵 Thread Below

orderings, nevertheless his projects is still valuable to me, because I see it as the meta frame in which the applied work sits.