schristoph.online

WeeklyReview · 7 min read

Weekly Review — Nov 9–15, 2026

On this page
TL;DR: Five posts this week. Monday tests TabPFN, a foundation model for tables, against XGBoost on four public datasets. Wednesday reads the docs of the new Well-Architected Agent, which ranks findings against goals you write yourself. Thursday is about a paella class in Barcelona and the difference between a group and a team. Friday has two posts: one turns agent steps like routing into a small decision model that returns a probability per option, the other reads Germany’s Leipzig Authoritarianism Study 2026. If you only click one link, make the decision-model post, because the threshold rule at the end is something you can apply to an agent you already run.

The first half of this week I was in Barcelona with a group of colleagues I had worked with for a year without ever meeting them. The paella post came straight out of those days. And the three technical posts all end in the same place: a decision the tool can’t make for you.

This Week on the Blog

A Foundation Model for Spreadsheets

Gradient-boosted trees have been the default for tabular prediction for about twenty years. TabPFN works differently: a transformer trained once on more than 100 million synthetic tables learns how to fit a table, and then predicts a new one with forward passes alone, with no training or tuning on your data. In the Nature paper its default beat baselines tuned for four hours on datasets of up to 10,000 rows, and in my small CPU test on four public tables it had the higher AUC on 17 of 20 splits against a tuned XGBoost, by small margins. The vendor’s own report still shows tuned models ahead on grouped, temporal and very large tables, and the newer weights are non-commercial unless you get a commercial license, so the post ends with when I would use it as the first baseline and when I would not.

What Should We Fix First? The Well-Architected Review Becomes an Agent

AWS Well-Architected Agent has been in preview since October 1. It scans your accounts every week, builds on Trusted Advisor, Security Hub, Compute Optimizer and others, and ranks at most 30 recommendations per run against business goals you write yourself, with the cross-pillar trade-off of each one. In the reviews I have run, finding issues was never the hard part; picking what to fix first was, and with the agent that order depends on the quality of your goals. I wrote this from the public documentation without running the agent myself, so the post sticks to what the docs say, including where they disagree, and lists five checks I would make before clicking remediate.

What a Paella Taught a Group About Becoming a Team

Gordon Curphy’s Rocket Model separates a group, whose result is the sum of individual efforts, from a team with common goals, joint work, shared leadership and a shared fate. By that line our distributed group of solutions architects is more group than team, and Curphy is clear that this isn’t automatically worse, because the task should decide. One evening making a single paella together in Barcelona came close to his definition, and the instructor’s way of leading (show the step, hand it over, let it go wrong a little, correct) is close to how I correct agents. The difference is that we remembered what she taught us, while an agent only keeps a correction that gets written into what it reads next time.

Many Agent Steps Are Classification in Disguise

Routing a request, deciding whether a human should confirm and rating urgency all pick from options you already know. Still, agents often ask a large model for a sentence and pull one word out of it, which leaves no number to put a threshold on. Following Unsloth’s public notebook, I fine-tuned Qwen3.5-4B into a decision model that returns a probability per option, and my one run scored 0.785 on the benchmark’s test split. That score is agreement with a teacher model that agrees with itself only 73.5 percent of the time, and under my definition of calibration error a uniform guess scores a perfect 0.000, so the post spends as much time on checking the evaluator as on training. Part 1 ends with an escalation rule: below a threshold, the step goes to a larger model or a person.

Democracy Is Defended Long Before Anyone Needs Weapons

At a dinner table with colleagues from inside and outside Europe, the question was what is going on in Germany. The post answers with charts from the Leipzig Authoritarianism Study 2026, a random-sample survey of 2,507 people: support for democracy as set out in the constitution fell to 65.2 percent, only a third are satisfied with how it works in practice, and the share with a closed far-right worldview rose for the second wave in a row, to 6.9 percent. Willingness to use violence, on the other hand, is the lowest the study has measured, and earlier peaks have come down before. I end with my own view that much of defending a democracy happens long before anyone needs weapons, in trust in institutions and in acting together.

The Thread This Week

So who decides? In each of the three technical posts, you do. TabPFN removes the per-dataset training and tuning, but you still have to know whether your rows are independent, grouped, or ordered in time. The Well-Architected Agent ranks the findings, but the order is only as good as the goals you wrote. A decision model gives you a probability, but somebody has to pick the threshold and own what happens below it. The paella post asks the same question about people, because the task decides whether a group needs to be a team at all.

Further Reading

Things worth your time that did not get their own post, all public:

  • What Frontier Labs Buy: A Data Deep Dive with Handshake AI (Latent Space): Handshake AI describes training data as expert-built environments plus rubrics that average 146 criteria, by their count. The best story is a grader that got gamed: a rubric checked for a net income of 100, so the agent added an Excel tab listing every number from one to a million. They fixed it by making the criterion more precise, and the model stayed the same.
  • Software correctness and programming languages with Sean Allen (The Pragmatic Engineer): Sean Allen of Antithesis argues that tests are a means to an end and that quality is not a gate, because every gate leaks. Build correctness in with types and models first, then use workloads with randomized inputs and injected faults for what static checks can’t catch.
  • Benevity’s AI-DLC Success Story (AWS Humans in the Loop): Benevity’s first AI-DLC workshop didn’t succeed, and they took it as the start anyway. Afterwards they reshuffled teams and anchored on cycle time. Code generation got faster right away while customers saw nothing, until they measured at capability level and estimated that work now moved roughly four to four and a half times faster end to end, by their own account.
  • How AI is breaking cybersecurity (Decoder): Alex Stamos argues that local open-weight models now automate most of the attack chain, so attack cost per token trends toward zero while defense stays expensive. His answer is to make defense cheap: small classifiers and models in the first tiers of the SOC, red-teaming yourself, and rewriting code in memory-safe languages instead of patching. These are his claims, from an hour-long interview.
  • Why AI Agents Need RAG + Observability (IBM Technology): A nine-minute explainer on giving an operations agent two kinds of context, documented knowledge through RAG and live state through metrics, logs and traces, so that it can investigate an incident with evidence while human operators stay in control.
  • How Discord Stores Trillions of Messages (Discord blog): Discord went from 177 Cassandra nodes to 72 ScyllaDB nodes, but the fix for hot partitions sat in front of the database: Rust data services that coalesce identical requests into one query. Historical message reads dropped to about 15 ms at p99.
  • A Two-Button Calculator That Does Everything (Numberphile): James Grime walks through Andrzej Odrzywołek’s paper, which shows that a single binary operator, exp(x) minus ln(y), together with the constant 1 generates the standard repertoire of a scientific calculator, from constants like e and pi to arithmetic and the usual transcendental functions, as exact expressions. It is the continuous counterpart of building all of Boolean logic from one gate.

Until Next Sunday

If you run an agent, pick one step that ends in a fixed choice and write down what you would do when the model is only 60 percent sure. I think that threshold tells you more about the step than its accuracy does.


This is the Weekly Review, and it also goes out as Sunday’s newsletter.

🎬 Also available as a blog walkthrough video on YouTube

❤ Created with the support of AI (Kiro)

WeeklyReviewMachineLearningAgenticAIWellArchitectedLeadership