Weekly Review — Aug 10–16, 2026

TL;DR: Five posts this week, one habit worth keeping: read past the headline before you decide. A metric is not the goal it stands in for (Goodhart’s Law). A model being big is about trainability, not just memorizing (the Beach Test). “AI moved into the database” is really two features, and only one belongs there. A watermark on AI text is a signal, not a verdict (content provenance). And GuardDuty’s new SOC agent does the correlation while the human keeps the call. The machinery keeps getting more capable and more opaque, and the durable skill is seeing what it actually does, then deciding, on purpose, what stays yours.

First full week back, and the five posts turned out to rhyme. Each one takes a thing everyone reads one way (a number, a big model, a database feature, a watermark, an “AI analyst”) and looks at what it is actually doing underneath. Then it draws a line: here is what the tool handles, here is the judgment you keep.

This Week on the Blog

Laws & Disorder #5 — Goodhart’s Law: The Goal You Specify Is the Only One You Get

The moment a measure becomes a target, it stops measuring what you actually care about: people, AI agents, and optimizers all hit the number and quietly sacrifice the goals you never put a number on. The post traces the same failure across a team told to “deliver 10 workshops,” reinforcement-learning reward hacking, and code tuned for one metric while the rest regress. The fix is not to stop measuring, but to express intent as fully as you can, balance three or more signals, and keep human judgment in the loop, for agent objectives too.

The Beach Test: Why Big Neural Networks Stop Getting Stuck

A statistical physicist’s analogy: the jamming transition that makes densely packed sand rigid mirrors a transition a neural network’s loss surface crosses as it becomes overparameterized. Below a size threshold the surface is rough and training gets stuck in poor minima; above it, in the model classes studied, near-zero-loss valleys open up and gradient descent can flow to the bottom. The reframe is that size buys trainability, not just memory, and because generalization can arrive before memorization, “bigger” isn’t the same as “a very expensive lookup table.” A careful analogy-not-identity read of why big models train reliably.

AI Moved Into the Database. I’ve Spent 25 Years Getting Logic Out of It.

SQL Server 2025 on Amazon RDS can now store vectors natively and call Amazon Bedrock straight from T-SQL, and those two halves deserve opposite reactions. Storing vectors is fine, because vectors are data and data belongs in the data tier; the database making the outbound model call revives the logic-in-the-database anti-pattern, adding in-transaction latency and non-determinism to the tier that is supposed to be boring. Store the vectors in the database, keep the model invocation in your application.

Content Provenance Is Growing Up: What Anthropic’s Watermark Really Signals

Claude’s newer models now mark what they generate two ways: an imperceptible statistical watermark on text (a probabilistic signal living in the token choices, not a deletable hidden character) and a cryptographically signed C2PA manifest on generated files, both driven by the EU AI Act’s Article 50. The honest core is that the text mark is a provenance signal, not a verdict: paraphrasing weakens it and it can even be spoofed. But paired with C2PA and a transparency rule that rewards human editorial control, it finally makes “can I actually use this content?” something media teams can operationalize, and ARD’s C2PA-on-AWS pipeline is the reference.

When the SOC Analyst Is an Agent (GuardDuty AI Investigations)

GuardDuty’s new preview investigation agent takes a raw finding, pulls 90 days of related activity, maps it to MITRE ATT&CK, and hands back a disposition with a confidence score and the CLI commands to remediate. The interesting part is not the automated evidence-gathering, it is the redrawn loop: the machine does the correlation humans hate, and the human keeps the call. The caution that comes with it: confidence is not correctness, and recommended commands are still commands.

The Thread This Week

Every post this week refuses the surface reading. What a metric rewards is not the goal you meant by it. What “bigger” buys is trainability, not rote memory. Which half of a database feature belongs in the data tier is not the whole launch. What a watermark proves is a signal, not a verdict. What the SOC agent owns is the correlation, not the decision. The tools get more capable and more opaque each month; the durable skill is seeing clearly what each one is actually doing, and then deciding, deliberately, what judgment stays on your side of the line.

Further Reading

The best public things I read this week that did not become their own post. Three of them circle the same media-and-bots question that runs through the provenance post:

• My Blog’s Biggest Fans Are AI Bots. So I Fought Back. Ran Isenberg walks through spotting AI-bot traffic with the AWS WAF AI activity dashboard (scrapers vs tools vs agents) and the options once you can see it: challenge, block, or charge.
• Meet the Media Industry’s Nerd-in-Chief A profile of the innovation lead reimagining the publisher business model against AI crawlers that consume content without compensation. The strategic backdrop to the same fight.
• Attention-Only Transformers: an FFN-removal study Removing the feed-forward layers (about two-thirds of non-embedding parameters) costs only ~0.27% quality at matched size, with the remaining gap localised to parametric recall. A mechanistic argument for why retrieval-first (RAG) architectures keep winning.
• A summary and discussion of that FFN paper The practitioner thread that surfaced it, with the enterprise framing: retrieval-augmented isn’t just cheaper, it may be architecturally superior.
• Arc XP — “Ask The News” Instead of losing readers to general chatbots, publishers embed a Q&A that answers only from their own journalism (retrieve, re-rank, verify, cite), keeping first-party signal and editorial control on-site. A clean production-RAG pattern for media.

Until Next Sunday

Five posts, one question worth carrying into the week: when a metric, a model, a database feature, a watermark, or an agent offers you a tidy headline reading, what is it actually doing underneath, and which half of the judgment are you keeping? Where would you draw that line in your own stack?

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

About the Author

Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.

This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.

Learn more →

• WeeklyReview
• AI
• Agents