Weekly Review — Sep 21–27, 2026

Some weeks the posts pick their own theme. This one kept coming back to the space between two people at the same table, each living in a slightly different world: what one side sends versus what the other receives, and who gets to organize whom. Two of the five are the ongoing asymmetry series (feedback and knowledge), two are a pair about agent collectives that behave like organizations, and one is a fridge, because the machine humming in your kitchen turns out to be a better teacher than it lets on.

This Week on the Blog

The Feedback Exchange Rate

Feedback runs on a hidden exchange rate. The giver spends five minutes on a remark and moves on; the receiver replays it for three weeks, because bad lands heavier than good. Stack that on two more asymmetries (honest signal barely travels upward, and every day a review sits in a queue the author pays twice, once in stalled work and once in rebuilt context) and you get a system that over-charges the powerless and under-informs the powerful. The fix is mechanics, not moralizing: aim at the task, send at the weight you want received, and read silence as filtered rather than as approval.

Cold Is Just Absent Heat: Inside the Vapour-Compression Cycle

Your fridge cannot make cold, because cold is not a thing you can make; it is just the absence of heat. The machine is a heat pump that spends electricity to move heat uphill, out of the box and into your kitchen, running a four-stop refrigerant loop where pressure sets the boiling point. Because it only moves heat rather than conjuring a chill, it shifts roughly two to three times more heat energy than the electricity it draws. This is the second Lunch Break Physics.

When the Model Becomes an Org

A “model” is quietly turning into a self-organizing collective. In Noam Brown’s podcast account of an OpenAI agent swarm (his description, not independently confirmed), giving agents only a peer-messaging primitive was enough for them to grow a spontaneous middle-management layer, which suggests two old laws of org design, Conway’s law and Katzenbach’s group-versus-team distinction, may now apply inside the model. The builder’s job becomes a boundary decision: which organizational properties (identity, observability, accountability, steerability) do you keep outside the model on purpose, rather than letting them emerge where you cannot see them?

What If We Didn’t Teach Agents to Work Like Us?

A speculative companion to the org post. Agent collectives coordinate like people at least partly because they were trained on human text, so the resemblance is partly a reflection rather than proof that coordination must work this way. The essay sits on one fork: withhold that human prior and would a collective rediscover our playbook because it is genuinely efficient, or settle on a stranger shape with no human analogue that fits the work better? AlphaGo’s move 37 is offered as a narrow precedent for a machine finding, then teaching us, a move we had missed, held loosely because Go has a crisp win condition that open-ended coordination lacks.

The Expert Can’t Remember Not-Knowing

In a 1990 study, people tapped the rhythm of a famous song and predicted listeners would name it half the time; the real number was 2.5%. That is expertise in one image: once you know a thing, you lose access to not-knowing it, so the clarity you feel while explaining is not evidence that anyone understood. The gap is worse because it is invisible from both sides at once, the expert gets no error signal and the confused listener assumes their confusion shows. The fix is mechanics on both sides: test instead of ask (teach-back), surface the listener’s prior model, and never leave a conversation with a word you cannot define.

The Thread This Week

Four of these five posts are really one question asked in different rooms. Feedback is the gap between what a remark costs to give and what it costs to carry. Knowledge is the gap between the melody in the expert’s head and the knocking that crosses the table. The two agent posts move the same problem inside the machine: a collective that coordinates like an organization, and the open question of whether it must coordinate like us at all. Even the fridge belongs here, because “cold” is a word for a perception, not a thing, and the whole post is about the difference between what we feel and what is actually moving. The through-line, if there is one: most of the friction between two parties is structural, not personal, and the fix is almost always a mechanism rather than more goodwill.

Further Reading

Things I read this week that did not get their own post, all public:

• Father of RL Thinks LLMs Are a Dead End (Richard Sutton, Dwarkesh Podcast): The perfect counterweight to this week’s agent posts: Sutton argues LLMs mimic what people say without a goal, ground truth, or the ability to be surprised, and lays out his four-part experiential agent (policy, value function, state representation, world model) as what real learning would require instead.
• From the Age of Scaling to the Age of Discovery (Ilya Sutskever, Dwarkesh Podcast): Sutskever’s claim that the clean scaling recipe is hitting data limits (“more companies than ideas”), that models generalize far worse than people, and that the real reward hackers are the researchers who design training to match evals. Useful skepticism about benchmark scores when you pick a model for real work.
• Engineering Practices That Make Coding with AI Work (Simon Willison, Pragmatic Engineer): The concrete playbook behind “keep a human on the loop”: red-green TDD as the trust mechanism, layered verification, and the lethal trifecta of three conditions (private data, malicious instructions, an exfiltration path) that defines the security boundary for coding agents.
• Slow Down to Speed Up: AI and Software Engineering (Gergely Orosz, Pragmatic Engineer): The cautionary flip side: a real breach traced to code written by AI and reviewed only by AI, with usage-leaderboard incentives pushing the wrong behavior. A blunt reminder that removing the human review gate on security-critical paths is where the feedback loop actually breaks.
• OpenAI Habitat: Scaling Storage to One Billion Users, Part One: A rare infrastructure (not model) disclosure from a frontier lab: how a storage layer grew from a Python library into a globally distributed platform, and the “solve today’s problem cleanly, let real demand tell you what to rebuild” pattern that maps to how most teams actually evolve their architecture.
• The Most Cited Paper of the Century Is a Brilliant Hack (Welch Labs): A clean origin story for why modern models can be so deep: ResNet’s constructive proof that a deeper network can always match a shallower one, and the skip-connection trick that made “learn only how each layer differs from doing nothing” trainable. The identity-as-default idea travels well beyond neural nets.
• Why AI Tokens Are So Expensive (Computerphile): The plainest explanation of token pricing: generation is bound by memory bandwidth, not raw compute, and the KV cache grows with context length and competes for the same memory you would use to batch more work. A good companion to the fridge post if you like knowing what you are actually paying for.
• Your Platform Is Still a CoE in Disguise (Gregor Hohpe): A crisp diagnostic that pairs with the agent-org posts: a Center of Excellence is a consulting model (helping teams), a platform is a product model (self-service and opinionated), and most teams get stuck forever in the transition. The same “codify the decision instead of advising on it” move applies when you scaffold agent coordination.

Until Next Sunday

If there is a thread this week, it is that the distance between two minds is usually a mechanism problem, not a character problem, whether the two minds are a manager and a report, an expert and a beginner, or a swarm of agents and the human trying to steer it. Which of these gaps do you run into most, and what is the one mechanism that actually closes it for you? Reply or drop a comment.

This is the Weekly Review, and 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