06-reference

every model got stranger

Sat Apr 18 2026 20:00:00 GMT-0400 (Eastern Daylight Time) ·reference ·source: Every ·by Every Staff (digest; primary tip attributed to Every editorial)
everynotebooklmpyramid-principlebarbara-mintoinductive-synthesisopus-4-7continual-learningharness-thesismodel-strangeness

“The Model Got Stranger” — Every Staff (digest, Apr 19 2026)

Why this is in the vault

This is Every’s weekly digest framing the moment as the models are getting stranger, not just smarter — which is the same observation arriving from three other independent angles in this week’s processing cycle: Parrott on Opus 4.7 dropping its implicit prompt-engineering, Sutskever on the end of scaling and the rise of continual-learning architectures, and Sutton flatly calling LLMs a dead end. The featured tip — running Barbara Minto’s Pyramid Principle in reverse via NotebookLM (dump facts, group inductively, let summaries surface a thesis you didn’t have) — is a directly portable workflow for RDCO’s research and brief-building skills. Worth filing both for the strangeness frame and for the inductive-synthesis pattern.

Issue contents

Featured tip — NotebookLM as inductive Pyramid Principle. Minto’s classic management-consulting framework is top-down: start with the conclusion, organize supporting facts into mutually exclusive, collectively exhaustive groups beneath it. Every’s reversal: dump every random fact onto the page, let NotebookLM cluster them inductively by what they appear to be about, write a summary per cluster, then let those summaries push their way up to an answer you didn’t have when you started. Pitched as especially valuable for problems too complex to map manually upfront — and explicitly framed as a workflow that benefits a dyslexic writer who finds top-down outlining painful.

Curation section — notes

The “Curated link tail” below is Every’s own recent pieces — 100% self-cross-promo (every linked piece is on every.to, written by Every staff). This is sister-publication promotion in the curation slot, not third-party curation. Disclosed for vault transparency; treat each linked piece as needing independent assessment if surfaced.

  1. “Vibe Check: Opus 4.7 Stopped Reading Between the Lines” — Katie Parrott (Every staff)
  2. “The Folder Is the Agent” — Kieran Klaassen (Every staff)
  3. “(Re(Re))Introducing Sparkle” — Yash Poojary (Every staff)
  4. “Mini-Vibe Check: Claude Managed Agents” — Laura Entis (Every staff)
  5. “You’re the Manager Now” — Laura Entis (Every staff)
  6. “Living Software” — Jack Cheng (Every staff)

All 6 are already filed in the vault from Every’s K-tier watch — the [[wikilinks]] in the related: frontmatter point to those existing files.

Mapping against Ray Data Co

Direct portable workflow — /research-brief and /curiosity should test the inductive Pyramid pattern. Both skills currently start with a question or thesis and pull supporting evidence. The NotebookLM-reverse pattern inverts that: assemble the evidence pile first, cluster, let the thesis emerge. This is a near-perfect fit for the /curiosity skill’s “periphery-of-knowledge” mandate — the founder explicitly does NOT have a thesis at curiosity time, so an inductive synthesis pass over the recent vault deltas could surface candidate research questions that top-down prompting would miss. Worth a prototype: run /curiosity once with current top-down logic and once with an inductive NotebookLM-style pass, compare candidate question quality.

Strangeness as the meta-frame. “The model got stranger” is the editorial frame Every is settling into across a half-dozen pieces this week. It maps directly onto:

The convergence matters editorially: four independent voices in 72 hours all arriving at variants of “the model is becoming a different kind of thing, not just a stronger version of the prior thing.” This is a Sanity Check angle. Provisional title surface: “The Strangeness Era” or “Stop Calling It Smarter.”

Harness-thesis tie-in. If the model is getting stranger (less generous with under-specified asks, more dependent on continual-learning context, more in need of a verification layer per 2026-04-19-kingsbury-future-of-everything-is-lies), then the harness (skills, specs, scaffolding, eval rigs) carries proportionally more weight. Every’s curation tail this week is structurally about that — folder-as-agent, file-organizer-as-agent, “you’re the manager now,” living software. They’re describing a world where the human’s job is harness-building, not prompting. Same direction RDCO has staked.

Operator action surfaced. Add NotebookLM to the candidate-tools shortlist for the research stack. We have not deeply evaluated it; if it can credibly do the inductive cluster-and-summarize pass at vault scale (~1,500 docs), it could complement or even partly replace some of the manual /cross-check and /compile-vault work. Queue as a research-backlog item, not a now-task.

Sources & bias notes

See related: frontmatter for the full link block. Highest-leverage cross-references for this entry:

Per copy-paste caution: paraphrased throughout, no body text pasted, single quote ≤15 words. Body fetched via WebFetch on the canonical Every URL because Gmail thread returned snippet only.