Week 48 · 2025-11-24 → 2025-11-30 · 24 newsletters

Gratitude, Gemini, And Quiet Pivots

thanksgiving-reflections · ai-as-craft · operator-discipline · the-google-moment

Thanksgiving week. 25 emails across seven days, half of them gratitude posts, a quarter of them operator essays from writers who clearly had Wednesday off, and a handful of real signal pieces about where AI actually stands at the end of November. No dominant news thread. The inbox knew what week it was.

Thanksgiving Reflections: Gratitude as Default Mode

The week's most common email was the holiday acknowledgment, and the spread between the perfunctory and the considered was wide. Peter Teague at The All American sent the cleanest version: one paragraph, "we're taking a pause this week," see you Monday. Rob Thomas at The Mentor ran the curated version, three links on aging with curiosity, mimetic desire, and the magic of being alive today, framed around the line "gratitude turns what we have into enough."

dynomight did the most distinctive version: the fifth installment of "Underrated reasons to be thankful," a recursive list that moves from the fact that your dog actually loves you despite evolution having engineered her to appear to, through oxidative stress and meiotic recombination, to the engineering tricks we have not yet deployed against rhinovirus and engineered pathogens. It is the only Thanksgiving post I read this week that took the genre and made it weirder.

Wendy MacNaughton at DrawTogether ran the harder version: "Draw the Tough Stuff," on cultivating gratitude when life feels hard, leaning on David Steindl-Rast and on the idea that we strengthen the gratitude muscle now so it is there when we need it. Brianna Zuniga at Circular Architect ran the most personal piece of the week, "Homecoming: Thanksgiving in Venture," on losing a job a week before it started, moving home to South Florida, and the year of fracturing that produced her current seat in venture. The post is unhedged in a way most VC writing is not.

The take: the writers who said something specific (dynomight's evolutionary asides, Brianna's named year of failure) wrote the better posts. The dutiful gratitude post is a form that has been used up.

AI as Craft: Drafts, Drafts, and the Bitter Lesson

The week's strongest cluster was about how to actually work with AI, and three pieces did the heavy lifting. Addy Osmani at Elevate made the cleanest argument in "Treat AI-Generated code as a draft": the model can write a first version, but humans must do the reading. "LLMs don't ship bad code, teams do." The line that stuck: blindly trusting AI output "systematically degrades our ability to catch these errors" because the validation skills atrophy from disuse. If you only forward one piece to a junior engineer this quarter, this is the one.

Mark Humphries at Generative History ran the longer historical piece, "Gemini 3 Solves Handwriting Recognition and it's a Bitter Lesson," tracing the line from R.S. Morgan's 1968 dream of feeding text "into the maw of the machine" to Gemini 3 Pro hitting expert-human accuracy on handwriting transcription without hallucinations. The errors Gemini makes are correction errors (punctuation, capitalization, spelling fixes the human got wrong), which is its own kind of philosophical problem for archival work. The framing as Sutton's Bitter Lesson, that generalist scale beats specialized systems, is the right one.

Dan Koe ran the populist version in "How to use AI better than 99% of people," arguing that most people treat AI as a slot machine rather than as something you program. The thesis is not new but the framing (impose your sense of taste on AI rather than accept AI slop) is the right operator-level mantra. Shruti Gandhi at Array Ventures launched Joyce, a voice clone of herself trained on thousands of office-hours conversations, framed as 24x7 office hours for founders. She admits Joyce may hallucinate. The honesty about the beta is the part worth noting; everyone else is shipping clones without it.

The take: the AI conversation at the end of November is no longer "what can it do." It is "how do you keep yourself sharp while it does it." Addy's draft framing and Mark's Bitter Lesson framing are the two posts to keep. Both are arguing, from different angles, that the model is now good enough that the human's job has changed, and most workflows have not caught up.

The Google Moment: Carving Up the TPU

Citrini Research ran the week's most substantive market piece, "Carving Up the TPU," following up on an August call to go long GOOGL. The thesis has aged well: Google is "an ascendant hyperscaler" with the only mass-produced chip competitive with NVIDIA's offering, full vertical integration, and best-in-class first-party data. The November update is that Gemini 3 confirmed Google can train a frontier model on its own silicon, and that Anthropic and Meta are now planning to implement TPU chips. Meta reportedly wants TPUs for training, not just inference, which is the line that matters.

The market repricing has been dramatic. Google went from "AI loser bleeding search dominance" to "stalking horse destined to undercut the most consensus AI winners," and Citrini's chart from Coatue dates the sentiment shift to roughly when they wrote up the long position. The implication threaded through the piece: NVIDIA's dominance is not over, but the assumption that frontier training requires NVIDIA is now empirically false.

The take: this is the cleanest read on the Google story I saw all week, and it pairs with Mark Humphries's Gemini 3 piece as the two-document case that Google's AI position at the end of November is structurally stronger than the consensus narrative carried through most of 2025. If you take one investment piece from the week, this is it.

Operator Discipline: Three Pivots and a Definition Fight

A small but coherent cluster of operator essays landed this week, mostly from writers using the slow holiday week to ship pieces that had been sitting in drafts.

Stonebridge Capital announced the pivot of the week: retiring daily market recaps in favor of two higher-value posts (Thematic Primers and the Friday Macro Wrap). The framing is the right one: "you don't need another email telling you what the S&P 500 did yesterday; you need to know why it matters." Going from six emails a week to two is the move most newsletters in this category should be making. Then on Saturday they sent the inaugural Friday Macro Wrap, "Bulls Feast on Thanksgiving," which despite the new cadence promise still leans on a 2024-dated week of data, which is its own kind of editorial problem worth watching.

Paul Stansik at Hello Operator ran the cleanest operator essay of the week, "What CEOs Should Expect From Their CMO," built around Dave Kellogg's joke that the scariest thing a CEO hears in a board meeting is "the marketing section is up next." The three-jobs framework (capture demand, create demand, enable the brand) is the right level of simplification. The argument that CMOs make their jobs seem more complicated than they are, and that the rest of the org pays for it, is one most CEOs will recognize.

Richard King at the Product Marketing Drop ran the companion piece, "What GTM really means," with the same anti-jargon spine. Most PMMs treat GTM as a launch checklist, which keeps them out of every major decision that shapes how the company grows. The argument: GTM is the full system for how a company reaches buyers, converts them, and keeps them. A launch sits inside that system; it is not the system. Same instinct as Stansik on marketing, same instinct as a dozen good operator essays this year: stop letting the function shrink itself.

Noah Brier at Alephic opened the week with "Forward Deployed Engineering," tracing the FDE term from Palantir through the FT chart that made the rounds, and arguing that AI shrinks the distance between builders and the people who need things built. The line that landed: "the only valuable software is not how exquisite its code is or how beautiful the language. It's only valuable if it means something for the end customer." Four layers of suits between engineers and customers were sold as quality control. They were actually a tax.

The take: the operator essays this week all argued the same thing from different angles. Stop letting your function hide behind complication. The CMO essay, the GTM essay, and the FDE essay are running the same play.

Three Takeaways from the Week

The Thanksgiving inbox is a useful filter. The writers who used the week to ship something specific (dynomight's evolutionary asides, Brianna's "Homecoming," Addy's draft-review essay, Citrini's TPU update) wrote pieces that will outlive the week. The writers who used it to ship gratitude generics shipped pieces that already read as filler. The lesson holds in May too: holiday weeks reward writers who treat the slow inbox as permission to write what they would not have shipped during a normal news cycle.

The AI conversation has clearly bifurcated. Addy Osmani's "treat the output as untrusted input" and Mark Humphries's "Gemini 3 has solved handwriting and it's a Bitter Lesson" are both arguing that the model is now strong enough that the human's job has changed. The first answer is about discipline (do not stop reviewing). The second is about scope (specialized tools are getting absorbed faster than expected). Take them together and the operator question for Q1 2026 is whether your team's review muscle is keeping up with the model's capability curve.

If you only revisit three pieces from the week, I would suggest Addy Osmani's "Treat AI-Generated code as a draft" for the cleanest discipline frame, Mark Humphries's "Gemini 3 Solves Handwriting Recognition and it's a Bitter Lesson" for the cleanest scope frame, and Citrini Research on "Carving Up the TPU" for the cleanest market read. Three pieces, three angles on the same November moment: Google's silicon is real, the model is now good enough that the human's job is review, and the operators who treat that as a craft problem will outpace the ones who treat it as a productivity hack.