SWLW #711: The laws of this world (game theory), Should you still become a Software Engineer in 2026, and more.

Oren Ellenbogen · email · 2026-07-10 Reader 🖥️

AI Labor Disruption Capacity to Verify AI Output Career Development Hiring Signal Design Leverage & Positional Advantage

As AI compresses both the cost of writing code and the reliability of the signals people and organizations have used to judge value (interviews, hours logged, raw effort), this issue's throughline is that the decisive differentiators shift to what resists commoditization: disciplined distrust toward AI output, deliberate positioning/leverage over sheer effort, and self-directed habits (monthly growth checks, daily writing) that build judgment rather than borrow it.

practical · mixed — A curation newsletter: short editorial blurbs around seven linked picks plus two closing tweets. Practical because the through-line is prescriptive about how engineers and leaders should act (verify AI output, seek leverage, re-assess skill monthly, write daily); mixed because the thesis emerges from selection and framing rather than one sustained argument. Several of this issue's picks already have standalone full extractions in the corpus (the Favorite video, the game-theory/leverage piece, Yegge's interview piece, Bayes's predictions, and Huynh's five-habits piece) — this extraction treats the digest itself as the unit, records the editorial framing, and relates it to those existing extractions rather than re-deriving their content.

Structure

Part	Role
This Week's Favorite — 'Should You Still Become a Software Engineer in 2026?' (video), framed around Scott Hanselman's line that he treats AI-generated code exactly like an anonymous stranger's pull request.	thesis anchor — positions zero-trust, adversarial review of AI output (not blanket distrust of AI as a category, but the same skepticism applied to any unvetted contributor) as the load-bearing practice of the issue.
Culture picks — a vibe-coding humor tweet, Huynh's 'monthly growth check' habit, Sieger's leverage/position framing (linked article already fully extracted in the corpus), and Tunguz's 'AI agency' GTM pattern for ML SaaS startups.	evidence — examples of deliberately-chosen position/leverage and self-directed practice as alternatives to raw effort, including at the business-model level (Tunguz).
Peopleware picks — Bayes's 46 near-future predictions (an 'algorithmic takeoff' piece already fully extracted), Yegge's account of the broken hiring-interview signal (already fully extracted), and Puri's 'daily writing is the new weightlifting' framing.	evidence / counsel — the compressing value of conventional signals (interview performance, hours worked) and the compensating value of habits that can't be faked or delegated (writing as thinking).
Closing tweets — Sahil Bloom's grandfather ('consistently reliable' beats 'occasionally extraordinary') and a Carl Jung quote on quiet, conscientious work being noticed.	conclusion — crystallizes the through-line: durable, demonstrated reliability and genuine effort on real work outcompete episodic brilliance or borrowed signal.

Key terms

Adversarial / zero-trust AI code review
Hanselman's framing, foregrounded by Ellenbogen: AI-generated code gets exactly the scrutiny an anonymous stranger's open-source pull request would get — the same ingress checks, the same willingness to disbelieve, cross-checked across multiple models — rather than a presumption of competence because it came from 'AI.'
Leverage vs. effort (people / capital / media / code)
Sieger's game-theoretic claim, summarized in the digest: effort's payoff is linear, but leverage across four types — people (network), capital, media (visibility), and code (automation) — multiplies output exponentially; most senior engineers default to the one leverage type (people, via harder work within existing relationships) they already have instead of acquiring the others.
AI agency (GTM pattern)
Tunguz's observation, as summarized: ML SaaS startups adopting an agency-style delivery model — doing the client's work outcome directly rather than selling them software — capture 95-100% of a client's OPEX spend (vs. 5-10% as a pure vendor) while using automation to avoid the agency's usual linear headcount-to-revenue scaling.
Hiring-signal breakdown
Yegge's point, as summarized: a Google hiring-committee calibration exercise in which reviewers unknowingly rejected anonymized packets belonging to their own committee members shows the standard interview doesn't measure what it claims to, and the labor-market shift from scarcity to surplus removes the cover that used to hide this.
Monthly growth check
Huynh's practice, as summarized: asking monthly (not annually) what you can now do that you couldn't do a month ago, because specific technical skill decays faster than people assume and annual review cadence hides that decay.

Propositions

explicitScott Hanselman treats AI-generated code with the same zero-trust, adversarial review he would give an anonymous stranger's open-source pull request — cross-checked with multiple models — rather than extending AI any presumption of competence.conf: high
supportsShould You Still Become a Software Engineer in 2026? GitHub VP · p3 — Same Hanselman claim (AI cannot be held accountable, so the human reviewer bears full responsibility exactly as with an anonymous PR) restated at digest length; the full extraction carries the complete argument and context.
explicitEffort's payoff scales linearly, but position — leverage across people (network), capital, media (visibility), or code (automation) — multiplies output exponentially, and most senior engineers have all four leverage types available yet default to working harder within the one they already use (people).conf: high
supportsThe Laws of this World (Game Theory) · p2 — Digest blurb restates the source article's core claim verbatim in substance (effort linear, position/leverage exponential) — the full extraction already carries the complete proposition set and gaps for this piece.
explicitThe smartest ML-startup go-to-market move is to behave like an agency (delivering the client's outcome directly) rather than a pure software vendor, capturing 95-100% of a client's OPEX spend instead of 5-10%, while using automation to avoid the agency's usual linear headcount-to-revenue scaling.conf: high
explicitAs AI increasingly performs the 'knowledge' part of knowledge work, it becomes genuinely hard to locate where humans still contribute marginal value, and the practical question for any team or individual is where their own margin lands once that marginal human contribution approaches zero.conf: high
supports46 thoughts on the near future · p18 — Ellenbogen's 'where's your margin once it goes to zero' is the same open question the source piece poses about what fraction of humanity retains paid employment as AI absorbs knowledge work.
explicitA Google hiring-committee calibration test — in which reviewers unknowingly rejected anonymized packets belonging to their own committee members — shows that standard technical-interview signal is unreliable even by the standard of the people running the process.conf: high
supportsThe Last Technical Interview · p2 — Digest's retelling of the Google calibration test restates the source piece's claim that interviewer judgments are highly inconsistent and poorly correlated with on-the-job performance.
explicitA flawed hiring-signal process was tolerable overhead when candidates were scarce; now that the market has flipped to a talent surplus, the hidden cost of a broken interview signal is no longer masked by how hard it was to get anyone in the door.conf: high
supportsThe Last Technical Interview · p6 — Digest's 'labor market flipped from scarcity to surplus' framing restates the source piece's account of the converging pressures breaking the old hiring equilibrium.
explicitEngineers who are actually compounding skill run a self-assessment every few weeks rather than annually, asking what they can do now that they couldn't do a month ago, because specific skills decay faster than the default annual-review cadence assumes.conf: high
supports5 Things World-Class Engineers Do That You Don't · p1 — Digest blurb restates the source piece's first practice (monthly growth check vs. annual review) verbatim in substance.
explicitDaily writing functions as cognitive exercise whose value increases rather than decreases in an AI-generate-everything environment, because the ability to produce original insight and direct AI with deterministic guardrails is what remains genuinely scarce.conf: high
explicitConsistent, reliable follow-through compounds more over time than occasional extraordinary performance.conf: medium
implicitThe issue's curation implicitly treats disciplined distrust of AI output (Hanselman) and seeking leverage over raw effort (Sieger) as compatible advice, without reconciling the tension that sustained, hands-on adversarial review of AI code is itself a form of effort that does not obviously scale the way the leverage framing promises.conf: medium
extendsTeam Topologies: Organizing Business and Technology Teams for Fast Flow · p4 — The digest's unacknowledged tension (adversarial AI-review effort doesn't obviously scale) is a concrete instance of Team Topologies' cognitive-load argument: review/verification capacity, not raw execution capacity, becomes the binding constraint once AI supplies cheap execution.
implicitThe issue implicitly assumes that honest, verifiable signal (hiring, output vs. effort, skill growth) is getting harder to produce credibly at the same time AI is making it easier to fake superficially — an unstated tension between the Yegge pick (signal collapse under AI-assisted gaming) and the Hanselman pick (confidence that disciplined process can restore trustworthy signal).conf: medium

Arguments

Engineers should extend the same adversarial scrutiny to AI-generated code that they would give any unvetted contributor, rather than treating 'AI wrote it' as either a reason for extra suspicion or extra trust. strength: medium

Explicit premises:

Hanselman does not trust an anonymous stranger's PR by default.
AI output, like a stranger's PR, has no track record attached to it.

Implicit premises:

The reviewer has, or can assemble, the domain knowledge needed to actually catch what adversarial review is supposed to catch — the review process itself doesn't replace the reviewer's competence.
Multiple-model cross-checking is a practically available option in the reviewer's normal workflow.

Inferences:

If true at scale, this implies review capacity (not code-generation capacity) is the real constraint on how much AI-assisted work a team can safely absorb — directly consistent with the cognitive-load framing in extract-team-topologies.

Narratives

The Google hiring-committee calibration test, in which bar-raisers reviewed anonymized candidate packets and unknowingly rejected their own committee colleagues.
→ Supports the proposition that standard technical-interview signal is unreliable even by the standard of its own practitioners, independent of any AI-era pressure.

Questions addressed

✅ What should an individual engineer or leader actually change about how they work, given that both AI capability and the reliability of traditional performance/hiring signals are shifting fast?
Verify AI output adversarially rather than trusting or distrusting it categorically; seek leverage (not just effort); re-assess your own skill growth monthly; keep a daily writing practice as a thinking discipline.
⭕ How does disciplined, effortful verification of AI output (Hanselman) coexist with the advice to escape linear effort in favor of exponential leverage (Sieger)?

Gaps

illogicalThe issue juxtaposes 'distrust AI output and review it adversarially' with 'stop relying on raw effort and find leverage instead' without noticing that rigorous adversarial review of AI-generated code is itself a form of effort that resists the leverage framing's promise of exponential, effort-decoupled return.
Read next: A treatment of AI-era verification cost as its own leverage category (e.g., tooling/process that multiplies review capacity rather than review hours) would resolve the tension — close to the ceiling-on-AI-assisted-work argument in feedback-loops/automated-signals literature.
incompleteThe digest repeats Bayes's 'where's your margin once knowledge work's human contribution goes to zero' provocation without connecting it back to its own earlier picks (Hanselman's human-in-the-loop review, Yegge's hiring-signal collapse) to ask what becomes of either practice if the premise is right.
Read next: extract-46-thoughts-on-the-near-future's own propositions on labor displacement and permanent-underclass risk are the natural next read for this unexamined connection.