WebinarNotes/wiki/queries/2026-07-28-webinar-theses.md

# Webinar Theses v2 — From Chat Box to Your Own Agentic OS

#query

Supersedes [[2026-07-22-webinar-theses]] (7 sources). This set is synthesized from all **10** sources plus the current deliverable state in `raw/notes/` (`Webinar script.md`, `Webinar Plan - From Chat Box to Your Own OS.md`, `my theses.md`).

## Question

"Refresh the webinar theses" — restate the candidate theses for the talk *from chat box to your own agentic operating system*, now that [[2026-07-24-youre-reading-way-too-much-code]] and [[2026-07-28-agentic-engineering-10x-developer]] have been ingested. (2026-07-28)

## What changed since v1

| | Change |
|---|---|
| **Strengthened** | Thesis 1 (harness not model) — Thorsten states the strongest form: *the dominant variable in output quality is the information you put in*, and he tells people to **stop tuning model choice**. This is now the best-evidenced claim in the vault and it is already the script's literal closing argument. |
| **Upgraded from assertion to evidence** | Thesis 5 (you build it, one small tool at a time) — [[explosion-of-internal-software]] supplies an outside, *non-engineer-shaped* case: a 20-person social club's ordering process encoded in **~2 hours of phone typing**, from a photo of a menu. The talk's least-provable claim is now its best-evidenced one. |
| **Weakened — needs reframing** | Thesis 3 (skills are the new memory) — a frontier practitioner ships 99%-AI-written code with **no skills, no MCP, no slash commands**. See T3 below for the reframe that survives him. |
| **New** | Three theses the earlier set had no source for: verification (T6a), shedding weight (T10), and variations-not-answers (T13a). |
| **New honest caveat** | **Token budget** — Thorsten names it as one of two winner/loser variables. The talk currently promises a skill gap can be closed by effort; this says part of it is closed by spending. |

---

## The refreshed set

Grouped by the job each does in the talk. Bold = recommended for the 30-min cut.

> **Reading the numbers.** **T** = thesis; the numbers are stable handles so the theses can be referenced from other pages and in conversation ("T3 needs reframing") without re-quoting them. T1–T15 follow v1's order where the thesis survived, so a v1 number still points at roughly the same idea. A **letter suffix** (T6a, T13a) marks a thesis added in v2 next to its nearest relative rather than renumbering everything — T6a sits with T6 (both about where value goes when production is free), T13a with T13 (both method). Introduced in v2; v1 used plain 1–14.

### Spine — what the talk claims

**T1. The model isn't the product — the harness is.** *(strongest in the vault)*
Same model at every station; only the harness around it grows. Two independent frontier voices now say the same thing: the harness *is* the difference ([[harness]]), and "the dominant variable in output quality is the information you put in, not the model or the effort level" ([[thorsten-ball]]). Corollary you can say out loud: **stop shopping for models.**
— [[harness]], [[context-as-scarce-resource]] · script closing arc ("the model never changed")

**T2. A chat box is an app you open; an OS is a system that runs around you.**
The rung ladder: stranger → doer → yours → teammate → knows you → always-on. Karpathy's framing (in the script's notes) is the same claim from outside: website → app you download → *self-contained, persistent, asynchronous entity working alongside teams*.
— [[levels-of-ai-usage]], [[personal-ai-operating-system]] · Plan through-line

**T3. Context you *author* beats context that's *inferred*.** ⚠️ *reframed — see "The honest tension" below*
The v1 form was "skills are the new memory." That form now has a live counter-example. This reframe is what **all four** practitioners actually agree on: Konstantin's skills, Allie's foundation docs, Eugene's anti-memory position ("a skill is a file you can read, edit, version and delete"), *and* Thorsten's `AGENTS.md` are all the same move — context a human wrote on purpose, beating context a system guessed. It keeps the script's Memory→Skills stations intact while surviving the dissent.
— [[skills-as-memory]], [[personal-ai-operating-system]] · script Memory + Skills stations

**T4. Context is the scarce resource — every rung is a technique for spending it wisely.**
Already dramatized in the script better than any slide could: *"the notebook is tiny. On purpose. Everything in it gets loaded into every single session — needed or not."* Then skills as two-stage loading: "the shelf can be huge — the desk stays clean."
— [[context-as-scarce-resource]], [[evolution-of-agent-tooling]] · script Memory→Skills transition

**T5. You don't buy your OS — you build it, one small tool at a time.** *(now evidenced)*
Tools made for exactly one person, in an evening, asked-for rather than written. The new outside evidence matters because it defuses the obvious objection ("sure, *you* can do that — you're technical"): Thorsten's example is a social club, a phone, and a photo of a menu, and the software it replaced was a spreadsheet.
*(Scoped 2026-07-29: [[maintenance-is-the-real-cost]] adds the honest boundary — the cost of software is maintenance, not writing, and its pendulum case is a company abandoning its own Jira clone within four months. T5 survives because its examples pass that source's build-vs-buy checklist: tiny, personal, no users but you, no SLA. The claim is "little tools you make for yourself" — not "replace your vendors." Worth one sentence in the talk; it inoculates against the sharpest pushback a technical audience member could raise.)*
— [[explosion-of-internal-software]], [[emacsification-of-software]], [[personal-ai-operating-system]] · script OS section ("I didn't write it — I *asked* for it")

### Stakes — why now

**T6. The cost of producing work is going to zero; value migrates to directing and verifying it.**
Judgment, ownership, taste and relationships are what stay yours. Now has a hard datapoint: **99% of AMP's code is AI-written** — from inside a shipping company, not a demo.
— [[code-as-throwaway]], [[product-ownership]]

**T6a. Verification is the new craft — and you get it by asking for checks, not just work.** *(new)*
The audience's real objection is "can I trust it?", and the current script has no answer. There is one: generate disposable work whose only job is to check the work you keep. Engineer form: 100 lines of slop verifying every shipped line ([[make-more-cheap-code]]). Non-engineer form: fresh-agent misread tests, parallel interpretations, checker skills, synthetic-candidate simulations ([[2026-07-24-non-engineer-throwaway-verification]]). Delegated form: *"you're async anyway — ask the agent for proof"* ([[async-by-default]]).
— **Designed in full at [[2026-07-28-verification-beat-design]]** (placement, drafted script copy, three options by cost).

**T7. The gap between AI users and everyone else compounds — and is becoming irreversible.**
The person who builds their OS this week fears no release, because each capability slots into a system that already knows them. *(Thorsten names **token budget** as a second winner/loser variable, but that is a claim about metered agent-fleet work; for this audience the budget is one consumer subscription — keep the thesis on the skill gap. See [[enterprise-ai-reality]].)*
*(Measured 2026-07-30: this thesis is no longer prediction-only — Stanford SWEPR's 46-vs-46-team analysis shows the gap growing 4.8% → 19% (4×) from April 2023 to July 2025. One citable stage line: "Stanford measured it: the gap quadrupled in two years." See [[2026-07-30-stanford-swepr-widening-gap]]. **Added to the script's closing arc 2026-07-30** — T7 is now dramatized, upgrading it from Q&A material to an on-stage beat.)*
— [[2026-07-14-gap-between-ai-users-irreversible]] · [[2026-07-30-stanford-swepr-widening-gap]]

**T8. The more the world is mediated by AI proxies, the more valuable real human connection becomes.**
The "market of one" raises, not lowers, the price of being human.
— [[connections-as-moat]]

### Obstacles — what the audience actually hits

**T9. Adoption is blocked by friction, not resistance.** People aren't against AI — the setup is. Remove three clicks and they come.
— [[2026-07-14-nina-interview]], [[2026-07-14-yulia-interview]]

**T10. Ask which of your processes only exist because *you* were the bottleneck.** *(new)*
Thorsten's knife, translated for a business audience: backlogs, status meetings, approval queues, the spreadsheet everyone re-keys. His test — *would this exist if agents had always been available?* His suggested exercise is a company-internal doc titled **"Software Is Dead — Now What?"**; the audience version is one honest list. Strong candidate for "Do this tonight."
— [[shedding-weight]]

**T11. Even advanced users hit structural walls: no durable memory, integrations that dead-end, drift on loose specs.**
— [[2026-07-21-larysa-interview]], [[integration-dead-ends]]

**T12. Leave less room for imagination.** Every gap in your instructions gets filled — invisibly. Now with a concrete, teachable structure instead of a principle: **set the standard → state intent → riff on the design → specify the process → set the constraints.** Thorsten's own gloss is the most quotable line for a non-technical crowd: *"This is how I would talk to a senior engineer. This is the Slack message I'd send."*
— [[leave-less-room-for-imagination]]

### Method — what to do

**T13. Solve first, then skillify.** Don't design up front — solve the task once in conversation, then freeze the working recipe. (~3 messages of correction, or >5 tool calls, = it's skill time.) The script already demonstrates this exactly: *"save what we just did as a skill."*
— [[solve-first-then-skillify]]

**T13a. Ask for fifteen options, not one answer.** *(new)*
Generation is free, so the human's job moves from producing the artifact to **choosing among artifacts**. Thorsten's orb icon came from 15 AI-generated variants across 18 palettes; he picked one. This is the most immediately actionable thesis in the set for a non-technical audience — it needs no codebase, no skill, no setup — and it is the concrete form of "taste at AI speed."
— [[make-more-cheap-code]], [[product-ownership]]

**T14. The assistant does the research; you do the judgment.** The Insights Collector meta-punchline: this talk was mined out of AI-processed interview notes.
— script §3 / Plan §3

**T15. Walk in a week what took the industry three years.** One hour of foundation docs · one skill from your #1 recurring annoyance · one real file tonight.
— [[levels-of-ai-usage]], [[personal-ai-operating-system]]

---

## The honest tension (Q&A ammo — read this before you present)

**Someone may ask whether skills are necessary at all.** They are right to. [[thorsten-ball]] runs a 99%-AI-written codebase with no skills, no MCP servers and no slash commands; his context lives in the codebase and a team-maintained `AGENTS.md`. The vault records three readings and settles none ([[skills-as-memory]]):

1. **Situational** — he works daily in one codebase he controls, so his context can live in the code. Your audience has no codebase; skills are how a non-engineer gets the same effect. *(Strongest answer, and honest, but note he never scopes the claim himself.)*
2. **Premature abstraction** — skills are scaffolding for models that needed it, and a strong model plus a rich prompt may simply beat a skills library.
3. **Same thing under another name** — his `AGENTS.md` and AMP's curated sub-agents *are* two-stage authored context; the dispute is over who curates, not whether curation is needed.

**The safe framing on stage is T3 as reframed above** — *authored beats inferred* — which is true under all three readings. Don't claim the skills mechanism is settled; it isn't, and the strongest counter-example is a frontier practitioner rather than a skeptic.

Other live tensions, if the room is technical:
- Personal vs company-managed vs vendor-managed harness — [[enterprise-ai-reality]]
- Built-in agent memory as anti-feature (Eugene) vs persistent context docs used happily (Allie) — [[skills-as-memory]]
- Tight specs (T12) vs wide latitude — [[think-wider-not-bigger]]
- Model choice as a real lever (Eugene runs 4.7 over 4.8) vs a distraction (Thorsten) — [[leave-less-room-for-imagination]]

---

## Recommended cut for the 30-min format

Seven load-bearing theses, mapped to script beats. Changed from v1: **T3 reframed**, **T6a added**, T12 promoted (it now has a teachable structure), T9 demoted to Q&A.

| # | Thesis | Script beat |
|---|---|---|
| T2 | App you open → system that runs around you | whole spine |
| T4 | Context is the scarce resource | Memory → Skills transition (already scripted) |
| T3 | Context you author beats context inferred | Memory + Skills stations |
| T6a | Ask for checks, not just work | **currently missing — see gaps** |
| T1 | The harness is the difference, not the model | closing arc (already scripted) |
| T5 | You don't buy it — you build it | OS section (already scripted) |
| T15 | Walk it in a week | "Do this tonight" |

## Gaps in the current script this refresh exposes

1. **No verification beat.** The script demonstrates capability at every station and never once shows the agent being *checked*. For an HR/BA audience whose first question is "can I trust it?", this is the biggest hole — and T6a fills it cheaply (one line in the Skills station: a second skill whose only job is to check the first).
2. **No "what stays yours" beat.** T6 and T8 are in the thesis set and in the Plan (§4), but the script's closing arc is entirely about the harness. The talk currently ends on capability, not on the human.

*(A third gap — "the token-budget caveat is unsaid" — was proposed and **withdrawn 2026-07-28**. For this audience the budget is one consumer subscription, which is obvious and would land as a disclaimer. The corpus supports the withdrawal: [[eugene]] runs 7 parallel project-agents on a $200 plan and [[allie-miller]] runs ~100 agents, neither reporting a cost ceiling. Thorsten's token-budget variable describes metered agent-fleet work, not subscription use — see the scoping note on [[enterprise-ai-reality]].)*

## Evidence trail

- [[overview]] — through-line, agree/diverge map, vault-level open questions
- [[2026-07-28-lint]] — flagged v1 as two ingests stale and effectively orphaned; this page is the fix
- Sources: all 10, principally [[2026-07-28-agentic-engineering-10x-developer]], [[2026-07-24-youre-reading-way-too-much-code]], [[2026-07-14-skills-based-on-git]], [[2026-07-14-gap-between-ai-users-irreversible]], [[2026-07-21-larysa-interview]], [[2026-07-14-nina-interview]], [[2026-07-14-yulia-interview]], [[2026-07-14-sebastian-eugene-interview]]
- Deliverable state (raw, authored): `raw/notes/Webinar script.md` (ladder: Chat box → ReAct → Tools → Memory → Skills → Process → OS), `raw/notes/Webinar Plan - From Chat Box to Your Own OS.md`, `raw/notes/my theses.md`

## Follow-up questions

- Does T6a earn a station, or one line inside the Skills station? (Recommend the latter — a checker skill is one sentence of demo and costs no new level.)
- Should T10 ("which processes only exist because you were the bottleneck?") open the talk instead of closing it? It reframes the audience's own work before any capability is shown.
- The falsification test still unrun: same task, with and without a skill, in a non-engineer's hands. It would settle the T3 tension and would itself make a strong demo.

## Changed existing pages?

Yes — [[2026-07-22-webinar-theses]] marked superseded and pointed here. No concept or entity pages changed; this is synthesis. `index.md` and `log.md` updated.