# sarver-personal-ai-should-actually-be-personal-2026-09-29

## Veille

Long thread published on X on **September 29, 2026** by **Ryan Sarver** (@rsarver), titled *"Personal AI Should Actually Be Personal"*; roughly **1,700 words**, 91.5k views at time of collection. Starting point: **Amazon** blocked **Muse** (**Meta**) from its shopping site, and users lost access to their shopping agent, *"a side effect of a dispute I had no say in."* **(A)** Definition: a personal AI must be **owned** (open source, data in controlled storage, portability), have **a single principal**, offer a readable contract and inspectable trade-offs, run wherever one chooses, and be pleasant to use. **(B)** Observation on the current landscape: Muse opts users into training by default, **Instinct** carries a *"perpetual and irrevocable"* license; estimated cost gap of **$3,000 to $7,000 per user per year** against $20 a month. **(C)** A reversal: the owned stack can end up better, because the user gives it *"every corner of your life."* **(D)** A five-layer stack (clients, harness and connectors, memory, two-tier models, machine) and a call to build a productized version of **OpenClaw**. Responds to **Brad Gerstner** on trust; extends the body of work on shopping agents and markdown memory.

## Titre Article

Personal AI Should Actually Be Personal

## Date

2026-09-29

## URL

https://x.com/rsarver/status/2104928906650304618

## Keywords

personal AI, personal AI, personal agent, dual principal, single principal, data ownership, OpenClaw, Hermes, Muse, Meta, Instinct, Amazon, agent blocking, platform wars, training on user data, steering, silent steering, markdown memory, LLM wiki, GBrain, open-weights models, model routing, cost per user, rented machine, data sovereignty, business model, consumer product

## Authors

Ryan Sarver (@rsarver), auteur du fil sur X ; fonction non indiquée dans le texte.

## Ton

**Profile**: an opinion piece from a heavy OpenClaw user who describes himself as a customer of the companies named and writes to find a project owner (*"if you're building it, I'd love to hear from you"*).

**Style**: an argument structured as definition, case, technical stack, call to action. Short formulations (*"If you don't pay for the product, you are the product"*, *"OpenClaw feels like Linux. We need Apple experience"*), comparisons with web search and Linux/Apple.

**Epistemic stance**: an openly held thesis, figures reported from third parties or personal use ($3,000 to $7,000 per year heard from Anish, $2 versus $46 a month from an uncited breakdown, $10 a month for the machine). The author notes that Muse and Instinct are good products and that their terms are public; the inference about Meta's economic balance is posed as a question.

## Pense-betes

- **Trigger**: Amazon's blocking of Muse the previous week, with a message citing access by an unauthorized agent. The user, a customer of both, is not consulted.
- **Two principals**: an agent hosted by a platform serves both the user and the company that built it, and the latter wins out. Hence the criterion of a **single principal**.
- **Five criteria for "personal"**: owned, single principal, known contract (business model, data, inspection of what it weighed), free and reversible choice of where it runs, pleasant to use.
- ⭐ **Steering more serious than in search**: with ten links, the user actively chooses; with an agent, they delegate and just want the result, so steering happens **silently** through an opaque algorithm. Requirement: built-in preferences and inspectable decisions.
- **Economics**: cost of $3,000 to $7,000 per user per year versus $20 a month; raises the question of how the subsidy gets repaid by an advertising-based business.
- **Reversal of the data advantage**: normally the company holding the most data has the best product; here ownership gives the fullest access, and memory **compounds** over time.
- **Five-layer stack**: clients (Muse and Instinct are ahead); harness and connectors (OpenClaw, Hermes) with approvals and a log; readable, exportable memory (Karpathy's LLM wiki, Garry Tan's GBrain); two-tier models (open-weights by default, frontier invoked deliberately, ~$2/month versus ~$46); personal or rented machine (~$10/month).
- **Observation on OpenClaw**: solid foundation (chaired by Dave Morin), but too many settings for a general audience; needs a version that makes the choices, keeps them reversible, and adds mobile, desktop, and messaging surfaces.
- **Intended revenue model**: money comes only from the user; payment for updates, purchased software, machine and model calls sold by companies for whom this is the sole business.
- ⚠️ **Gaps**: no description of routing between models (called *"mostly hand-built"*); cost figures unsourced; no project owner identified.

## RésuméDe400mots

In this thread from September 29, 2026, Ryan Sarver argues that what is today called personal AI is not, in fact, personal. He starts from an incident: Amazon blocked Muse, Meta's assistant, from its shopping site, displaying a message describing the access as an unauthorized agent. Users, customers of both companies, lost a core agent use case, shopping, because of a dispute they had no part in. He sees this as the start of platform wars in which the consumer will be caught in the crossfire.

His central argument concerns the dual-principal problem. An agent hosted by a platform works for the user and for the company that built it, and it is clear which one wins out. He cites Muse's terms, where training on conversations is enabled by default, and Instinct's, which grant a perpetual and irrevocable license covering screenshots and keystrokes. He does not claim these companies are cheating: they disclose what they do. But a personal AI is informationally "maximalist," and it is precisely the one that should least be entrusted to a third party.

He sets out five requirements: owning the agent (open source, data under control, portability), having only a single principal, knowing the contract and being able to inspect its trade-offs, choosing where the agent runs, and having a pleasant product. Silent steering strikes him as more serious here than in search: with an agent, the user delegates and simply wants the outcome, so the steering happens **silently** through an opaque algorithm. He reports that assistants reportedly cost $3,000 to $7,000 per user per year, against the $20 a month charged, and questions how an advertising-funded company would recoup that subsidy.

His paradox is that ownership can produce the better product: the user hands everything to what they own, memory accumulates, and the most complete access is what drives quality. But OpenClaw feels like Linux; what's needed is an Apple-like experience.

He lays out the five-layer stack: polished clients; a harness with connectors, approvals, and a log; memory readable in markdown, in the vein of Karpathy's LLM wiki or GBrain; two tiers of models, an open-weights default (about $2 a month versus $46) and a frontier model invoked on demand; and a personal or rented machine (about $10 a month).

He calls for building a productized version of OpenClaw, whose revenue comes solely from the user, who can leave with everything.

## GrapheDeConnaissance

- Ryan Sarver —publie→ article Personal AI Should Actually Be Personal (DOCUMENT, 0.95)
- Amazon —s_oppose_à→ Muse (TECHNOLOGIE, 0.93)
- Meta —a_créé→ Muse (TECHNOLOGIE, 0.95)
- article Personal AI Should Actually Be Personal —affirme_que→ un agent hébergé par une plateforme a deux mandants et le second prime, donc l'IA dite personnelle ne l'est pas (AFFIRMATION, 0.93)
- article Personal AI Should Actually Be Personal —recommande→ IA personnelle possédée, à mandant unique, à contrat lisible, exécutée où l'utilisateur choisit (AFFIRMATION, 0.92)
- Muse —observé_dans→ entraînement sur les conversations activé par défaut (AFFIRMATION, 0.88)
- Instinct —observé_dans→ licence perpétuelle et irrévocable couvrant captures d'écran et frappes, y compris pour l'entraînement (AFFIRMATION, 0.88)
- agent personnel —mesure→ coût de 3 à 7 k$ par utilisateur et par an contre 20 $ par mois facturés (MESURE, 0.8)
- article Personal AI Should Actually Be Personal —affirme_que→ l'orientation silencieuse d'un agent est plus grave que celle d'un moteur de recherche, car l'utilisateur délègue le choix (AFFIRMATION, 0.9)
- article Personal AI Should Actually Be Personal —s_oppose_à→ Brad Gerstner (PERSONNE, 0.8)
- OpenClaw —est_basé_sur→ pile de l'IA personnelle (CONCEPT, 0.75)
- Muse —s_inspire_de→ OpenClaw (TECHNOLOGIE, 0.85)
- Hermes —concurrence→ OpenClaw (TECHNOLOGIE, 0.7)
- LLM wiki —est_instance_de→ mémoire en markdown lisible et modifiable (CONCEPT, 0.88)
- Andrej Karpathy —a_créé→ LLM wiki (METHODOLOGIE, 0.88)
- Garry Tan —a_créé→ GBrain (TECHNOLOGIE, 0.88)
- routage à deux niveaux de modèles —réduit→ le coût mensuel à environ 2 $ par utilisateur contre environ 46 $ par défaut (MESURE, 0.78)
- Dave Morin —dirige→ fondation OpenClaw (ORGANISATION, 0.85)
- article Personal AI Should Actually Be Personal —affirme_que→ "OpenClaw feels like Linux. We need Apple experience." (CITATION, 0.93)

---
Canonical: https://www.thekb.eu/en/fiches/sarver-personal-ai-should-actually-be-personal-2026-09-29/