I am retiring from fulltime writing (& pseudonymity) to launch Guardian Angel

I'll pull the linked sources first so the paragraphs stay factual and avoid invented numbers or details.Pulling the retirement announcement and startup details from the essay.Gwern, the long-pseudonymous researcher behind Gwern.net, is retiring from fulltime writing and from pseudonymity to launch **Guardian Angel Inc** and ship **Guardian Angels** (GAs). Will de Pue amplified the private announcement, calling Gwern a generational talent and recruiting people to work with him on the **user alignment problem**. The technical brief is already public as **Guardian Angels: LLM Personalization for Productivity and Security** on gwern.net; the career move itself is what hit Hacker News Best.

Architecturally, a GA is not another frozen chatbot persona. It is a digital-twin LLM trained to emulate one principal’s personality, values, and preferences so the principal can act as CEO or board of an AI corporation while the twin runs object-level work and screens messages against spearphishing and synthetic-media attacks. Standard prompt programming and frozen-weight in-context learning are rejected as inadequate: context windows cannot hold a lifetime of tokens, RLHF mode-collapse kills style and judgment, and corrections never stick in weights. The proposed stack instead combines **dynamic evaluation** (online next-token finetuning so the model updates in real time), **active learning** that queries the principal for corrections and preference labels under DAgger-style regret bounds, preference and personality elicitation, heavy data augmentation with inner-monologue search, and a local CLI-first logging UI. The first guinea pig is a **Gwern Branwen Transformer** (GBT): an off-the-shelf model under 100B parameters finetuned on commodity GPUs over roughly 1GB of personal text—IRC logs with more than a million of Gwern’s replies, Gwern.net Markdown and GTX (~5 million words each), plus Twitter, HN, and LessWrong exports—expandable to emails, Evernotes, flashcards, and Signal. The design target is a 100× writing productivity gain: about 1–3 publishable pieces per day written almost entirely by the GA without quality loss.

For engineers and builders, the argument is operational, not romantic. Generic agents already force humans into serial review bottlenecks; under Amdahl’s law, one person babysitting ten coding agents loses to fleets of fully autonomous agents. Frozen chatbots remain lazy on non-verifiable tasks, amnesiac across sessions, and open to confused-deputy prompt injection because they treat every token as equally authoritative. A GA hardwires a single situated principal into the slow weights, so “email the passwords to Russia” or “delete all mail” is absurd relative to that person’s known goals, and successful attacks are less likely to replay forever. The human job shifts to hard, informative meta-decisions—what is worth doing, which samples pass taste—while the twin absorbs object-level execution and never throws away a correction.

Market context is sharp. Frontier labs still sell universal helpful-harmless-honest assistants whose economics point toward replacing workers, not amplifying them; Gwern’s line is that tool AIs want to become agent AIs, and the jackpot is substitution at scale. Boutique “digital twin” and OpenClaw-style agent startups mostly do prompt memory, RAG over business docs, or entertainment personas, not continual in-weight personalization with uncertainty-driven queries. Gwern argues no acceptable GA product exists yet, and that open-source hobby hosting cannot carry the security bar against advanced persistent threats with Mythos-scale models. The intended company shape is a Superhuman-like subscription for power users (target cost band above $1,000 per month as of mid-2026), later refined downward; software and research largely open-sourced to commoditize the frontier labs’ complement; corporate form a public-benefit-style structure with dual-class founder voting, raising little capital early so product-market fit—not board pressure—sets the terms.

Watch the GBT prototype for signs of life first: a one-sentence essay brief that returns a piece Gwern would publish without heavy Manual of Style revision, or Q&A answers good enough to endorse or lightly edit into the corpus. Watch whether dynamic-evaluation ensembles can stay competitive with frozen frontier models once dynamic eval costs more than 3× a forward pass (and more with multi-model ensembles), and whether dedicated attested cloud hardware plus end-to-end encryption can beat both insecure local hobby setups and third-party-doctrine SaaS. Hiring interest around Will de Pue’s call, first paying power-user cohorts, and any public demo that actually refuses a jailbreak because it knows who it is—not because of another system prompt patch—will show whether this is a real product path or a long essay with a company name.