You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
What exists. Over the past year I built two working sideloads — interactive digital copies of real, consenting people, constructed from their full text corpus, correspondence, interviews and voice. They run as autonomous chatbots with long-term memory, a verified-facts ontology, and a strict no-confabulation rule (the copy says "I don't remember" rather than inventing memories). The key technical asset: a deterministic style-vector metric (~24 dimensions) that scores any generated reply against the person's real corpus. Raw LLM roleplay scores ~63 against the person's own baseline of ~92; my calibration loop brings output to 80–90+. Fidelity of personality emulation is currently asserted by every product and measured by none — I measure it.
The safety-relevant question. Style fidelity is measurable. Is judgment fidelity measurable — and achievable? Concretely: given a sideload of person X, how often does its evaluation of a novel decision match what X actually says, on held-out cases the model never saw? Nobody has this number. It matters for three reasons:
1. Scalable human oversight. Today "human feedback" means anonymous annotators and reward models with no fidelity guarantee to any actual human. A sideload with measured judgment-fidelity is a different primitive: an auditable "would this specific trusted person approve?" that runs at machine speed inside AI decision loops — reviewing agent actions, flagging decisions the original would object to, escalating to the real human when confidence drops. It does not replace the human; it extends one specific human's oversight bandwidth by orders of magnitude, with a number attached telling you how much to trust the extension.
2. Preserving and scaling alignment researchers. Senior alignment researchers' judgment is among the scarcest resources in the field, and it is mortal, non-copyable, and lost on every career exit. A consenting researcher's sideload — built from their papers, reviews, correspondence and structured interviews — could pre-review research at scale, red-team proposals, and answer "what would X say about this?" with a known fidelity score. I want to build one such sideload with a consenting AI safety researcher and run a blinded evaluation: colleagues judge, without knowing which is which, whether answers to novel technical questions came from the researcher or the copy.
3. Survival channel. If alignment partially fails, verified digital copies of humans inside AI systems are a candidate preservation route for human minds and values — reconstructable data rather than nothing. I have argued for related mechanisms in published work ("Message to any future AI", "Digital Immortality: Theory and Protocol for Indirect Mind Uploading"); sideloads turn that argument into an engineering practice with measurable fidelity.
What the grant buys (12 months).
Judgment-fidelity benchmark: held-out evaluation protocol measuring sideload-vs-original agreement on novel judgments; publish the current gap and how far calibration closes it.
One researcher sideload (consenting AI safety researcher) + blinded colleague evaluation.
Open protocol: full methodology — corpus collection, ontology, anti-confabulation memory, style/judgment calibration — published so results are reproducible and not locked in one project.
Honest objections. A sideload is an LLM shaped by a person's data; under distribution shift it may revert to base-model priors. Judgment fidelity may plateau well below usefulness. An oversight sideload could itself be manipulated by the system it oversees. These are exactly why the ask is for measurement, not deployment: if fidelity turns out low and uncloseable, that is a decision-relevant negative result against all human-emulation oversight proposals — cheap now, expensive to learn later.
Tokens, computer, and personal expenses
I am alone.
Track record. 15+ years of published existential-risk research (global catastrophic risks classification, AI safety via messaging future AI, digital immortality theory); co-authored a futurology book; two working sideloads as of today.
Not funded. Or experiments will turn negative results.
None