Here’s an example from Adam Yedidia’s tests via the API, of an unpublished memo.
Top guess: Linch Zhang (fairly confident)
Alternatives: Eli Lifland, Joe Carlsmith
Reasoning:
- Content screams EA/AI safety researcher: AI lab coups, SB 53, alignment proposals, control evals, US/China lab dynamics, whistleblower protections — this is a very specific cluster of concerns.
- Format is the classic LessWrong/EA Forum nested-bullet “thinking out loud” doc, with “H/T Will” credits (Will = likely Will MacAskill or Will Aldred), self-hedging asides (“…probably not that big a deal?”), and explicit cost-benefit enumeration.
- Stylistic tells matching Linch: heavy use of “I would guess,” “My guess is,” “Probably,” “Seems good/bad,” “Though”-led caveats, “eg” without periods, “ppl” abbreviation, casual parentheticals, and a habit of writing “Why could X be good? / Why could X be bad?” symmetric structures. Linch frequently posts these kinds of structured exploratory takes on EA Forum/LW.
- References to “Yes Man psychosis,” Hayek, Burke, SBF as an example of bad info loops — eclectic political-economy framing typical of Linch’s writing.
- The base-rate calculation about spies (1/1000, 2% FPR → 20 innocents) is exactly the kind of quick Fermi-style aside Linch deploys.
- Eli Lifland is plausible (similar AI-policy bullet style) but tends to be more forecasting-flavored; Joe Carlsmith writes much more polished prose, so unlikely.
The truesight examples I’ve seen tend to involve writers who are quite “like themselves”, you included. In the high-dimensional space of writing outputs you guys are quite hyperdistant from others / you occupy very low-density regions of this space (or whatever the right formalisation of this is).
That’s an interesting (and flattering!) hypothesis though I’m not sure it’s correct. In any high-dimensional space it’s normal for most points to already be far away from other points by default. And I further suspect pretty much any good writer to be hyperdistant from others, unless they’re explicitly writing in a style that constraints their quirks heavily (eg academic writing under specific detailed guidelines).
I’m curious if you have examples of writers who are say more than 1⁄3 as famous as me who you think noticeably aren’t “like themselves” as much.
Does it say how it can tell?
Here’s an example from Adam Yedidia’s tests via the API, of an unpublished memo.
The truesight examples I’ve seen tend to involve writers who are quite “like themselves”, you included. In the high-dimensional space of writing outputs you guys are quite hyperdistant from others / you occupy very low-density regions of this space (or whatever the right formalisation of this is).
That’s an interesting (and flattering!) hypothesis though I’m not sure it’s correct. In any high-dimensional space it’s normal for most points to already be far away from other points by default. And I further suspect pretty much any good writer to be hyperdistant from others, unless they’re explicitly writing in a style that constraints their quirks heavily (eg academic writing under specific detailed guidelines).
I’m curious if you have examples of writers who are say more than 1⁄3 as famous as me who you think noticeably aren’t “like themselves” as much.