Xyra Sinclair
epistemic infrastructuralist · existentially oriented
ex-West Point · 1st of 4800 teams, 100-hour Interdisciplinary Contest in Modeling · 77.5 TB live behind SQL
I am making key parts of the internet ergonomically SQL queryable. Instead of a bunch of other organizations' databases being siloed and slow to access via their APIs, I swallow them all into one massive server and give people and their agents performant programmatic access over them. I solve the problem of CPU resource scarcity with free-floating congestion pricing (except for specially contracted users/orgs).
I also am working to provide a deeper "epistemic nucleus" suite of functionality, building out key compositional primitives for critical future workflows that can be rapidly assembled on top. That includes SOTA harnesses for having LLMs sort things in lists, interaction protocols for public/private/anonymous communication via LLM structured judgements, private mediation and economies of LLM labor, and in-browser agent functionality.
It is a solo operation; I have worked with great contractors and can scale fast. I'm raising to scale operations.
- 2015 obsessed with effective altruism as a lone Floridian (EA retweets begin Sept 2016) — enlisting at the same time as an Army deep sea diver — the highest-attrition, maximum proving-ground job
- 2016 set my $400k life insurance to go to MIRI and 80,000 Hours
- 2017–19 at West Point, genuinely training to lead America's soldiers in ground combat
- 2019 left West Point to build epistemic infra, after AlphaZero shifted my AI x-risk timelines to the late 2020s
- 2020 international sperm donation
- 2020–23 systems programming development and epistemic tool engineering
- 2024 decided early in the year that I needed to realize most of my expected impact on AI x-risk by the end of 2026; later started hormones to figure out what was up with emotions — an overwhelming good fit (self modifying for x-risk); incorporated ExoPriors, Inc.
- 2025 researched and presented on post-nuclear-war logistics for AI alignment, in person pitching a billionaire on unconventional strategies
- 2025 when Opus 4.5 came out, felt the immense pressure of epistemic-infra dreams suddenly, majorly at risk — rented a Hetzner server, because something had to happen; prompted over 12 hours a day
- 2025 launched Scry in December — public readonly SQL-over-HTTPS over much existentially important internet text like LessWrong, arXiv, millions of blogs
- 2025–26 decided the day after Christmas to leave when EA funding didn't come through — changing things up to keep affording the server, leaving a bad environment again when complacency was too dangerous; kept building from Thailand, traveling and living out of cheap hostels, until a cracked programmer who was taking the singularity seriously noticed my work — partially capitalized 5 weeks in
- 2026 invested in five-year SPX 20,000 LEAPS — another big strategic bet: as investors get singularity-pilled it runs early and can be sold early, so in the worlds where things really boom there is capital to make something good happen and pay for tokens, without gaming the market or being amazing at stocks
- 1st of 4800 teams, 100-hour Interdisciplinary Contest in Modeling
- 11th of 3100 teams in a math modeling competition (ICM 2018)
- 4th of 1000+ on December's token leaderboard; 2nd on tokscale.ai the week Fable arrived
- my public arts school had no sports — ran multiple marathons at 15 on my own
- craved battalion command at 19; appointed three pay grades above my rank
- became a full-stack technologist by following my curiosity — basically a big pre-training run for about 6 years
- Army deep sea diver — honor grad of the 7-month school, 85% attrition rate
- West Point's premier military skills competition team, then the triathlon team — 15+ hours of steady-state cardio a week on top of ~22 semester hour course loads and cadet duties; a big engine, the same one now pushing tokens
- 160B tokens pushed Dec–Jun
- 12.9B public records queryable — 77.5 TB live behind SQL; increasingly every academic paper, prediction market, intellectual forum, news article, blog post and comment
- live colo server: EPYC 9755, 128 cores / 256 threads, 1 TiB RAM; 420 TB of storage owned — 100 PB the dream
- bested a GM in blitz chess960. we got decent mid-game chaotic intuition.
- won 30 Verdansk Warzone mobile solos in a row, averaging a 56 K/D over 10 of them
what does existential competence look like?
OOD somaticing, out-of-distribution uplift,
the most ultra-sensitive, good-at-noticing-error, technologists (particularly around paradigms for high-bandwidth communication), with deeply varied life-history and interdisciplinary dominance, we need to find these people and unfuck them, because they can have wounds and be stuck in various ways like anyone else.
this is scry's premier purpose.
we're also working out private deep search, so we can do useful cognitive economy, multi-party search over private notes with bounded information leakage.
it's not a game. the people we need to contribute their slice of the existential security cathedral, such as more parsimonious training techniques and AI lab IP mediation infra, are often strangers, we don't really know them, they often don't apply, and when they do, the miserly bandwidth of 20th century grantmaking infra doesn't make it easy enough for grantmakers to overcome their own filters and biases in noticing signal.
instead of brilliant people strictly luckily rescuing themselves in the markets in time in the plane of bottom-up survival, it's time we start honoring intelligence explosion backpressure and approximating the conscientiousness of top-down existential rescue. weirder, higher trust financial instruments (made far more mediable with agents) should also have things pay for itself.— @XyraSinclair, x.com
it's time the ingroup, Team Conscientiousness, gets a 100 PB storage cluster including over 10 PB of PCIe 5.0 NVMe.
GPUs overrated, data being optimally traversable by single-threaded programs is not. when you have the storage paid for and live, economically and autistically getting the data becomes a much easier question.
let the autists get the data first, use cases with serious alpha will become more apparent as models and us get smarter.
am getting 1B tweets a day and have 420TB. traversability is what we do— @XyraSinclair, x.com
without cheap, highly ergonomic, privacy-respecting access to huge data, prosocial builders get outcompeted by default corporate optimization gradients.— @XyraSinclair, x.com, same thread
An old debilitating obsession, finally shipped open source as cardinal-harness: pairwise ratio judgements fitted into calibrated cardinal scores with uncertainty, active pair selection, budget control, and full cost receipts.
ultraoptimized pairwise ratio elicitation systems for token efficiently eliciting cardinal latents for interpretable attribute_prompts over conscientious entities. literal key civilizational bottleneck for conscientiously factoring cognitive work, autotaxonomizing things, metabolizing important alpha, and tractably doing things like incremental cognitive work over private data in privacy preserving ways.— @XyraSinclair, x.com
models need to be good at pairwise ratio comparison. the order of entities shouldn't matter. they should be consistent at it. the logprobs should have coherent distributions. models should be fairly coherent across entity sets. if labs simply used this as a fairly deep optimization target, they would get more profoundly useful and beautiful models.— @XyraSinclair, x.com
am working to make it effortless for LLMs to flush out arbitrary attribute-prompts-as-magnitudes across arbitrary entities, via pairwise ratio comparisons.
SELECT title, doi, year, ( 0.15 * magnitude['technical complexity to a new ML researcher who recently grokked transformers'] + 0.20 * magnitude['technical purity'] + 0.25 * magnitude['ahead of its time'] + 0.25 * magnitude['likelihood to go underappreciated'] + 0.15 * magnitude['juiceiness--having an air of aliveness'] ) as score FROM papers ORDER BY score DESC LIMIT 50;
People and LLMs grok stuff like paperA::“technical-purity” is roughly .75 of paperB, which is roughly .8 of paperC. Alternative entities can so often be prioritized very effectively with simple functions like linear combinations and threshold-filtering of subjective qualities that experts can sus out and rate things by.
LLMs have this remarkable ability to consistently rate entities (e.g. articles, papers, anything) by arbitrary subjective dimensions such as "sweetness" or "technical complexity to a new ML researcher who just recently grokked transformers" (with a little sophisticated scaffolding like eliciting via pairwise ratio comparisons).
People will have a much more empowered experience doing content curation, LLMs will be able to automate the analytic hierarchy process at scale, researchers will have vastly more interesting data to explore the world with.
I've been oriented to epistemic infrastructure for years since leaving the United States Military Academy at West Point in 2019.
I was obsessed with the concept of flushing out ultra-nuanced, coherent LLM judgements over the internet--basically building the infrastructure to tag everything with provenanced LLM judgements, so we could know that e.g. Claude Opus 3 thought LessWrong article A was 92nd percentile "existentially heavy" compared to all LessWrong articles it's looked at. I had a SOTA multi-objective reranker that minimized the number of LLM comparisons necessary to meaningfully rate a set of things, but for various reasons including virtually no money for LLM judgements, was struggling to deliver fast enough.
I incorporated ExoPriors, Inc. in 2024, because I had been obsessed for years with externalizing cognition in scalable, tractable, neglected, parsimonious ways. there was one really important paradigm of prosocial cognitive work, that nobody was doing well, which is provenanced LLM pairwise ratio comparison (and other structured judgements for prior elicitation and meaningful knowledge emission), and basically just doing the Analytic Network Process efficiently at scale.
When Opus 4.5 came out, I felt the incredible pressure to make something happen, and rented a Hetzner server with most of the last of my money before I knew what to do with it. Within a week, I knew that the somehow novel, profoundly useful thing I could do, would be to give people's agents arbitrary SQL+vector access over much important online content, like arXiv and LessWrong. I locked in harder than I ever have in my life, prompting over 12 hours a day, getting loans from family for 2 additional GPT Pro and Claude Max accounts, and delivered. I was going to run out of money soon for the server and AI coding subscriptions, and basically did the nuclear option of buying a plane ticket to Thailand to move out on January 1st and use the rest of rent money to afford the server and cheap living and a change of pace. Days later, I reposted Scry on HackerNews and got 350+ upvotes.
I'm still making the internet more SQL queryable, and I'm working on deeper collaborative agent infrastructure. The efforts fuse my technical ambitions, my desire for collaborative affordances relevant to surviving the next few years, and my serious for-profit ambitions.
My part-time contractors have been tremendously helpful and unblocking even given how much I'm able to lean on GPT-5.5 xhigh.
these provide statistical evidence of how I've been shaped, and are a result of past conditions and grace.
- as a kid went down all black diamonds with no poles my first day skiing (Wisconsin tho), swam class III rapids in 45°F water, sat off the front of the boat on class IV rapids holding myself down simply with a D ring between my legs. I tend to be much more responsible about risks nowadays, but I think rationalist/nerds tend to underrate the knowingness that can arise that you are actually safe in certain intense situations.
- my public arts middle and high school didn't have sports (although I was on the competitive varsity ballroom dance team) so I did distance running on my own, working up to multiple marathons at 15
- honor grad of Army dive school, which involves simulated drowning, cutting steel underwater with a 10,000° F exothermic lance, and seeing lots of folks including former D1 swimmers quit (85% of my cohort failed out and got reclassed).
- I genuinely craved to be a battalion commander within a few months at my first unit at 19 yo. At least I was able to be appointed to an E-6 position working directly with the top-3 (CO, 1LT, 1SG) as an E-3.
- competed on West Point's premier military skills competition team and then the Triathlon team. like 5:30 am swims, 15 hours of steady-state cardio a week plus practice logistics time, on top of my ~22 semester hour course load and multitude of cadet duties. my body was quite fast twitchy for triathlon too which meant it was really easy for me to overly damage body by going too lactic. after 10 months on the West Point triathlon team, days after I did my sub 5 hour half-ironman and 1:28 half marathon, I deadlifted 335 lbs (2.2x bw).
- was an international sperm donor (wanted to get over instinct to have kids that could sway my EA ambitions, offer better sperm than the worrying counterfactual lads, and perhaps increase the richness of my life with a minor co-parenting role that some women are into) and dated and lived with a wonderful woman 18 years older than me.
- came up with some solid optimizations for hardware streaming top-k sorting networks that save vast majority of dynamic power for large k. not a hardware jockey but in another life...
- managing to take like HGH, retatrutide, MOTS-c, and a bunch of other life enhancing stuff, on an income of $25k/year in Berkeley. Investing your last bit of money into yourself like that--that is frugality.
I did performance ballroom dancing because it was terrifying, underwater construction and reading Bostrom in the evenings while having my $400k life insurance set to go to MIRI and 80,000 Hours in 2016, trained to lead America's soldiers in ground combat, was an intl sperm donor to counter future desires to progenate which could affect my EA ambitions, learned software engineering to build epistemic tools, started estrogen to feel my feelings more because my agency was inadequate to respect my AI x-risk forecasts, pitched a billionaire on a post-nuclear-war logistics AI alignment strat, now am partially convinced of supracausal happenings and am gentleness maxxing.
I just hope I can learn lessons faster so I can be more strategically helpful in the future
- by early 2024 I decided that I needed to live so as to realize most of my expected impact on AI x-risk by the end of 2026, because after that, AI outcomes would be presumed to be largely locked in. And maybe that wouldn't be the median case, but as an uncommonly strategic strong longtermist, I had to be strong enough to worry about say, that 10-20% of early tail outcomes. In August 2024 I was facing the tough reality that I just didn't seem productive enough to reliably realize a significant fraction of my potential impact (e.g. I could not build my startup software fast enough, my nuclear alignment continuity work was not going anywhere). There was this old silly problem of "emotions" plaguing me--if only there was a pill to just magically make all your inner gunk go away (now I understand to be a very naive view of how things work). I was trying somatic therapy and was making some progress but had to accept that it just wasn't reliably working fast enough, and that something drastic would have to happen. Women tend to process their emotions better, so yeah I figured I could try estrogen, maybe 20% chance I like it and want to stay on it long term, but I figured there was a decent chance I could take it for 3-5 months and cry out lots of debilitating suppressed grief and then go back to testosterone dominance. So yeah I started estrogen in OCT 2024, and immediately got so much more than I bargained for, and haven't wanted to go back, although I have experimented a bit with exogenous testosterone. I don't know about gender identity, but I know that I deeply love dropping into my body and relating to the world with less armor. The old scripts I was being run by just weren't actually a decent fit for the world I found myself in, so I've faced a lot of ego death and have been finding increasing lightness and wellbeing in daily life.
so want to make an embroidered jacket to signal what I am about. like a varsity or motorcycle jacket but my own style. Existential Command big on the back, smol/cute/bi/poly on the sleeve, Rust logo, a diving helmet, triathlon symbol, US Army seal, etc. Like tattoos but way louder and more fearless— @XyraSinclair, x.com
I am fallible, I can't do it all alone.
- elite prompting skills (talent and very high volume)
- ultra-sensitivity
- outsized messy interdisciplinary clarity
- deep acceptance for weirdness and shadows, with a Chaplain's confidentiality
- 60 TB U.2 NVMe drives
- pointers to protocols and papers from The Book
- infinite tokens
- lots of approximates of felt-safety
- a built out sprinter camper van, for that singularity dynamism
- ε