# Teddy Wright — my content bundle Generated 2026-05-23 from theodorewright.dev. My own writing, research, models, and updates — every essay, paper, model writeup, and weekly note on the site as one markdown file. AI-Research pipeline output is a separate bundle (see /bundle-ai-research.md). **About:** I have a problem where I think about way too many things at once and can't quite pin one down. So writings here span evolutionary biology, game theory, philosophy, the extremity to which modern society is different than what we evolved in, and whatever else I feel connects to the strange place the earth is and how we all appeared with consciousness here and have to deal with it. **Site status (as of 2026-04-29):** Trimming the front pages and tightening the model and dashboard rosters. **Contact:** theodorewrightwork@gmail.com · https://substack.com/@theodorealan · https://github.com/theodorewright11?tab=repositories # Writing ## information is now dirt cheap with ai — the scarce resource is now logic and reasoning *2026-03-19 · tier: mine* AI hasn't just made information cheaper to find — it's collapsed the entire cost structure of acquiring, distributing, and repackaging knowledge. The thing that used to differentiate people (knowing stuff) is now basically free, which means the actual differentiator is logic: the ability to select, frame, judge, and reason about information. i was writing a research paper recently and something clicked that i hadn't fully processed before. i was using ai heavily through the whole thing — giving it context, having it generate sections, iterating on those sections with custom logic chains and framings i wanted for specific parts. standard workflow at this point. but then i started thinking about the person on the other end. someone who gets this paper and wants to understand it. they could use ai to target specific parts at different levels — a high-level overview, a mid-level dive into methodology, a deep dive into a couple of the results. the paper just needs to have the in-depth material there to reference. ai handles the rest. and that's when i realized the shift is bigger than just "ai helps you write faster." what's actually happening is this hyper-modularization of information at differing levels of complexity, because ai can just so easily switch between these and translate across them. it changes how you'd even think about crafting a document — you could almost throw out formalities and conventions and just lean fully into what would be most functionally useful for downstream processing. this sounds like a writing trick but it's actually pointing at something structural. the old model of information exchange looked like this: you create a document at one resolution. a research paper is dense. a policy brief is compressed. a presentation is skeletal. if you want five versions for five audiences, you do five times the work. each version is a fixed-resolution artifact — one level of detail, one assumed audience, one structure. the new model: you create a dense information substrate — the maximally detailed, structured version — and ai generates arbitrary projections of that substrate at whatever resolution a given audience needs. you write the substrate once. the projections are generated on demand and cost almost nothing. this is basically what happened to web design a decade ago with responsive design. you stopped building separate sites for mobile and desktop and started building one site that adapts. same shift, applied to information itself. but here's the part that i think most people are missing. this isn't just about efficiency. it's about what becomes valuable when information is this cheap. think about the old cost structure. acquiring information used to be expensive — you needed education, research access, experience, the right networks. distributing it was medium cost — writing, publishing, teaching. so society built itself around treating information as a scarce, high-value resource. "i know things" was a legitimate differentiator. credentials, degrees, expertise — all fundamentally certificates that say "this person has information you don't." ai just crushed that whole structure. acquiring information is now approaching free. distributing it is approaching free. repackaging it at different complexity levels is approaching free. the entire information layer of the economy just got commoditized. **so what's left? what's still scarce?** **logic. judgment. reasoning. the ability to look at a pile of information and decide what matters, why it matters, what's missing, what's wrong, and what to do about it. selection, framing, synthesis, evaluation — these are still expensive because they require genuine understanding, not just pattern matching over a corpus.** there's an analogy from thermodynamics that i think captures this well. information is like energy — abundant but diffuse. raw information everywhere, just like solar radiation hitting the earth's surface constantly. but energy doesn't do useful work on its own. you need a gradient — a structured difference — to extract work from energy. that's what an engine does. logic is the gradient. it's what converts raw, abundant information into useful intellectual work. without it, you just have heat — lots of data, no output. and there's a game theory angle too. in a world where everyone has access to the same information (because ai equalizes access), competitive advantage shifts entirely to who reasons better about what everyone knows. the game moves from "who knows what" to "who thinks better about what everyone knows." some work in mechanism design (myerson, maskin) formalizes how equilibria shift when information asymmetries collapse — the strategic landscape changes fundamentally when the information layer is no longer the bottleneck. what's wild to me is how miscalibrated most of society still is for this. the education system still mostly rewards information retention. "learn this material, reproduce it on the test." professional credentials still mostly certify "this person was exposed to information in a structured way." people still treat "i read about this" or "i know about this" as meaningful differentiators in conversations and careers. this is basically price stickiness applied to mental models. information used to be expensive, so people valued it highly. it's now cheap, but the mental model hasn't updated. it's like still paying 2005 prices for a phone call because you remember when long distance was expensive — except applied to how we think about knowledge itself. the person who reads 100 articles and forms an opinion is now less valuable than the person who reads 3 and builds a framework for evaluating which articles matter and why. the professor who lectures facts is less valuable than the one who teaches students how to evaluate, synthesize, and decide. the worker who "knows the system" is less valuable than the one who can redesign the system when conditions change. now i want to be fair to a potential objection here. information isn't worthless. you still need the lego blocks to build anything. the end product of good logic is still... information, in some sense. it's just information that's been run through selection and reasoning to become something more structured and useful than the raw inputs. so it's not that information doesn't matter — it's that *raw, unprocessed* information has lost almost all of its standalone value. the value has migrated to the processing layer. another fair objection: maybe this doesn't change much if most people can't actually do the logic part. and yeah — that's real. the gap between "information is cheap" and "everyone suddenly reasons better" is enormous. humans are slow to change. institutions are slower. it might take a generation for this to really permeate. but the people who get it now and adjust have a compounding advantage over those who don't, precisely because the shift is still underrecognized. the part that really excites me is the epistemic implication. if you take this seriously — that the constraint on intellectual progress has shifted from information access to reasoning quality — then the highest leverage intervention isn't producing more information. it's improving the logic layer. teaching people to reason better, building tools that scaffold judgment, designing systems that help humans do the selection-and-framing work more effectively. that's where the bottleneck is now. and it's a bottleneck most of the world hasn't even identified yet because they're still operating on the old model where information was the scarce thing. what makes this different from past information revolutions — the printing press, the internet, search engines — is that those shifted access. they made it easier to *find* information. this one is shifting something closer to the qualia of working with information. the actual felt experience of thinking, writing, researching, and building is different now. it's not just that i can google something faster. it's that the entire texture of intellectual work has changed — how ideas get generated, tested, recombined, and packaged. that's not an access revolution. that's a cognitive one. and because it's happening at the level of experience rather than just infrastructure, most people won't fully register it until they've felt it themselves. which means there's a window — right now, while this is still being assimilated — where the people who lean into these shifts and actually restructure how they think and work have a compounding advantage that's hard to overstate. not because they're smarter, but because they've updated their model of what intellectual work even is while everyone else is still using ai to do the old thing slightly faster. # My Research ## Can we get otherwise unattainable insights from human text and comments using AI? *2026-06-01 · upcoming · University of Utah · SPUR Summer 2026* Authors: Teddy Wright (University of Utah), Vineet Pandey (Professor, School of Computing, University of Utah), Sara Yeo (Professor, Communication, University of Utah) Exploring how AI can analyze human text and comments in research contexts. ### Abstract Abstract coming soon. --- ## How is AI permeating into the workforce and what is there to do about it? *2026-04-15 · in-progress · working paper · in progress · Utah Office of AI Policy* Authors: Teddy Wright (Utah Office of AI Policy), Alice Schwarze (Head of Research, Utah Office of AI Policy), Zach Boyd (Director, Utah Office of AI Policy · Professor, BYU Mathematics) Working paper measuring AI automation exposure across the U.S. workforce — which occupational tasks are technically substitutable, where exposure concentrates, and how that maps to wage value at risk. ### Abstract Abstract coming soon. --- ## Companion dashboard — How is AI permeating into the workforce and what is there to do about it? *2026-04-01 · in-progress · live dashboard · in development · Utah Office of AI Policy* Authors: Teddy Wright (Utah Office of AI Policy), Alice Schwarze (Head of Research, Utah Office of AI Policy), Zach Boyd (Director, Utah Office of AI Policy · Professor, BYU Mathematics) External: https://aea-dashboard-75lzrqja9-theodorewright11s-projects.vercel.app/explorer Companion dashboard for data exploration — interactive tool for digging into AI's task-level workforce exposure across occupations, work activities, and time. ### Abstract Companion dashboard to the workforce automation exposure paper. It measures how AI capabilities map onto the U.S. occupational task structure: for a given occupation, work activity, or job category, what share of the work could be affected by current AI systems — and how many workers and wage dollars does that represent. The dashboard triangulates across five independent AI scoring sources (Anthropic Economic Index conversation and API data, MCP server logs, and Microsoft Copilot exposure data) so no single methodology drives the result, and combines that with O*NET task structure and BLS employment and wage statistics. Eight pages cover occupation- and work-activity-level exploration, two-group side-by-side comparisons, time-series trends, and task-level diffs between dataset versions. Built for Utah's Office of AI Policy. Intended for policymakers, workforce analysts, and informed members of the public. The companion paper makes the substantive argument; this dashboard is the place to interrogate the underlying numbers. --- ## Why are people miscalibrated on AI? *2026-04-01 · published · University of Utah · AI + Ethics Workshop 2026 · University of Utah · AI + Ethics Workshop 2026* Authors: Teddy Wright (University of Utah), Valen Cole (University of Utah) Paper: https://theodorewright.dev/papers/AI_Miscalibration_Research.pdf A framework decomposing view formation into Optimization Target × (Information Quality × Processing Quality), with the Self-Concealing Property as key insight — identity-protective cognition is phenomenologically indistinguishable from rigorous skepticism. ### Abstract This paper develops a framework for understanding why people form miscalibrated views about artificial intelligence. We argue that miscalibration is not primarily an information deficit problem but an optimization target problem: individuals whose goal is identity preservation rather than accuracy deploy their cognitive capacity selectively, producing motivated reasoning that resists correction. We formalize this as a multiplicative model — View = Optimization Target × (Information Quality × Processing Quality) — and identify failures along each component, distinguishing between external factors that degrade information and processing independent of the individual and target-driven distortions that systematically bias both from within. We describe a critical property of identity-protective cognition, the self-concealing property, by which the bias is phenomenologically indistinguishable from the epistemic rigor it displaces, making standard interventions of "more awareness" insufficient. We further argue that miscalibration is compounded by misspecification of the object of the view itself — misattributing outcomes to AI rather than the human system, and treating "AI" as a single entity rather than a bundle of capabilities entering distinct domains. We then outline interventions organized around what organizations can change about the environment and what individuals can do given the limits of self-correction. --- ## Project Iceberg *2025-09-01 · contribution · working paper · Sep 2025 · MIT* Authors: Ayush Chopra (MIT · Project Iceberg), Santanu Bhattacharya (MIT · Project Iceberg), DeAndrea Salvador (NC General Assembly · Future Caucus), Ayan Paul (Project Iceberg), Teddy Wright (Utah Office of AI Policy), Aditi Garg (Project Iceberg), Feroz Ahmad (Project Iceberg), Alice Schwarze (Utah Office of AI Policy), Ramesh Raskar (MIT · Project Iceberg), Prasanna Balaprakash (Oak Ridge National Laboratory) External: https://iceberg.mit.edu/report.pdf Nationally recognized AI labor market simulation study, led out of MIT. Skills-centered exposure index built on a 151M-worker agent simulation. ### Abstract Artificial Intelligence is reshaping America's over $9.4 trillion labor market, with cascading effects that extend far beyond visible technology sectors. When AI automates quality control in automotive plants, consequences spread through logistics networks, supply chains, and local service economies. Yet traditional workforce metrics cannot capture these ripple effects: they measure employment outcomes after disruption occurs, not where AI capabilities overlap with human skills before adoption crystallizes. Project Iceberg addresses this gap using Large Population Models to simulate the human–AI labor market, representing 151 million workers as autonomous agents executing over 32,000 skills across 3,000 counties and interacting with thousands of AI tools. It introduces the Iceberg Index, a skills-centered metric that measures the wage value of skills AI systems can perform within each occupation. The Index captures technical exposure — where AI can perform occupational tasks — not displacement outcomes or adoption timelines. Analysis shows that visible AI adoption concentrated in computing and technology (2.2% of wage value, approximately $211 billion) represents only the tip of the iceberg. Technical capability extends far below the surface through cognitive automation spanning administrative, financial, and professional services (11.7%, approximately $1.2 trillion). This exposure is fivefold larger and geographically distributed across all states rather than confined to coastal hubs. Traditional indicators such as GDP, income, and unemployment explain less than 5% of this skills-based variation, underscoring why new indices are needed to capture exposure in the AI economy. By simulating how capabilities may spread under alternative scenarios, Project Iceberg enables policymakers and business leaders to identify exposure hotspots, prioritize training and infrastructure investments, and test interventions before committing billions to implementation. Iceberg is built with the AgentTorch framework. # Models ## How to parent optimally *2026-04-28 · draft* Confidence, stakes, reversibility, disclosure, etc. Draft. A four-axis framework for parenting decisions where the right move depends not just on expected value but on how reversible the decision is, how high the stakes are, how confident the actor is, and how much of the reasoning is disclosed to the affected party. --- ## Why certain harm and suffering gets more attention than others *2026-04-28 · draft* Vividness, recency, identifiability, narrative fit, etc. Draft. The premise: perceived harm is actual harm passed through a salience filter — vividness, recency, identifiability, narrative fit. Two harms of equal magnitude can produce wildly different moral weight depending on how legible they are to attention. Formalization and an interactive dashboard will land here once the literature scan is done. --- ## Why continuing is rational once you exist *2026-04-27 · published* Once you exist, you hold an option on future states. The asymmetry between starting and continuing is structural, not psychological. ## Intuition Once you're alive there's more rational incentive to keep living than there would have been to start living in the first place — and this isn't sunk cost bias but a real structural difference in the decision architecture. Once you're alive you hold an option — the right to keep experiencing future states. That option has value independent of current conditions because the future is uncertain and you can't re-enter once you exit. From outside existence, you don't hold the option, so the comparison is asymmetric for a structural reason, not a psychological one. ## Equations Pre-existence (no option value): `EV(exist) = G − B + w(E − C)` In-existence (option value added): `EV(continue) = G_future − B_future + w(E_future − C_future) + OV` The option value itself: `OV = ∫ max(future good states, 0) · f(states) d(states)` You integrate probability-weighted future good outcomes across all possible states; bad states contribute zero because you retain the choice to exit later. ## Variables - **G** — expected good experiences - **B** — expected bad experiences - **E** — positive externalities generated for others - **C** — costs imposed on others - **w** — weight on others' utility relative to your own (0–1) - **OV** — option value ## Try it ## Key insights - OV is only zero when the future is perfectly certain to be uniformly bad — almost never true. - Depression artificially truncates the right tail of future good states, collapsing *perceived* OV without changing actual OV. - Pre-existence EV can be negative while in-existence EV is positive — entirely because of the OV term. - This vindicates the asymmetry without appealing to sunk cost. ## Consent-adjusted natalism extension `EV(create life) = p · k · (G − B + w(E − C)) + (1 − p)(−B)` where `p` is the probability the person would endorse their own existence, `k ∈ [0,1]` is a quality-of-circumstances scalar, and `(1 − p)(−B)` is the moral penalty for imposing harm on a non-consenting subject. Anti-natalism and pro-natalism disagree specifically on: whether `G − B` is reliably negative in bad circumstances, what `p` actually is empirically, and how much unique irreplaceable value each life generates. # Updates ## Site scaffold + research migration *2026-04-28 · weekly* Got the site structure into place: content collections for research, models, AI Research pipeline, writing, and these updates. First polished model (Option Value) live. This week was mostly site infrastructure. Stood up content collections for research, models, AI Research (the LLM Iterate pipeline), writing, and updates. Migrated existing research entries off hardcoded markup. Shipped the first interactive model — Option Value — with sliders for G/B/E/C/w/OV. Next week: simplify front pages, get the AI miscalibration paper PDF wired up, start the first AI Research stage on social-output-efficiency.