Theory of Change
Orthogonal pursues "formal-goal alignment": designing a fully formalized mathematical goal that would produce good outcomes when maximized by any optimizer, including a superintelligent one, then building an AI that pursues that goal. Founder Tamsin Leake:
"You do not align AI; you build aligned AI... I do not expect that current AI technology is of a kind that makes it easy to 'align'; I believe that the whole idea of building a strange non-agentic AI about which the notion of goal barely applies, and then to try and make it 'be aligned', was fraught from the start."
The approach explicitly rejects prosaic alignment (interpretability, RLHF, scalable oversight) as "net harmful when published" -- claiming these methods don't scale to superintelligence and create capability externalities. Instead, Orthogonal proposes building an entirely new AI system from scratch with alignment built into its mathematical foundations.
The concrete research agenda is QACI (Question-Answer Counterfactual Interval): a scheme where an AI scores actions by asking counterfactual questions about information blobs located across possible computational universes. A more recent direction is ESP (Epistemic State Prior), proposed as an alternative to Solomonoff-style priors for avoiding demonic hypotheses.
The org explicitly acknowledges very low success probability: "we still mostly die. I do not expect that our plan saves most timelines... but we will have significantly increased the ratio of worlds that survive."
What They Do
Orthogonal has produced approximately 6 public posts over its roughly 3-year existence, plus an unknown number of deliberately unlisted posts (hidden for exfohazard reasons). The core technical output is:
- QACI formalization paper (Jun 2023, with Julia Persson): Full mathematical specification of the QACI alignment goal using set theory, quantum Turing machines, blob location functions, and counterfactual insertion.
- Epistemic State Prior paper (Aug 2024): Proposes modeling a user's epistemic beliefs rather than using computational priors, to avoid malign Solomonoff induction. Described by the author as "pseudomath" and "sketches."
- Strategy/philosophy posts: Theory of change explanation, LDT application to arms race dynamics, binary AI outcome framing.
Many posts are deliberately unlisted to prevent capability exfohazards. The QACI table of contents references multiple inaccessible posts. A Discord server (~296 members) may contain unpublished discussion.
No publications since August 2024. Last LessWrong comment by Tamsin Leake was December 2024 (about PauseAI, not Orthogonal research). No public update on organizational status. The org may have quietly wound down.
Key People
Tamsin Leake -- Founder and sole confirmed full-time member. Former indie game developer (France). Went through Refine incubator (Conjecture, Aug-Oct 2022). Online handle: "carado."
The founding announcement (Apr 2023) claimed "several promising researchers intending to work fulltime." Only Julia Persson (co-authored one paper) and "mesaoptimizer" (editorial assistance) have been publicly identified. No evidence the planned team materialized. The announced fellowship was never publicly launched.
Money and Incentives
Total confirmed funding: $30,000. This is the only publicly verifiable funding Orthogonal has received.
| Source | Amount | Date | Notes |
|---|---|---|---|
| LTFF grant (to Tamsin Leake personally) | $30,000 | Feb 2023 | "6 months research stipend" |
| Every.org donations | Unknown | Ongoing | Tax-deductible via fiscal sponsorship |
- No Coefficient Giving / Open Philanthropy grants (zero across 480+ AI safety grants in their database)
- No Survival and Flourishing Fund grants
- No follow-on LTFF grant in the May 2023-March 2024 payout period
- No 990 financial data (uses Every.org fiscal sponsorship, not own nonprofit)
- No corporate or lab funding ties -- completely independent but also completely unsupported
- No venture capital, government contracts, or compute credits
Legal structure: Fiscal sponsorship under Every.org (EIN 61-1913297). Note: "Orthogonal Productions Inc" (EIN 88-2978002) is a Minneapolis theater company, not this organization.
The $30K stipend covered approximately 6 months at a modest rate. How the work was funded after August 2023 is unknown. France-based cost of living may have extended runway.
What Others Say
Dan MacKinlay (independent researcher): "QACI remains speculative and hard to parse. There are a lot of axioms to buy into to make it look even remotely feasible. It's not obvious how to connect 'blobs' to real-world referents, or why this formulation really sidesteps Goodhart's Law. Even insiders hedge on whether it's the right line to pursue."
Anonymous LW commenter comparing QACI to HCH: Identified shared failure modes including need for inner alignment, risk of "Chinese whispers" fidelity loss, convergence problems across iterations, memetic selection pressure, amplified failure probability, and vulnerability to acausal tampering.
Community engagement: A comprehensive 2022-2023 survey of alignment approaches (75+ hours of work covering MIRI, ARC, Anthropic, DeepMind, Redwood, Conjecture, and others) does not mention Orthogonal, QACI, or Tamsin Leake at all. This reflects near-zero visibility in the broader alignment community.
Manifold Markets: "Will QACI turn out to be a viable alignment plan?" prediction market shows roughly 20-51% probability (varying by market version), reflecting deep ambivalence.
No Refine evaluator (Steve Byrnes, Vanessa Kosoy, Evan Hubinger, Ramana Kumar, John Wentworth) has publicly commented on Tamsin's work. No MIRI researcher has publicly engaged with QACI despite Orthogonal explicitly claiming to share MIRI's worldview and seeking collaboration.
What's Absent
- 19+ months with no publication (last: Aug 2024). No public explanation of status.
- Team never materialized as described in the founding announcement. Remains effectively a solo project.
- Fellowship never launched despite being prominently announced.
- MIRI collaboration never happened despite being a stated goal.
- No peer review of the QACI formalization by independent mathematicians, despite the approach's entire value proposition being mathematical rigor. Tamsin self-describes as "bad at math."
- No endorsement from any major AI safety funder or established alignment researcher.
- No comparison with davidad/OAA, the other major "formal alignment" approach which has received ARIA funding.
- Total funding received is opaque -- no 990s, no annual reports, no financial disclosures.
Recommended Reading
Tamsin Leake's LessWrong profile and comments (https://www.greaterwrong.com/users/tamsin-leake) -- Most candid window into the founder's thinking. Detailed QACI explanations, exfohazard views, PauseAI advocacy, criticism of AI labs.
Dan MacKinlay's agent foundations assessment (https://danmackinlay.name/notebook/agent_foundations.html) -- The strongest independent external critique of QACI.
Orthogonal's theory of change (https://www.alignmentforum.org/posts/4XcADCLDDguyej2N7/orthogonal-s-formal-goal-alignment-theory-of-change) -- The core document explaining why this org exists and what it rejects about mainstream alignment.
Formalizing the QACI alignment formal-goal (https://www.alignmentforum.org/posts/MR5wJpE27ymE7M7iv/formalizing-the-qaci-alignment-formal-goal) -- The actual technical output; form your own view of the math.