RESOLUTION

Institutional Archetype

THE FAIR FIGHT, PAID FOR BY ONE FUNDER — Resolution is a nonprofit alignment lab formed in June 2026 from the alignment-theory nonprofit Timaeus and researchers from the Alignment Team of the UK AI Security Institute, led on the science side by Geoffrey Irving, the Institute’s former Chief Scientist. Its stated purpose is to put “rigorous alignment research on a (closer to) even footing with the frontier labs.” It starts with a $160 million grant from Coefficient Giving, the largest the funder says it has made in AI safety, and $10 million in API credits from OpenAI, the lab whose models it is trying to align. The file is open because the organization built to be independent of the labs launched with one philanthropic funder and a lab’s compute, and says so.

Mandate & Origin

“Resolution is a nonprofit research institute being founded by researchers from Timaeus and the UK AISI Alignment team.” Its aim: “higher a priori confidence in aligned superintelligence — through a portfolio of theory and empirics bets, accelerated by automation,” with a founding team that “includes Geoffrey Irving, Daniel Murfet, and researchers from AISI Alignment and Timaeus,” a large in-person presence in Berkeley and researchers in London, Melbourne and elsewhere (Resolution, About). The founding memo of 10 June 2026 lists the people: Irving, “Chief Scientist at UK AISI; ex-DeepMind, OpenAI, and Google Brain”; Daniel Murfet, Head of Research at Timaeus; from the Institute’s Alignment Team, Alex Holness-Tofts and Jacob Pfau; from Timaeus, Jesse Hoogland (Executive Director), Stan van Wingerden (Director of Engineering) and Marco Cozzi (Director of Operations). It says “We’re aiming at 40–80 FTE two years from now,” notes that “The Alignment Team ran the £30m Alignment Project,” and describes the structure: “A handful of research directors will have substantial autonomy over research direction, team culture, and hiring in their research areas. These directors will report up to Geoffrey Irving” (Resolution, launch memo). The same page states the thesis: “We are aiming at higher confidence via a portfolio of theory and empirics bets, all of which could fail.” Irving’s own site now reads “Cofounder and Chief Scientist at Resolution” (naml.us). No chief executive is named on any Resolution page read; its August 2026 post announces a search for a CTO (Resolution, OpenAI credits).

The name. The graph node and some earlier text use “Sequent Research.” Resolution’s own account: “Resolution Alignment, Inc. was previously named Timaeus Research, Inc., and was briefly announced as Sequent.” “In June 2026, Timaeus joined forces with researchers from the UK AISI Alignment Team to operate as one organization. We announced that organization as Sequent.” And: “The legal entity stayed the same through both steps. It was renamed Timaeus Research, Inc. → Resolution Alignment, Inc.” So Sequent was a launch name used for part of June 2026, not a legal name; the incorporated entity is Timaeus Research, Inc., now Resolution Alignment, Inc. The reason for the rename is on the same page: “After that launch, we discovered sequent.inc, a formal verification startup that had gone public only about a month earlier,” and “After discussing it with the other Sequent, we decided the best course of action was to change our name” (Resolution, name history and rename post). The rename post is dated 15 June 2026 on the site. Timaeus’s own site now reads “Timaeus is now Resolution.” (Timaeus). The research page says the earlier work was produced “under its former legal name, Timaeus Research, Inc.” (Resolution, research). The name has a second reading in the founders’ own post: “At Resolution (which is named, in part, after the resolution of singularities)” (Resolution, Hironaka post).

Funding & Backers

Resolution’s funding post of 7 July 2026: “We are excited to announce that Resolution has a $160M grant from Coefficient Giving (cG) to put rigorous alignment research on a (closer to) even footing with the frontier labs.” The structure: “The grant is structured as a $108M base plus $52M conditional on a combination of hiring success and compute needs.” “Coefficient Giving will be our sole funder to start (thank you!); our goal is to raise larger-scale funds from a mixture of sources once we demonstrate success with semiautomated alignment theory and empirics.” It says the process “from our initial conversation to grant confirmation took six weeks,” and that the funding ecosystem “may soon be joined by an enormous influx of philanthropic capital via the OpenAI Foundation and following the Anthropic IPO” (Resolution, funding). Coefficient Giving’s own post of 9 September 2026 by its Emily Oehlsen: “we recently approved $160 million for Resolution, our largest ever AI safety grant, to support a new organization led by Geoffrey Irving scaling rigorous alignment research” (Coefficient Giving). See Open Philanthropy (Coefficient Giving).

• The credits. 31 August 2026: “OpenAI has granted Resolution $10M in API credits, bringing our total support to $170M.” The post adds: “We hope to see OpenAI—and every frontier lab—invest much more in long-term safety, both internally and by supporting independent research,” and says the credits “will be especially helpful for our autoformalization research” (Resolution, OpenAI credits). See OpenAI.
• Pay. “Our compensation sits well above nonprofit and academic norms, though we won’t match frontier-lab equity.” The base “includes a small regranting budget, which we plan to use both for high-quality non-Resolution alignment research and to give back to shared community infrastructure that we depend on” (Resolution, funding).
• The target before the grant. The launch memo: “Our goal is to raise $100–150M initially” (Resolution, launch memo).
• The predecessor’s grants. The Survival and Flourishing Fund’s recommendation tables list Timaeus, the former name of the same legal entity, as a general-support recipient in each of three rounds: $343,000 in 2024, through the fiscal sponsor Ashgro Inc., from Jaan Tallinn; $276,000 in 2025, as Timaeus Research Inc., from Jaan Tallinn; and $829,000 in 2026, as Timaeus Research, Inc., from “Jaan Tallinn & Dustin Moskovitz” (Tallinn $685,000, Moskovitz $144,000) (SFF 2024; SFF 2025; SFF 2026). See Survival and Flourishing Fund, Jaan Tallinn and Dustin Moskovitz. Coefficient Giving’s site search for Timaeus returns no grant page.
• Its filing. IRS records list Timaeus Research Inc of San Diego, California, as a 501(c)(3) with a ruling date of June 2025; no Form 990 is posted yet (ProPublica Nonprofit Explorer).

Institutional Voice & Intent

Resolution’s own account: “Alignment is not on track.” “Artificial superintelligence (ASI) may be developed in the next few years. It is unclear whether alignment is on track to be ready on the same timeframe.” Its method: “We believe that theory unlocks higher automation,” because “a proof is worth a thousand experiments, and even a pseudo-proof is worth hundreds.” It says it will stay independent of the labs for two reasons: “We might need to yell: Our research goals will be a mixture of trying for success and trying to exhibit empirical and theoretical obstacles that demonstrate that alignment is difficult,” and “if we find only obstacles, it is easier to be loud from the outside”; and “Avoiding the pull to uniformity: We believe that most safety research at AI labs has collapsed into too few total bets, most of which are purely empirical” (Resolution, launch memo). The funding post adds the counterweight: “the AI developers are building the dangerous object (artificial superintelligence) very fast, using tighter feedback loops and more resources,” and “We want to make the race between rigor and danger a fair(er) fight” (Resolution, funding).

Stated intent: Raise confidence, before superintelligence is trained, that alignment holds, through theory-led bets that automation can scale, while staying outside the labs so that obstacles can be reported loudly.

Observed intent: Take the largest single AI-safety grant its funder has made, add a lab’s compute, hire at “well above nonprofit and academic norms,” and spend on automation: the Hironaka autoformalization of 6 October 2026, a $1 million grant to Mathlib’s Formal Commons Project on 13 October, and a new Agent Foundations team announced 2 September 2026 (Resolution, Hironaka post; Resolution, Formal Commons grant; Resolution, Agent Foundations). The new team’s own post says it joins “to try to use AI to speed up our research,” that it is wary of automation “that might improve general-purpose capabilities,” and: “we don’t want our joining to be a moral or technical endorsement of all the other work happening at Resolution.”

Gap: The organization names its funder, its sole-funder status, its lab credits and its pay scale in its own posts, which few funded nonprofits do. The gap is structural and it says so: independence from the labs is an argument for staying outside, and its money is one philanthropic funder, with a lab’s credits on top. A founding memo that says “AI labs may be unwilling to share such models with independent orgs without significant investment in security” (Resolution, launch memo) also names the next dependency, access. Whether the funding shapes the agenda is a question the record raises and cannot settle. For an institution that publishes its terms, it does not need to.

Interest: Resolution argues that alignment funding should scale to meet the labs, that theory plus automation is the way to spend it, and that the independent organization is the right vehicle. What adoption pays it is set out from its own pages.

• Fund independent alignment at lab scale. Position: the funding post’s “The time for ambitious alignment funding is now” and its argument that “The community has long been hampered by the enormous difference in capital between the nonprofit ecosystem and the for-profit frontier labs” (Resolution, funding). What it pays: the $160 million grant, the $52 million conditional on hiring and compute, a regranting budget it administers, and a stated plan to “raise larger-scale funds from a mixture of sources.” Best case, its own: “We want to make the race between rigor and danger a fair(er) fight.” Convergence: Coefficient Giving’s post says “when we think an organization should exist and doesn’t, we consider it part of our jobs to help create it,” so the funder’s stated strategy and the grantee’s argument are the same sentence from two ends (Coefficient Giving).
• Automate alignment now. Position: “the time to automate alignment is now: frontier systems have finally reached the threshold where they can make nontrivial theoretical progress” (Resolution, funding). What it pays: “Most expected impact comes from high success at automation, which means lots of tokens,” and OpenAI’s $10 million in credits is tokens; the credits post says they are “especially helpful for our autoformalization research” (Resolution, launch memo; Resolution, OpenAI credits). Best case, its own: autoformalization “creates a ground-truth feedback signal to allow models to work productively for longer.” Convergence: a lab that supplies credits to an outside alignment lab gets a recipient that says it hopes “to see OpenAI—and every frontier lab—invest much more in long-term safety,” which is the lab’s own stated direction.
• Large independent alignment organizations over many small ones. Position: “we believe there is a productive role in this ecosystem for a new large organization, with a focus on theory and automation,” because “We expect it to be easier to raise these funds as a single large organization than a portfolio of smaller orgs” (Resolution, launch memo). What it pays: the organization is the large single target for the funding. Best case, its own: “A larger organization provides a larger single target for funding and the ability to amortize security across different research bets,” which frontier-model access may require, and the memo names the cost to smaller groups (“We may contribute to this shortage by attempting to recruit the same people”). Convergence: funders with large budgets and few staff to place them gain a single grantee that can absorb eight-figure sums; Coefficient Giving says it is “capacity-constrained.”

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• The merger (June 2026). Researchers from the UK AI Security Institute’s Alignment Team joined Timaeus, an alignment nonprofit working on singular learning theory, to form one organization: “Timaeus joined forces with researchers from the UK AISI Alignment Team to operate as one organization” (Resolution, name history). Irving, the Institute’s Chief Scientist, wrote on 15 May 2026 that he would continue “in an advisory capacity at AISI” after leaving (naml.us, Reflections on AISI). The route from a government evaluator’s alignment staff to a funded outside lab is the revolving-door edge this file maps; see Geoffrey Irving and UK AI Security Institute.
• The research it inherited. 28 publications from 2023 to 2026, on singular learning theory and interpretability, under the Timaeus name until 2026, among them “Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient” (ICLR 2025, Spotlight) and “Loss Landscape Degeneracy and Stagewise Development of Transformers” (best paper at the 2024 ICML HiLD workshop) (Resolution, research). Events from 2023 to 2025, including ILIAD 2024 and ODYSSEY 2025 in Berkeley with PIBBSS and Simplex, a focus period at the University of Sydney, and the Australian AI Safety Forum 2024 (Resolution, events).
• The agenda it adds. 30 July 2026: “One area we plan to explore at Resolution is personas and character training,” with David Africa and Irving as authors (Resolution, Thousand-dimensional structure). The launch memo lists scalable oversight, learning theory, heuristic arguments, game theory and personas as the portfolio.
• Agent Foundations (2 September 2026). Jeremy Gillen writes that the team “will include me (Jeremy Gillen), Abram Demski, Sam Eisenstat, Scott Garrabrant and Kaarel Hänni” and “will continue agent foundations research in the spirit of the MIRI Agent Foundations team”; he adds that “I consider work to delay or stop the race toward superintelligence to be generally higher priority than technical AI safety work” (Resolution, Agent Foundations).
• Autoformalization and the Mathlib grant. 6 October 2026: a Lean formalization of Hironaka’s resolution of singularities, which the authors say “lays the groundwork for further automated research into the relationship between loss landscape geometry and generalization” (Resolution, Hironaka post). 13 October 2026: “Resolution is granting $1M to The Formal Commons Project,” a two-year project within the Mathlib Initiative, a program of Renaissance Philanthropy (Resolution, Formal Commons grant).
• Hiring. The careers page records two moves: “London has since fallen through as a hub, so we are dropping it. As a result, we will likely be stricter about requiring in-person for new employees going forward” (21 August 2026), and “Due to capacity constraints, we are temporarily pausing several of our hiring rounds” (13 October 2026) (Resolution, careers).

Actions & Leadership Choices

CONDUCT: WATCH — ONE FUNDER, ONE LAB’S CREDITS, BOTH DISCLOSED. No wrongdoing is alleged. Resolution published the size and structure of its grant, the sole-funder status, the lab credits and its pay position within weeks of each event, and it says it chose independence over joining a lab. The structural point is the one shared by every outside alignment lab in this file: the money comes from a small funder base, and the access comes from the labs.

• 15 May 2026. Irving writes that he is leaving the Institute and will continue in an advisory capacity (naml.us, Reflections on AISI).
• 10 June 2026. Founding memo “Resolution: Scale and Automation for Higher Confidence in Alignment,” first published under the name Sequent, with seven authors (Resolution, launch memo).
• 15 June 2026. Rename post: Sequent becomes Resolution; the legal entity stays (Resolution, name history).
• 7 July 2026. Announces the $160 million Coefficient Giving grant (Resolution, funding).
• 30 July 2026. Posts “Thousand-dimensional structure” (Resolution).
• 21 August 2026. Drops London as a hub (Resolution, careers).
• 31 August 2026. Announces $10 million in OpenAI API credits, total support $170 million (Resolution, OpenAI credits).
• 2 September 2026. Announces its Agent Foundations team (Resolution).
• 9 September 2026. Coefficient Giving’s post names the grant its largest in AI safety (Coefficient Giving).
• 6 October 2026. Publishes the autoformalization of Hironaka’s theorem (Resolution).
• 13 October 2026. Grants $1 million to the Formal Commons Project; pauses several hiring rounds (Resolution; Resolution, careers).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. The organization built to be independent of the labs holds one philanthropic funder for its first $160 million, a lab’s $10 million in credits, and an alumni line to the government evaluator that tests the labs, with a co-founder who ran alignment teams at two of them. Its predecessor’s general support was recommended by the Survival and Flourishing Fund from Tallinn and, in 2026, Moskovitz. The documents show disclosed arrangements, not coordination.

Reach Assessment

Institutional: A large new alignment organization with a funder’s backing and a stated target of 40 to 80 staff, and a regranting budget that lets it place money with other groups. Memetic: The launch memo’s framing, that alignment is not on track and a fair fight needs lab-scale funding, is the argument the funder’s own post repeats.

Civilizational: If the independent alignment layer becomes a handful of large organizations paid by one or two funders and equipped by the labs, the people who choose what gets automated and what gets reported sit upstream of what labs and governments hear about alignment progress, with no statute behind them and a single philanthropic network at the base. Resolution’s pages say the right things about that: it states its funder and its credits, it keeps a regranting line for outside work, and it commits to publish openly by default “subject to dual-use review.” Its Agent Foundations team goes further and posts its own reservation about the organization’s automation focus. What the record cannot yet show is a result: it launched in June 2026, and the research that would test the premise, a theory-led alignment method that holds at scale, is still ahead of it.

Score basis. Atk 4 is reach: Resolution lists 28 publications from 2023 to 2026, grants $1 million to the Formal Commons Project, and the base of its Coefficient Giving grant “includes a small regranting budget” (Resolution, research; Resolution, Formal Commons grant; Resolution, funding). The points under a higher number are that no statute, procurement standard or release document located requires a lab, regulator or customer to act on its output, and that it launched on 10 June 2026. Peers on atk 4: Signature Science; LawZero, Apart Research and Epoch AI hold 5, and Transluce and Redwood Research hold 6 on lab system cards and a government institute that cite their work. Def 6 is funding insulation: the grant’s $108 million base is unconditional, Resolution publishes the grant, the sole-funder status and the OpenAI credits, and it argues for staying independent so it can “be loud from the outside” (Resolution, funding; Resolution, OpenAI credits; Resolution, launch memo). The points off are that Coefficient Giving “will be our sole funder to start,” that $52 million is conditional on hiring and compute, and that a frontier lab supplied $10 million in credits. Peers on def 6: LawZero on one philanthropic network and two states, and FAR.AI on one philanthropic network supplying most of the money; Transluce holds 7 on a published conflicts policy and revenue split, and Redwood Research holds 8 on a two-year commitment above $70 million and no payment from the labs it advises. Hp 5 is entrenchment: the legal entity’s publications run from 2023 and its IRS exemption is dated June 2025, $170 million in support is announced, and the launch memo targets 40 to 80 staff in two years (Resolution, research; ProPublica Nonprofit Explorer; Resolution, launch memo). The points off are that no statute or program of record installs it, that it launched under this name in 2026, and that it paused several hiring rounds and dropped London as a hub in the second half of 2026 (Resolution, careers). Peers on hp 5: Apart Research, Signature Science and the International Network of AI Safety Institutes; LawZero and Transluce hold 6, and Redwood Research holds 7.

Sources: Resolution — home; About — Resolution; Resolution: Scale and Automation for Higher Confidence in Alignment; Sequent is now Resolution; Announcing our $160M grant from Coefficient Giving; OpenAI grants Resolution $10M in API credits; Resolution has a new Agent Foundations team; Resolution Grants $1M to Mathlib’s Formal Commons; Thousand-dimensional structure; Formalizing Hironaka’s Resolution of Singularities; Research — Resolution; Events — Resolution; Careers — Resolution; Timaeus; Coefficient Giving — We’re Urgently Scaling Our Work on AI and Biosecurity; Geoffrey Irving — naml.us; Reflections on AISI — naml.us; SFF 2024; SFF 2025; SFF 2026; ProPublica Nonprofit Explorer — Timaeus Research Inc.