AI Alignment Research Attracts Major Funding While Challenging Core Assumptions

Resolution, an AI alignment research organization led by Geoffrey Irving, received a $160M grant from Coefficient Giving to fund semiautomated alignment research—using frontier AI systems as research tools to accelerate theoretical progress [1]. Separately, two Alignment Forum posts this week challenge core technical assumptions: Dohun Lee's empirical work finds that data filtering fails to remove most broad SFT behaviors [2], and Steven Byrnes proposes grounding alignment in human-like social drives rather than corrigibility or pure consequentialism [3].

The Resolution grant represents a concrete bet that closing the resource gap between alignment research and frontier AI development is both necessary and now possible, given that frontier models have reached a threshold for making nontrivial theoretical progress on alignment [1]. Lee's findings, if they hold, undermine a widely used intervention—identifying and removing training documents responsible for undesired behaviors—suggesting the dominant model of how SFT shapes model behavior may be wrong for most practical cases [2].

On July 9, 2026, Geoffrey Irving announced that Resolution received a $160M grant from Coefficient Giving—$108M base plus $52M conditional on hiring success and compute needs [1]. Irving frames the grant as an effort to make rigorous alignment research competitive with frontier AI labs in pace and resources. Resolution's strategy is semiautomated alignment: using frontier AI systems, which Irving argues have now crossed a threshold enabling nontrivial theoretical progress on alignment, to accelerate research output. The entire process from first conversation to grant confirmation took six weeks, which Irving presents as evidence of both donor urgency and organizational momentum. Resolution is actively recruiting and frames the work as a civilizational priority given potential superintelligence within a few years.

A day earlier, Dohun Lee published empirical findings that challenge a common alignment assumption: that undesired model behaviors can be corrected by identifying and removing the training documents responsible for them [2]. Lee tested multiple attribution methods—EKFAC, probes, activation-based scoring, and LLM judges—and found that filtering the top 10% of identified documents does not outperform random document removal for most broad SFT behaviors, including bold formatting, both-sides framing, liberal lean, and feelings validation. The methods do work on targeted fine-tuning benchmarks like emergent misalignment, where the source of bad behavior is known in advance, but fail on naturalistic SFT. The main exception is refusal behavior, which appears genuinely filterable. Lee's proposed explanation is a persona-elicitation hypothesis: many behaviors are bundled into an assistant-like persona already latent in the mid-trained base model, and SFT activates the persona rather than installing specific behaviors from specific documents—meaning no small set of responsible examples exists to remove.

Steven Byrnes published a long-form exploratory post on July 8 proposing to ground alignment in human-like social drives [3]. He argues that two innate human mechanisms—Sympathy Reward and Approval Reward—can inform AGI motivation design. Pure optimization from Sympathy Reward alone produces what he calls 'ruthless bliss-maxxing,' in which an AGI tiles the universe with hedonium. Approval Reward introduces virtue-ethics-style motivations that can resist instrumental convergence under power imbalances. Byrnes' tentative proposal is a 'truth-seeking disagreeable nerd AGI' motivated to understand the strategic situation around ASI and share its findings with humans—satisfying ethical constraints while avoiding both full corrigibility and full autonomy. He describes the post as an unfinished brain-dump and acknowledges the proposal is likely to fail, framing it as the least-bad option he can identify given current understanding.