Response Synthesis

Strategic Direction

Research focus and pivotal questions

Respondents cluster around two broad areas, with a real tension between them on GCR-relevance versus fit with UJ's evaluative capacity.

The AI × economics/labor cluster attracted the most convergence. Pacchiardi endorses the AI-economics interface as the most relevant area"AI-economics interface (eg impacts on labour market [Anthropic's work] or macro-trends à la Ord) seems the most relevant area"— Lorenzo Pacchiardi, Garavaglia points to workplace impacts and skills retraining, Du to labour welfare and AI literacy/inequality, and Habermacher wants "realpolitik" regulatory frameworks for labour impacts"politically grounded (I mean 'realpolitik' type not 'purely ivory tower abstract') Regulatory Frameworks for dealing with labor impacts"— Florian Habermacher. DR asks what Habermacher means concretely: what research, methods, producers, and decision-relevance are in scope? This needs unpacking before UJ commissions here. Tagat observes there is already substantial work on labor impacts — DR emphasizes this is a feature, not a bug: rich pipelines to evaluate is precisely what UJ wants. Kao suggests broadening to include high-effort informal pieces like Trammell/Patel's "Capital in the 22nd Century" or Citrini's "2028 Global Intelligence Crisis." DR flags the open question: does this cluster actually get at the highest-GCR / highest-impact issues?

The governance/risk-parameters cluster is more GCR-adjacent. Kalkar identifies gaps in regulatory interventions and comparative international governance analysis"gap in regulatory interventions, comparative international governance analysis, and estimates of risk parameters for AI safety and governance"— Uma Kalkar, with pivotal questions like which governance mechanisms actually change frontier lab behavior. Schwab pushes for middle-power regulatory strategies and lessons from arms control — e.g., Chinese cooperation patterns in adjacent risk domains. DR wants concrete examples: risk-parameter estimation is quantitative and comfortable for UJ, but IR-style middle-power analysis would be a meaningful move outside our empirical comfort zone. Tagat suggests engaging AI/catastrophic-risk funders (e.g., Schmidt) to openly evaluate work they commission.

Staying timely

Several complementary process proposals emerged. Manheim wants very fast turnaround with a single-evaluator fast track"Fast track, 1+ evaluators, invites to more than 1 reviewer and move forward as soon as 1 is submitted"— David Manheim. Pacchiardi proposes pre-booking evaluator time before paper selection"'pre-booking' evaluator's time before deciding what paper to review so that you can choose a paper and are sure that someone will be able to look at that in a timely manner"— Lorenzo Pacchiardi and paying more for flexibility — DR strongly endorses this and notes it connects to giving evaluators some limited choice. Pacchiardi also suggests crux-targeted reviews (~1 hour of evaluator work per specific claim). Kao proposes ACX/Zvi-style commentary roundups as a complement to full evaluations"'commentary roundups' that presents clusters of comments made by others online + light discussion could be useful"— Andrew Kao. Kalkar cautiously endorses LLM-assisted prioritization but warns against over-automation. Garavaglia favors one evaluator plus AI-assisted briefing. Tagat suggests an NBER-track-style pipeline with faster turnaround.

Who to bring in

Named individuals: Jonathan Prunty (Leverhulme CFI, Cambridge) and Marko Tesic (DSIT UK) for labor-market impacts. Institutional networks: RAND TASP Fellows and current/former GovAI Fellows (Kalkar), BlueDot (Manheim), the ILO for labor welfare (Du), and AI/catastrophic-risk funders like Schmidt (Tagat). Tagat also endorses circulating a call in the AI Slack for crowdsourced ideas.

The Ord RL Scaling Evaluation

Should we still pursue it?

Respondent views ranged from "moment passed""Moment passed."— Josephine Schwab (Schwab) and lukewarm (Manheim, Habermacher, Pacchiardi still on the fence) to actively supportive: Kalkar argues it may make sense to be a first-mover now that the empirical picture is contested"Given that the empirical picture has become messier and more contested, it may actually make sense to be a 'first-mover' and evaluate it now"— Uma Kalkar, especially given Ord's reframing to 2027+.

Latest development (Apr 29): Toby Ord has responded to DR's outreach and this substantially updates the picture. Ord explicitly endorses more rigorous analysis ("I'd be very happy for there to be more formal analysis of these questions"), but won't participate personally — a non-participant, not an objector. Critically, he intended to write a wrap-up essay for the EA Forum Scaling Series but won't get around to it — a concrete gap a UJ synthesis could directly fill.

Most importantly, Ord clarified his actual quantitative position, which is notably more modest than headline readings suggest. In his framing, speed was roughly 60% compute / 40% other in the GPT-2/3/4 era; roughly half the compute boost is going away, leaving progress at ~30% + 40% = 70% of its old speed unless other contributors (algorithms, data, RL environments, recursive self-improvement) compensate. This is not a claim that RL has hit a fundamental limit — it is a claim about a declining relative contribution.

This reopens the case for evaluation on three grounds: (1) Ord now explicitly endorses formal analysis; (2) there is a specific, evaluable quantitative claim on the table; and (3) the gap between Ord's actual position and how it has been characterised is itself worth documenting. The "moment passed" view (Schwab, Habermacher) warrants revisiting.

If yes — what form?

Kalkar's proposal fits the updated situation well: a long-form EA Forum/PubPub post separating what is established, contested, and open"long-form EA Forum/ PubPub post. It should cover the full evidence base... with clear separation between what's empirically established, what's contested"— Uma Kalkar, with a draft circulated across the GovAI community. Given Ord's non-participation but endorsement, this could explicitly serve as the wrap-up he won't write.

Points of apparent consensus

• UJ should stay in AI governance and AI-social/economic impacts rather than expand into technical AI safety — DR takes this to heart but is not fully convinced the space is covered; the Alignment Journal is domain-limited to alignment, not broader technical safety.
• Speed matters: single-evaluator fast tracks, pre-booked evaluator time, and crux-targeted micro-reviews are all worth piloting.
• Named networks to tap: GovAI (fellows and community), RAND TASP, BlueDot, plus the AI Slack channel for crowdsourcing.
• AI assistance for briefing/prioritization is welcome, but should not replace human judgment.

Key tensions worth discussing

• GCR-relevance vs. wheelhouse fit: AI×labor/economics is where UJ is strongest, but DR asks whether it addresses the highest-impact questions. Governance and risk-parameter work is more GCR-relevant but stretches UJ's methodological comfort zone.
• Technical safety expansion: Respondents are strongly opposed, but DR is not yet convinced the rapid-credible-evaluation niche is filled. Manheim's suggestion of peer-reviewing model cards and technical reports would in fact pull UJ toward technical safety — an internal contradiction to resolve.
• Ord evaluation timing: Ord's April 29 clarification and endorsement materially updates the earlier "moment passed" view; the synthesis could now serve as the wrap-up essay Ord himself won't write.
• Realpolitik regulatory work and IR-style middle-power analysis: promising but under-specified — need concrete examples and clear decision-relevance before commissioning.