Category intelligence

Research Briefing — May 23, 2026

11 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research and commentary cluster around post-AGI economics, AI safety methodology, and timely model evaluations, with limited novel technical contributions.

Transformative AI Economics

  • *The AI Industrial Explosion (Part 3)* models post-AGI growth by reoptimizing production recipes for cheap labor and capital-intensive factor prices.
  • *Will we really put data centers in space?* offers concrete cost/thermal analysis of orbital data centers, pushing back on speculative Musk-era claims.

AI Safety & Alignment

Model Reviews & Practice

Strategy & Timelines

Key Themes

AI Safety & Alignment · 5Transformative AI Economics · 2AI Strategy & Timelines · 2Model Reviews & Practice · 2

Primary evidence

Top Ranked Signals

Research LessWrong May 22

The AI Industrial Explosion — Part 3: Going faster

By djbinder

60 score
AI Analysis

Third installment in a series modeling post-AGI economic growth, examining how reoptimizing production recipes for post-AGI factor prices (cheap labor, fast capital reproduction) would accelerate economic doubling beyond what fixed-recipe models suggest. Uses 2017 US input-output tables as baseline.

In Part 1, I found that a fully automated economy using today's production methods could double roughly every year. In Part 2, I modeled the transition from today's economy to that maximum-growth composition and found that energy production could double within about four years. Both parts held production methods fixed: each sector continues using exactly the recipes it uses today, with robots replacing human workers. That assumption is too conservative. Today's production recipes were chosen at
AI EconomicsTransformative AIAGI Implications
Research LessWrong May 22

Will we really put data centers in space?

By Avi Parrack

55 score
AI Analysis

Analysis of the technical and economic feasibility of orbital data centers (ODCs) for AI compute, examining whether claims by Musk and others about space-based AI are realistic. Concludes that cost-competitiveness depends almost entirely on Starship reusability achieving Falcon-like economics (~$250/kg to orbit).

AbstractSeveral major technology companies have announced plans to operate AI data centers in orbit. Elon Musk recently claimed: “the lowest-cost place to put AI will be space […] within two years, maybe three.” If a meaningful fraction of new AI compute really is placed in space within a few years, that would be a fairly big deal for AI governance and strategy. Here we try to disentangle the hype from reality and provide a sober assessment of the technical and economic feasibility of orbital da
AI InfrastructureAI StrategyCompute Economics
Research LessWrong May 22

Gemini 3.5 Flash Looks Good For How Fast It Is

By Zvi

55 score
AI Analysis

Zvi's review of Google's recently-released Gemini 3.5 Flash, arguing it's the best at its speed point but not preferable to Opus 4.7 or GPT-5.5 for most uses. Covers other Google I/O announcements.

Google once again has a model worth at least some consideration. Gemini 3.5 Flash is likely the best model out there at its particular speed point, as long as you don’t mind that it is a Gemini model. So for cases where speed kills, this can be a reasonable choice. Otherwise, I don’t see signs you would want to use it over Opus 4.7 or GPT-5.5. Google also had some other offerings for I/O Day, which this post will also cover. Introducing Google Gemini 3.5 ‘Flash’ Google introduced Gemini 3.5 Flas
Language ModelsModel EvaluationGoogle AI
Research LessWrong May 22

Which technical AI safety fields are going to be automated first?

By Chamod Kalupahana

50 score
AI Analysis

Analysis of which technical AI safety subfields are most likely to be automated first by frontier labs, using feedback quality and economic incentive as the two key factors. Notes Anthropic's use of Mythos and UKAISI evaluations as early signals.

I’m transitioning into technical AI safety, and I find myself thinking a lot about what fields I want to research and where I’ll have the biggest impact. One thing I’ve found myself thinking about a lot recently is what fields are likely to be automated.This seems pretty likely since frontier labs will likely be automating capabilities research as a part of automated R and D, and safety research won’t be far off. Some initial examples of this are Anthropic is investigating automated alignment re
AI SafetyAutomated ResearchAI Alignment
Research LessWrong May 22

Counting Arguments in AI Safety

By Samuel Ratnam

50 score
AI Analysis

Examines the structure of 'counting arguments' in AI doom reasoning (vast goal space → most goals are unfriendly), drawing parallels to Bertrand's Paradox to question whether the chosen measure is principled. Engages with prior LessWrong critiques.

cf. www.lesswrong.com/posts/YsFZF3K9tuzbfrLxo/count... , www.lesswrong.com/posts/yQSmcfN4kA7rATHGK/many-... A counting argument is a style of argument that looks something like this:We are drawing from a space where there are many more Xs than YsTherefore, absent any strong reason to expect Ys, we are much more likely to get XsFor example, when trying to answer the question “what is the probability that super
AI SafetyProbability TheoryAI Alignment
Research LessWrong May 22

We made a map of the doom debate

By Sean Herrington

45 score
AI Analysis

AI Safety Camp project produced an interactive tree-structured map of AI threat pathways, allowing users to set probabilities for each branch and identify cruxes with others. Aims to systematize the P(Doom) debate by decomposing assumptions.

This was produced as a part of the AI Safety Camp 2026 "Assumptions of the Doom Debate" project, led by Sean Herrington, who was also the lead author on this post. The other participants have equal contributions and are listed in no particular order. It is the first in a sequence we intend to publish over the coming weeks. TL;DR:We have created a breakdown of AI threat pathways, which can be accessed at lifeuniversesafety.com/doom-assumptions/index.htmlThis breakdown is in a tree format,
AI SafetyExistential RiskAI Governance
Research LessWrong May 22

AI is Not Normal Technology

By Olivia Scharfman

45 score
AI Analysis

Response to Narayanan and Kapoor's 'AI as Normal Technology' essay, arguing AI is fundamentally different and citing biosecurity as the most concrete catastrophic risk. Uses chess history as a parallel.

Last year, Arvind Narayanan and Sayash Kapoor published a now well-circulated essay, AI as Normal Technology. The essay is still popular, I think, primarily because people would like it to be true, myself included. It is a terrifying proposition to acknowledge how different AI may be from “normal” technology. However, acknowledge it we must –– the urgency of developing proper governance and technical progress on alignment cannot be overstated.I think the essay has signif
AI GovernanceAI RiskAI Safety
Research LessWrong May 22

Notes on Collaborating with Claude Opus

By Nissa Seru

40 score
AI Analysis

Practical notes on prompting/collaboration patterns with Claude Opus 4.7, including how reasoning behind instructions improves compliance, the value of labeled response sections, and pitfalls of negatively-framed instructions. Practitioner observations.

INTENT: Share elements of my mental model regarding collaboration with Claude Opus models. Not intentionally scoped to a specific model version, but my experience is generally with the latest model version available (4.7 as of time of writing items 1-4 on 5/22/26)Accompanying an instruction with the why significantly improves: Observed rate of the instruction being visibly salient to Claude Quality/nuance of instruction execution (baselined on 1a)A standing instruction to break replies into labe
Prompt EngineeringHuman-AI CollaborationClaude
Research LessWrong May 22

Proposal for "Timelines to what": DIAL distribution

By tlevin

35 score
AI Analysis

Linkpost proposing the 'DIAL distribution' framework (decision importance adjusted for leverage) for reasoning about AI timelines, advocating distributions over point estimates and weighting by decision-relevance.

This is a linkpost for my post on the Coefficient Giving Substack today, which bundles a few points about AI timelines that I think are important to keep in mind when making decisions. TLDR, when making decisions with respect to AI timelines, usually you should:Use distributions, not point estimates/deadlines;Use "decision importance" rather than capability thresholds;Remember to account for leverage.I coin "DIAL distribution," where DIAL stands for "decision importance adjusted for leverage," a
AI TimelinesDecision TheoryAI Strategy
Research LessWrong May 22

Strong Longtermism Is Simply Correct

By Bentham's Bulldog

30 score
AI Analysis

Philosophical defense of strong longtermism, arguing the vast scale of potential future people means most expected impact of our actions concerns the long-run future. Standard utilitarian/longtermist argument with minor framing updates.

Crosspost. 1 Strong Longtermism explained Strong Longtermism is the idea that the most important features of our actions concern how they affect the long-run future. The case for it is very simple. The future could contain ridiculously large numbers of people—10^58 by some estimates, far more by others. While humanity might die off soon, we might survive for billions of years. What we do today has some chance of affecting things over cosmic timescales—vaster than empires and
LongtermismEthicsAI Philosophy
Research LessWrong May 22

Insurance Premiums To The Moon

By PossiblyElaine

15 score
AI Analysis

Discussion of rising US insurance premiums, breaking down where premium dollars go and considering causes. Not AI-related.

(This post does not belabor why most people want or need insurance. That has been extensively discussed: What makes buying insurance rational?, When Is Insurance Worth It?, Money threshold Trigger Action Patterns.)We the people have not been loving our insurance premiums, which has outpaced both inflation and wage increase in the US. Many insurance companies cite policy changes or natural disasters as the reasoning, while a growing population thinks corporate greed is the dominant factor (possib
EconomicsInsurance