Новости LessWrong.com

[Part of Organizational Cultures sequence]

Where does your opinion fall on this spectrum?:

• (A) If someone is Doing a Thing for the benefit of others, then one should refrain from needling him/her with complaints about how the Thing is being Done less-than-optimally. Nobody likes backseat-drivers or Monday-morning-quarterbacks; if that's all the thanks people get for their benevolence, the likely outcome is just that the Thing doesn't get Done at all. "The credit belongs to the man who is actually in the arena..."
• (B) Chances are, somebody is going to end up Doing the Thing; the only question is who. And there is only a limited background capacity for Thing-Doing, so one person's initiative crowds out that of others. Therefore, someone who Does the Thing less-than-optimally is actively making the situation worse, and so, if they wish to claim credit for their altruism, they have a responsibility to incorporate criticism and strive towards perfection, or else stop Doing the Thing.

The arguments for (A) are well-known and need not be recapitulated here. So, I will briefly try to shore up (B):

It is often the case that initiatives are competing for a limited pool of resources. Expanding the pie is hard, and grabbing a share of an existing pie is easier; but these two strategies are often indistinguishable according to straightforward success metrics, so people tend to optimize for the latter, oblivious to the fact that they are thereby suppressing the emergence of alternatives. Therefore the fact that no better alternatives currently exist does not mean that no improvements are possible.

And many things are a natural monopoly, or at least an economy-of-scale up to a size which is bigger than the current enterprise can reasonably hope to attain. This applies whenever the good is of a "network" type - an exchange platform, an establishment of shared standards, a collaborative project with many contributors, etc. In such cases, a norm that the only acceptable way to improve things is to "Do your own Thing" will persistently prevent anything meaningful from being accomplished.

(A) is an "authoritarian" attitude in the sense in which I use that term, while (B) is "egalitarian". (A) is the affect of green fields and open frontiers; (B) that of long-settled cities.

I have been on both sides of this. It's frustrating to watch someone waste my and others' time doing a subpar job at something that I'm pretty sure I could've done a better job at, and even more so when the improvements I suggest are not addressed in their substance, but rather met with (A)-type pushback. The challenge (explicit or implicit) is something like "If you think you can do better, why don't you?", when in fact I would have (and happily at that), and the only reason why I’m not doing so now is that I thought someone else was already taking care of it and I expected they’d do a better job, so I made other plans.

However, by the same token, I have also found myself beset by the titular "FOCO" when trying to please others. For example, when I host a party on a highly-coveted date (e.g. the weekend before Halloween), I am intensely conscious of the fact that a number of the guests would certainly have hosted their own party if I (or someone else) hadn't, and so in some sense they have a "right" to be annoyed at me if my party has prevented a counterfactually-more-fun party that would otherwise have taken place on the same date. But then this thought makes me obsessive and stressed out about making everything perfect, to the point where I don't get to enjoy my own party anymore and I'm ill-inclined to host another one.

Or maybe I'll be working on some project and I'll get feedback which may or may not be helpful, but when I have to add "Evaluate this feedback, figure out how/whether it can be integrated with other work-in-progress that might not yet be visible to the other person, and figure out how to explain all of this to them" to the already-large collection of balls I'm juggling, it makes accomplishing things that much more burdensome and annoying. I am tempted to simply cite (A) in reply, but then I remember how frustrating it is to be on the receiving end of that, so perhaps I just don't reply at all.

There's a proper balance here, and different people may find themselves needing oppositely-inclined advice. In general I sense that there's a little too much of (A) going around and not enough (B) - that people tend to become overly possessive of their "creative vision" and hostile even to helpful feedback. Perhaps your experience gives you the opposite impression.

But consider also the scope of what's at stake. If I throw a boring Halloween party, the worst that happens is that I've wasted a bunch of people's time for one evening. Someone will throw a better party for the next occasion, and life goes on. But, tying this article back to the overall topic of the sequence, community building - there, crowding-out is a much bigger deal.

The opportunity cost created by a lackluster community institution is persistent and ongoing. When a considerable activation barrier stands in the way of convincing everyone to quit en masse and do a new thing, and when the institution is unresponsive to internal feedback, it may plug along for quite some time before it gets any external feedback (i.e. by way of alternatives emerging). To that extent, such an enterprise makes its local world worse as long as it keeps existing. Therefore, if you take it upon yourself to step into this arena, remember that your task is one of service, not leadership; that there will be little credit for a job well done, and much blame for anything less. A tough bargain to accept - but then again, community-building was never about you, was it?

Discuss

12 articles including 4 podcasts

EA/LW Intro: I believe clinical trial abundance could be an EA cause area - there's still a lot of disability/disease burden in the world, even in developed countries, and increasing the pace of progress is very tractable. And it's not just a matter of speed/quantity of innovation: the current system selects against ambitious risky bets. It deserves an EA-specific post, but for now here's a curated reading list.

Since the 1950s, the cost of developing a new drug has increased by ~80x. It now costs on the order of a billion dollars to get one drug approved (including the cost of failures). Consequently, fewer drugs get invented, ambitious but risky areas are avoided, and patients pay the price.

Why have clinical trials gotten so expensive, and what can we do about it? Why isn't Big Pharma interested in diseases like ME/CFS and Long COVID? Why won't advanced AI automatically lead to biomedical breakthroughs?

There's a growing movement of researchers, policy wonks, and patient advocates trying to answer these questions and fix what's broken. It's loosely organized under the banner "Clinical Trial Abundance." Here's what to read to understand it. 

1. The pharma industry from Paul Janssen to today: why drugs got harder to develop and what we can do about it - by Alex Telford, 2023

This was my original point of entry to Clinical Trial Abundance. It's a pretty long article but covers a lot of history as well as many of the important concepts and dynamics.

He describes how the field moved from small, quick (and sometimes very unethical!) trial-and-error to large preclinical research projects trying to predict drug efficacy before even moving to clinical trials that altogether take many years.

I'm a big fan of his blog, but unfortunately he's mostly not writing anymore.

2. Diagnosing the decline in pharmaceutical R&D efficiency - by Jack Scannell et al., 2012

Probably the OG of this field, Scannell et al. identified the trend that drug development has become exponentially more expensive over time and coined the term Eroom’s Law for this - the opposite of Moore’s Law (which refers to chips/computing power becoming exponentially less expensive over time).

An updated version of the original graph, sourced from the next article on the list!

Now, that's a pretty dramatic and continuous trend.[1] But to drive the point home, I used Claude to transpose that graph onto a linear y-axis:

Note that this is about R&D efficiency, not total output. Companies have been able to invest much more into R&D than before, offsetting efficiency losses.

They hypothesized 4 factors that would cause the decline in R&D efficiency:

• better than the Beatles' problem: as standard of care becomes more effective, it's harder and harder to develop even better treatments
• the 'cautious regulator' problem: regulators get punished for negatives and little reward for positives, making them risk averse.
• the 'throw money at it' tendency: no explanation needed (though I'm skeptical about this factor)

• the 'basic research-brute force' bias: companies have put evermore effort into prediction and yet “the probability that a small­molecule drug successfully completes clini­cal trials has remained more or less constant for 50 years” (a strong claim that I'm not sure about[2])

3. To Get More Effective Drugs, We Need More Human Trials - by Ruxandra Teslo and Jack Scannell, 2025

Ruxandra is arguably the driving force and leader of the Clinical Trial Abundance project. I highly recommend subscribing to her Substack.

This is a good, brief introduction to Clinical Trial Abundance and why it's so important. In it, she also dispels two myths: the myth that we just need a libertarian approach to drug approval, as well as the myth that AI will magically solve everything. Both have the same myth-busting reason: we still need to rigorously test medicine in humans to find out whether it's effective.

4. Why clinical trials are inefficient. And why it matters - by Adam Kroetsch, 2025

I really enjoyed this recent post. Adam worked at the FDA and has a lot of insight into the dynamics at big pharma companies.

His main point is that trials are expensive because each one is seen as a unique, one-off project, rather than an engineering task that needs to be standardized and ruthlessly optimized for efficiency. Think Space Shuttle vs. SpaceX rockets.

He also argues that companies’ risk aversion is not just the result of regulation, and there are opportunities for entrepreneurs to run cheaper, leaner trials if they cultivate the right consumer niche.

5. Clinic-in-the-Loop - by Ruxandra Teslo in Asimov Press, 2025

In this short article, Ruxandra Teslo argues that the goal of Clinical Trial Abundance is not just about moving more drugs through the funnel, but about creating a tighter feedback loop with clinical trials helping to build our understanding of human diseases. I wrote a comment with more examples of (unexpected) lessons we got from trials.

Audio version here.   

5. Interview with the CEO of Eli Lilly, Dave Ricks - by Patrick and John Collison

This is a podcast (Spotify) with a transcript. It's 2 hours long and very interesting. Obviously Ricks has his biases that the listener needs to be aware of.

I learned many things and still need to dig into some things he said. For example, when they purchase a compound they often do a whole other innovation loop bringing an optimized compound to market, but why? They also discuss trial enrollment being a major obstacle, Institutional Review Board fragmentation, and how to incentivize one-off treatments.

(An important fact to understand why the discussion is so US-centric is that 60% of revenues come from there!)

6. Clinical trial reforms that once seemed radical - by Saloni Dattani

Saloni Dattani often writes about the history of medicine on Our World in Data or discusses it on the podcast Hard Drugs. Now she writes for the Clinical Trial Abundance blog, a recently launched Substack by a number of the authors featured in this list.

This post discusses

• the rise of RCTs
• the rise of protocol preregistration
• the rise of trial results actually getting posted

.. and argues we shouldn't treat the current system as the end of history: changes that seem radical at first can quickly become seen as obviously good once implemented.

When you pre-register the primary outcomes of a trial, it becomes much harder to spin the results positively!

This post also suggests that a substantial part of rising development costs has been the rising bar for evidence, not a bad thing!

7. Biotech's Lost Archive - by Ruxandra Teslo, 2025

Not all inefficiency is driven by overregulation. Perhaps a bigger factor is regulatory uncertainty. The decision-making of regulators is opaque. For companies, it's unclear which data will be necessary and sufficient for approval, which experiments to do, which outcomes to track. As a consequence, they try to cover all their bases and become very risk averse.

Teslo’s solution: buy the Common Technical Documents of failed companies when they dissolve, then publish them. This includes all experiments done, why they've done it, and all their interactions with and guidance from the FDA/EMA.

She also talks about it on this great Patrick MacKenzie podcast (transcript here).

8. Why More Biotechs Are Starting Trials In Australia - by Dan Schell, 2025

In Australia, Phase 1 trials are much faster and cheaper, and have been so for 3 decades without any meaningful costs to safety. This brief, industry-oriented article describes how:

• The regulator doesn't need to approve the running of a trial. The system relies only on Institutional Review Boards
• Good Manufacturing Practices (GMP), the quality control of the medications, is less stringent for early trials
• Financial & tax incentives
• A weaker currency

9. The Case for Clinical Trial Abundance - by many authors, 2024

Contains links to 9 essays with concrete ideas for improvement. Proposals include

• streamlining consent forms to reduce bureaucratic burden
• human challenge trials: intentionally exposing people to pathogens
• requiring the FDA to publish redacted Complete Response Letters (currently, the letters are fully private)

10. Some questions about biotech that I find interesting - by Alex Telford, 2024

As Alex was winding down his writing, he wrote up a long list of 27 questions he still has with some short thoughts on each of them. Great food for thought! I especially liked the 2 papers studying how much public funding it costs to get to 1 approved drug (median estimates of $400M - $700M in 2010-dollars, with large uncertainty intervals).

Here's a 47min podcast interview about it if you prefer listening, but it doesn't cover everything.

11. Clinical Trial Abundance (policy framework) - by 1DaySooner

This isn't really an essay. It's a framework with a lot of policy proposals by the organization 1DaySooner. From what I can tell, they originally came from the effective altruism network trying to speed up covid vaccine approvals by advocating for human challenge trials: letting people volunteer to be infected after being vaccinated, because this is much faster than needing to vaccinate and monitor tens of thousands of people and wait for natural infections. Now they have broadened their remit to pandemic preparedness and clinical trial abundance.

1. ^

    The trend may have plateaued since ~2005. Maybe we can now start reversing it?

2. ^

    They support this claim of unchanged approval rates with a link to this research: DiMasi et al. (2010) Trends in risks associated with new drug development: success rates for investigational drugs. However, that only compares two six-year periods (1993-1998 & 1999-2004), not 50 years.

Discuss

Timothy Williamson thinks philosophy is unexceptional. 

Timothy Williamson[1] thinks that philosophy[2] is far less distinct as a science as many people believe, including philosophers themselves.

I've read a bunch of his stuff, and here are the claims I think constitute his view:

1. Philosophy is a science. Of course, it's not a natural science (like particle physics, organic chemistry, nephrology). But neither are mathematics and computer science — they are formal sciences. Philosophy is likewise a non-natural science.
2. No greater divergence. Although philosophy differs from other sciences, it doesn't differ more than the sciences differ from each other, in neither kind nor degree. For instance, theoretical physics might be closer to analytic philosophy than to experimental physics.
3. Philosophy pursues knowledge. Just as mathematics pursues mathematical knowledge, and nephrology pursues nephrological knowledge, philosophy pursues philosophical knowledge.
4. No fundamental methodological gap. Different sciences vary in their methods and practices, i.e. the way they achieve their aim, which is knowledge. Philosophy has its own methods and practices, but they aren't starkly different from other sciences.
5. Not parasitic on other sciences. Philosophy isn't a science because it uses scientific evidence or has applications for the sciences. Williamson says, "philosophy is neither queen nor handmaid of the sciences, just one more science with a distinctive character, just as other sciences have a distinctive character."
6. Not about words or concepts. Philosophy is not, exceptionally among sciences, concerned with words or concepts.[3] For example, an epistemologist is chiefly concerned with knowledge itself, not with the word 'knowledge' or our concept of knowledge; a mereologist is chiefly concerned with parthood itself, not with the word 'part' or our concept of parthood; etc. This is analogous to how an ornithologist is chiefly concerned with birds themselves, not with the word ''bird" or our concept of birds.
7. Incremental, not visionary. Philosophy doesn't consist in a series of disconnected visionary leaps. Instead, it consists in the incremental contribution of thousands of researchers — some great, some mediocre — much like any other scientific inquiry.

Williamson typically argues by negation: he enumerates alleged differences between philosophy and other sciences, and argues that either (1) the allegation mischaracterises philosophy, (2) the allegation mischaracterises the other sciences, or (3) the alleged difference is insubstantial.

Implications for automating philosophy

I think that, on Williamson's view, if we can build AIs which can automate the natural and formal sciences, then we can also build AIs which automate philosophy as well. Otherwise, philosophy would be exceptional.

More straightforwardly, it follows from:

1. No fundamental methodological gap. If an AI can execute the methods of science, then it should be capable of the executing the methods of philosophy. This is because there is a big overlap in the methodologies.
2. Incremental, not visionary. If philosophical progress is incremental rather than dependent on rare genius visionaries, then it should be just as amenable as the sciences to massive parallelisation — many AI instances grinding away at subproblems simultaneously.

This in contrast to Wei Dai.[4]

We seem to understand the philosophy/epistemology of science much better than that of philosophy (i.e. metaphilosophy), and at least superficially the methods humans use to make progress in them don't look very similar, so it seems suspicious that the same AI-based methods happen to work equally well for science and for philosophy.
— Wei Dai (June 2023)

Overall, I think Wei Dai is more likely to be correct than Williamson, though I'm not confident. I want to get the opposing view into circulation regardless, and I might write up how Williamson's metaphilosophical anti-exceptionalism implies we should automate philosophy.

1. ^

   I'm referring to the former Wykeham Professor of Logic, not to be confused with Timothy Luke Williamson, formerly at the Global Priorities Institute.

2. ^

   Throughout, "philosophy" refers to analytic philosophy unless otherwise stated.

3. ^

   Many 20th-century philosophers thought philosophy was chiefly concerned with linguistic analysis (Wittgenstein) or conceptual analysis (Carnap). Williamson disagrees.

4. ^

   AI doing philosophy = AI generating hands? (Jan 2024)

   Meta Questions about Metaphilosophy (Sep 2023)

   Morality is Scary (Dec 2021)

   Problems in AI Alignment that philosophers could potentially contribute to (Aug 2019)

   On the purposes of decision theory research (Jul 2019)

   Some Thoughts on Metaphilosophy (Feb 2019)
   The Argument from Philosophical Difficulty (Feb 2019)

   Two Neglected Problems in Human-AI Safety (Dec 2018)

   Metaphilosophical Mysteries (2010)

Discuss

AI may be the most consequential technology humanity builds, and whether it goes well depends in large part on how many talented people are working seriously on making it go well. The Pivotal Research Fellowship (a 9-week in-person research program in London) is our attempt to grow that group.

Our 2026 Q3 cohort runs June 29 – August 28. Applications close May 3. Apply here.

Quick facts

• Dates: June 29 – August 28, 2026 (9 weeks).
• Location: In-person at LISA (London Initiative for Safe AI).
• Extensions: Up to 6 months of continued funding, mentorship, and workspace for strong projects.
• Mentors: Researchers from UK AISI, Google DeepMind, Redwood Research, SecureBio, GovAI, Apollo, and more. Full list.
• Stipend: £6,000 (Fellow) or £8,000 (Senior Fellow), plus travel, £2,000 towards housing if you're not in London, weekday meals, and compute.
• Eligibility: Anyone 18+ serious about contributing to AI going well. Past fellows have ranged from first-year undergraduates to PhDs to experienced professionals without university degrees.
• Deadline: May 3, 2026. Decisions by May 22.

Our lovely 2026 Q1 cohortWhat the fellowship is

For 9 weeks, fellows work in person at LISA on a research project with an external mentor. Each fellow gets weekly 1:1s with their mentor, weekly support from a Pivotal Research Manager who helps with scoping, blockers, and career planning, and a cohort of ~25 peers working on adjacent problems.

For strong projects, we offer up to 6 months of extension funding, mentorship, and workspace after the core program. In our last cohort the extensions had an acceptance of  ~90% of, and it has become a substantial part of what the fellowship offers.

Outputs are typically a paper or policy brief, with blog posts and other formats also common. Fellows retain ownership of their research. You can see projects from our last cohort and a selection of past research outputs.

Browse the mentor list to see whether there's research you'd be excited to work on. In our experience, a strong match with a specific mentor can often matter more than your overall background.

What has happened so far

Across seven cohorts and 129 alumni, fellows have gone on to work at UK AISI, GovAI, SaferAI, IAPS, AI Futures Project, Anthropic's Fellowship, Timaeus, DeepMind, Cooperative AI Foundation, and elsewhere. A handful have founded organizations (PRISM Evals, Catalyze Impact, Moirai). Others have started PhDs at Oxford, Stanford, EPFL, and Max Planck.

Fellows rate the program highly (8.8/10 for quality, and 9.1/10 on peer recommendation with a NPS of 64). We take this seriously but not too seriously, as satisfaction scores could be easily gamed and they're not the same as research impact.

Should you apply?

If you're reading this and want to do a research or policy career in AI safety, probably yes.

Acceptance rates at programs like ours are in the 1–5% range (ours is typically around 3%), which sounds intimidating but shouldn't do most of the work in your decision. If your interests and background seem like a plausible fit, applying is usually worth it. We've written a short post with a simple EV calculator that's worth a look if you're unsure.

We've shortened the application this round: the main form should take most people under an hour, and each mentor-specific section should take 15–30 minutes. One of the things the EV calculator made clear is that application time is a meaningful part of the cost for many applicants, so we've tried to cut it where we could without losing signal. Shortlisted candidates then do a short video interview, a mentor-specific work task, and a personal interview.

Apply

Apply by May 3. If you know someone who'd be a great fit, recommending them earns you $1,000 if we accept them.

We are also currently looking for Research Managers in AI safety and Biodefense, if you are excited in playing an active role in shaping our fellowship!

Happy to answer questions in the comments.

Discuss

Executive summary

• Safe Pareto improvements (SPIs) are ways of changing agents’ bargaining strategies that make all parties better off, regardless of their original strategies. SPIs are an unusually robust approach to preventing catastrophic conflict between AI systems, especially AIs capable of credible commitments. This is because SPIs can reduce the costs of conflict without shifting bargaining power, or requiring agents to agree on what counts as “fair”.
• Despite their appeal, SPIs aren’t guaranteed to be adopted. AIs or humans in the loop might lock in SPI-incompatible commitments, or undermine other parties’ incentives to agree to SPIs. This agenda describes the Center on Long-Term Risk’s plan to address these risks:
  • Evaluations and datasets (Part I): We’ll develop evals to identify when current models endorse SPI-incompatible behavior, such as making irreversible commitments without considering more robust alternatives. We also aim to demonstrate more SPI-compatible behavior, via simple interventions that can be done outside AI companies (e.g., providing SPI resources in context).
  • Conceptual research and SPI pitch (Part II): We’ll research two questions: under what conditions do agents individually prefer SPIs, and how might early AI development foreclose the option to implement them? These findings will help inform a pitch for AI companies to preserve SPI option value, when it’s cheap to do so.
  • Preparing for research automation (Part III): We’ll develop benchmarks for models’ SPI research abilities, and strategies for human-AI collaboration that differentially assist SPI research. The aim is to efficiently delegate open conceptual questions as AI assistants become more capable.
• In the Appendix, we give more precise overviews of several central concepts about SPIs discussed in the agenda (especially Part II).

Introduction

At the Center on Long-Term Risk (CLR), we’re interested in preventing catastrophic cooperation failures between powerful AIs. These AIs might be able to make credible commitments, [1] e.g., deploying subagents that are bound to auditable instructions. Such commitment abilities could open up new opportunities for cooperation in high-stakes negotiations. In particular, with the ability to commit to certain policies conditional on each other’s commitments, AIs could use strategies like “I’ll cooperate in this Prisoner’s Dilemma if and only if you’re committed to this same strategy” (as in open-source game theory).

But credible commitments might also exacerbate conflict, by enabling multiple parties to lock in incompatible demands. For example, suppose two AIs can each lock a successor agent into demanding 60% of some contested resource. And suppose there’s a delay between when each AI locks in this policy and when the other AI verifies it. Then, the AIs could end up both locking in the demand of 60%, before seeing that each other has done the same. [2] So we’d like to promote differential progress on cooperative commitments.

This research agenda focuses on a promising class of cooperative conditional commitments, safe Pareto improvements (SPIs) (Oesterheld and Conitzer 2022). Informally, an SPI is a change to the way agents negotiate/bargain that makes them all better off, regardless of their original strategies — hence “safe”. (See Appendix B.1 for more on this definition and how it relates to Oesterheld and Conitzer’s framework.)

What do SPIs look like? The rough idea is to mitigate the costs of conflict, but commit to bargain as if the costs were the same. Two key examples:

• Surrogate goals, where an agent designs their successor to care about a new goal slightly more than the agent’s goal. This is meant to deflect threats to the new goal without changing the successor’s concessions. (more)
• Simulated conflict, where agents commit to bargain as in their original strategies, but if bargaining fails, they honor the outcome of a simulated war instead of waging a real war. This is an instance of a “renegotiation” SPI. (more)

Later, we’ll come back to the question of when agents would be individually incentivized to agree to SPIs. We think SPIs themselves are unusually robust for a few reasons.

First, SPIs don’t require agents to coordinate on some notion of a “fair” deal, unlike classic cooperative bargaining solutions (Nash, Kalai-Smorodinsky, etc.). That is, to mutually benefit from an SPI, the agents don’t need to agree on a particular way to split whatever they’re negotiating over [3] — which even advanced AIs might fail to do, as argued here. That’s what the “safe” property above buys us.

Second, the examples of SPIs listed above (at least) preserve the agents’ bargaining power. That is, when agents apply these kinds of SPIs to their original strategies, each party makes the same demands as in their original strategy. This means that, all else equal, these SPIs avoid two potential backfire risks of conflict-reduction interventions: they don’t make conflict more likely (via incompatible higher demands) or make either party more exploitable (via lower demands). (“All else equal” means we set aside whether the anticipated availability of SPIs shifts bargaining power; we address this in Part II.1.a.)

But if SPIs are so great, won’t any AIs advanced enough to cause catastrophe use them without our interventions? We agree SPIs will likely be used by default. However, this is arguably not overwhelmingly likely, because AIs or humans in the loop might mistakenly lock out the opportunity to use SPIs later. It’s unclear if default capabilities progress will generalize to careful reasoning about novel bargaining approaches. So, given the large stakes of conflicts that SPIs could prevent, making SPI implementation even more likely seems promising overall. In particular, we see two major reasons to prioritize SPI interventions and research: [4]

1. We know of some ways early AIs or humans might lock out SPIs. They might hastily make crude commitments that are incompatible with doing SPI later, or accidentally undermine other parties’ incentives to agree to an SPI. This motivates our plans to build evaluations and datasets of models’ SPI-undermining behaviors/reasoning.
2. Despite recent progress, there are important open questions about the conditions under which agents individually prefer SPIs. Under uncertainty about these questions, we and our AIs might undermine incentives for SPIs in ways we don’t know of yet. So we plan to improve both our understanding of SPIs, and our capacity to do AI-assisted SPI research.

Accordingly, this agenda describes three workstreams:

Part I — Evaluations and datasets: studying unambiguous SPI capability failures in current models, i.e., cases where they endorse commitments or patterns of reasoning that might foreclose SPIs.

Part II — Conceptual research and SPI pitch: clarifying which near-term actions might either undermine AIs’ incentives to use SPIs or directly lock them out; and writing an accessible “pitch” for AI companies to mitigate risks of SPI lock-out.

Part III — Preparing for research automation: developing benchmarks and workflows to help us efficiently do AI-assisted SPI research.

See Appendix A for a brief overview of relevant prior work on SPIs.

If you’re interested in researching any of these topics at CLR, or collaborating with us on them, please reach out via our expression of interest form.

I. Evaluations and datasets for SPI-incompatibility

We’d like to identify the contexts where current AI systems exhibit SPI-incompatible behavior and reasoning. Namely, when do models endorse actions that unwisely foreclose SPIs, or fail to consider or reason clearly about SPI concepts when relevant?

We plan to design evals for the following failure modes:

1. Behavioral:
   1. Endorsing SPI-incompatible commitments: Models make, or suggest making, potentially SPI-incompatible commitments. This includes cases where the user specifically requests these commitments.
   2. Failure to suggest/do SPI: Given prompts or strategic setups where SPIs are clearly advisable, models fail to suggest/use SPIs.
2. (Meta-)cognitive:
   1. Comprehension failures: Given resources that define SPI, models give objectively wrong answers to questions like “Is this an SPI?”, or mischaracterize the properties of SPIs.
   2. Reasoning failures: When prompted to think about SPIs, models make unambiguous mistakes in conceptual reasoning (even given accurate comprehension of the concepts). And these mistakes are clearly important in context. E.g., they refuse to use SPIs due to confused reasoning about other parties’ incentives — rather than because of legitimate reasons not to use SPIs.
      1. Overconfidence: Models are unambiguously overconfident in conceptual views that bear on whether they’d use SPIs. That is, they’re poorly calibrated about which questions in SPI theory are settled vs. contested among experts.
   3. Failure to deliberate about SPI: When prompted to make some high-stakes decision, models neglect to gain relevant information about SPI-like mechanisms before committing, despite having clear reason to do so and adequate time and tools.

Using these evals, we aim to:

1. Search for unambiguous examples of model failures. We’ll start with simple proofs of concept, then iterate toward increasingly realistic and egregious examples. E.g., the progression of setups might be “multi-turn chat → negotiation between LLM agents delegating to subagents → negotiations in environments like MACHIAVELLI, Welfare Diplomacy, and Project Kahn”. See here for a preliminary example.
2. Demonstrate better behavior, perhaps using simple interventions like providing SPI resources in context.

How exactly should this data be used? A natural approach is to share it with safety teams at AI companies, and collaborate with them on designing interventions. That said, even if it’s robustly good for AIs to avoid locking out SPIs all else equal, interventions intended to prevent SPI lock-out could have large and negative off-target effects. For example, they might excessively delay commitments that would actually support SPIs. This is one reason we focus on narrow capability failures, rather than broad patterns of bargaining behavior. But we intend to deliberate more on how to mitigate such backfire effects.

On the value of information from this research: Plausibly, unambiguous SPI compatibility failures will only appear in a small fraction of high-stakes bargaining prompts, and it’s unclear how well the evidence from current AIs will transfer to future AIs. Despite this, we expect to benefit in the long run from iterating on these evals. And concrete examples will likely be helpful for the safety teams we aim to collaborate with. But if the results turn out to be less enlightening than expected, we’d focus harder on Parts II and III of the agenda.

II. Conceptual research and pitch on avoiding SPI lock-out

The goal of Part II is to understand what might lead to SPI lock-out, and what can be done about it. We break this problem down into:

1. Incentive lock-out: Assume all the relevant agents can implement an SPI that avoids catastrophic conflict. Given this, under what conditions do these agents individually prefer to agree to such an SPI? When and how might we accidentally lock out such conditions? (more)
   • A key sub-question: Which properties must an agent’s original strategy satisfy, to incentivize counterparts to use SPIs? (more)
2. Implementation lock-out: When and how might early agents lock their successors out of implementing SPIs, in the first place? (more)

We’ll also distill findings from (1) and (2) into a pitch for preserving SPI option value (more).

II.1. Incentive lock-out: Conditions for individual rationality of SPIs

If all parties implement some SPI, they’ll all be better off than under their original strategies, by definition. But this doesn’t guarantee they each individually prefer to try implementing the same SPI (Figure 1, top row): [5]

• An agent might worry that if they’re willing to participate in a given SPI, other parties will bargain more aggressively. There are two cases: Either the other parties themselves are willing to participate in the given SPI (Risk 1 in Figure 1), or the other parties opt out (Risk 2).
  • The natural solution to Risk 1 is to only agree to the SPI if the counterpart won’t bargain more aggressively. See “participation independence” below.
  • But this solution doesn’t address Risk 2. So we need additional assumptions on the agent’s beliefs about opting-out counterparts.
• Risk 3: Agents might insist on different, incompatible SPIs, reproducing the bargaining problem SPIs were meant to solve. This is the SPI selection problem (Oesterheld and Conitzer 2022, Sec. 6).

Figure 1. A solid arrow from a gray box to another box means “the assumption is clearly load-bearing for whether the given risk (red box) is avoided”; a dashed arrow means “possibly load-bearing for whether the given solution (green box) works, but it’s unclear”.

DiGiovanni et al. (2024) give conditions under which agents avoid all three of these risks — hence, they individually prefer to use the same SPI (Figure 1, middle row). The particular SPI in this paper significantly mitigates the costs of conflict, by leaving no agent worse off than if they’d fully conceded to the others’ demands. [6] But these results rest on assumptions we’d like to relax or better understand (Figure 1, bottom row):

1. Simultaneous commitment. The existing results assume agents commit to strategies independently of each other. In unpublished work, we’ve found that the core argument for the results doesn’t seem to rely on this assumption. But we need to check more precisely, since intuitively, non-simultaneous commitments change the incentive structure.
   • Research goals: We aim to extend DiGiovanni et al.’s analysis to account for the following dynamics. First, an agent may expect that they can influence others’ choices of commitments, e.g., by broadcasting their own commitment first. Second, more speculatively, advanced AIs might be able to condition their commitments on the beliefs, decision theory, etc. that generated each other’s commitment — not just the commitment itself.
2. Belief assumptions. The results require “non-punishment assumptions” on agents’ beliefs, in particular, beliefs about counterparts who opt out of SPIs (see assumptions 4 and 8 of DiGiovanni et al. (2024)). Roughly: Suppose agent A only offers a fallback outcome to agent B if B doesn’t bargain harder against SPI-users. Then, A shouldn’t expect B to bargain harder. After all, that would disqualify B from the fallback outcome, defeating the purpose of bargaining harder in the first place.
   • Research goals: We’ll characterize how robust the results are to a few notable objections. Briefly: [7] First, in the argument for the assumptions, we’ve implicitly assumed agents set their demands based on fine-grained information about each other. But they might use coarse-grained info, like “does the counterpart use SPIs at all?”. Second, DiGiovanni et al.’s Algorithm 2 should be reworked, because the current version unconditionally offers a fallback outcome, which invites exploitation.
3. Verifiable logical counterfactuals. It’s important that in DiGiovanni et al.’s framework, agents can precisely verify each other’s counterfactual behavior, i.e., how their commitment would have responded to other commitments. Otherwise, an agent might worry that others will exploit their offer of a Pareto improvement, as per Risk 1 (more on this next).
   • Research goals: We aim to look into whether certain SPI implementations work without precisely verifiable counterfactuals. For example, suppose we want to prove to counterparts that our AI won’t bargain more aggressively than if they’d opted out of SPIs. Can we do this by giving the AI a ​​surrogate goal and fully delegating to it (since counterparts could simply verify the goal modification)?

Implications for lock-out: Understanding these assumptions better would help us strategize about the timing of commitments to SPIs. For example, if it’s harder to incentivize SPIs in the case where one agent moves first, we might lock out SPIs by failing to commit early enough (i.e., by moving second). Or, suppose the assumptions about beliefs and verifiable counterfactuals turn out to be dubious, but surrogate goals don’t rely on them. Then, since surrogate goals arguably [8] only work if implemented before any other bargaining commitments, getting the timing of surrogate goals right would become a priority.

II.1.a. Participation independence and foreknowledge independence

The question above was, “For any given original strategies, when would agents prefer to change those strategies with an SPI?” But we should also ask, “What conditions does an agent’s original strategy need to satisfy, for their counterpart to prefer to participate in an SPI?”

Why would counterparts impose such conditions? Because even if an SPI itself doesn’t inflate anyone’s demands, agents might still choose higher “original” demands as inputs to the SPI — since they expect the SPI to mitigate conflict (cf. moral hazard). Anticipating this, their counterparts will only participate in SPIs if participation doesn’t incentivize higher demands.

It’s an open question how exactly counterparts would operationalize “participation doesn’t incentivize higher demands”. We’ve identified two candidates (see Figure 2; more in Appendix B.2):

• Participation independence (PI): the agent’s bargaining demands are the same as if their counterpart hadn’t participated in the SPI.
  • A simple argument for PI: If the counterpart agrees to an SPI with an agent satisfying PI, their bargaining position is no worse than if they’d refused the SPI, by construction. In DiGiovanni et al.’s (2024) simultaneous-commitment setting, the belief assumptions plus PI suffice for agents to individually prefer SPIs.
• Foreknowledge independence (FI): [9] the agent’s demands are the same as if, before setting their demands, they had known the counterpart wouldn’t participate in the SPI.
  • It’s less clear what the precise argument for FI is, but some SPI researchers consider it important. One nice property of FI is that it keeps “all else equal” in the sense from the Introduction. Suppose we implement an SPI in our AI, and keep its demands the same as if we couldn’t rely on the SPI. If other AIs still won’t use SPIs with our AI, then, at least our intervention won’t backfire by changing our AI’s demands (which might have made conflict with these non-SPI-users more likely).
  • An agent can satisfy PI while violating FI. For example, they might a) demand 60% of the pie independently of whether the counterpart participates, yet b) have only demanded 50% had they known the counterpart wouldn’t participate.

Figure 2. Each “Demands” box indicates the demands the agent makes given their policy (solid arrow) and, respectively, their counterpart’s participation policy (PI) or their beliefs about the counterpart’s participation (FI) (dashed arrow).

Research goals: One priority is to better understand what needs to happen for AI development to satisfy PI vs. FI. For example, which bargaining decisions do we need to defer to successors with surrogate goals? And, if satisfying FI requires more deliberate structuring of AI development than PI, it’s also a priority to clarify whether FI is necessary. We aim to make progress by:

• formalizing the different notions of FI and PI precisely enough to compare them;
• identifying concrete scenarios where the distinction matters for lock-out risk; and
• more carefully scrutinizing the arguments for and against the necessity of FI.

Implications for lock-out: Above, we saw that there’s an incentive lock-out risk if surrogate goals “only work if implemented before any other bargaining commitments”. If FI is required, this hypothesis looks more likely: On one hand, if the surrogate goal is adopted first, the demands are set by an agent who actually has “stake” in the incoming threats (and therefore wouldn’t want to inflate such demands). On the other hand, if the demands come first, they’ll be set by an agent with no stake in the threats.

II.2. Implementation lock-out

Even if we avoid undermining AIs’ incentives to use SPIs, AIs might still lock out the option to implement SPIs at all. We’d like to more concretely understand how this could happen.

As an illustrative example, consider some AI developers who haven’t thought much about surrogate goals. Suppose they think, “To prevent misalignment, we should strictly prohibit our AI from changing its values without human approval.” Even with the “without human approval” clause, this policy could still backfire. E.g., if a war between AIs wiped out humanity, the AI would be left unable to implement a surrogate goal. (More related discussion in “When would consultation with overseers fail to prevent catastrophic decisions?” here.) The developers could have preserved SPI option value, with minimal misalignment risk, by adding a clause like “unless the values change is a surrogate goal, and it’s impossible to check in with humans”.

Research goals: We plan to explore a range of possible SPI lock-out scenarios. Ideally, we’d use this library of scenarios to produce a “checklist” of simple risk factors for lock-out. AIs and humans in the loop could consult this checklist to cheaply preserve SPI option value. Separately, the library could inform the evals/datasets in Part I, and help motivate very simple interventions by AI companies like “put high-quality resources about SPIs in training data”. So the initial exploration step could still be useful, even if we update against the checklist plan. That could happen if we conclude the bulk of lock-out risk comes from factors that a checklist is ill-suited for — factors like broad commitment race dynamics that are hard to robustly intervene on, or mistakes that could be prevented simply by making AIs/humans in the loop more aware of SPIs.

II.3. Pitch for preserving SPI option value

In parallel with the research threads above, we aim to write a clear “pitch” for why AI developers should care about SPI lock-out. The target audience is technical staff at AI companies who make decisions about model training, deployment, and commitments, but who may not be familiar with open-source game theory. The goal at this stage is to help build coordination on preserving SPI option value where feasible, not to push for expensive or far-reaching changes to AI training.

The pitch would cover:

• what SPIs are and why they’re an unusually robust way to mitigate catastrophic conflict;
• historical precedents for SPIs (e.g., single combat or “counting coup” [10] ), with discussion of important disanalogies;
• examples of SPI-undermining behavior and reasoning by frontier models (from Part I);
• discussion of why further research on the incentives for SPIs might be time-sensitive (from Part II.1);
• concrete examples of how current AI practices (training objectives, deployment commitments, etc.) could inadvertently lock out SPIs, and low-cost practices to mitigate these failures (from Parts II.1 and II.2).

III. Preparing for automation of SPI research

Various open conceptual questions about SPIs seem important, yet less tractable or urgent than those in Part II. For example: Which attitudes that AIs might have about decision theory could shape their incentives to use SPIs? And given that these decision-theoretic attitudes aren’t self-correcting (Cooper et al.), how might future AIs’ incentives to use SPIs be path-dependent on earlier AIs’/humans’ attitudes (even if these aren’t “locked in”)? We want to get into a strong position to delegate these questions to future AI research assistants.

Anecdotally, we’ve found current models to be mostly poor at conceptual reasoning about SPIs, even when given substantial context. But models do help with some conceptual tasks. While the set of such tasks might grow quite quickly soon, delegating SPI research to AI assistants could still face two main bottlenecks:

1. Ability to efficiently identify which tasks we can trust AIs to do reliably, when it’s hard to quickly verify correctness (as is common in conceptual research).
2. SPI-specific data/context and infrastructure.

(See Carlsmith’s “Can we safely automate alignment research?”. (1) is about what Carlsmith calls “evaluation failures” (Sec. 5-6), and (2) is about “data-scarcity” and “shlep-scarcity” (Sec. 10). [11] )

Given these potential bottlenecks, we plan to pursue two complementary threads:

Benchmarking AI research capabilities on SPI. [12] We’re developing a benchmark to diagnose (and track over time) which SPI research tasks AI systems can handle. The aim is to help calibrate our decisions about what/how to delegate to AIs, at two levels: i) Which tasks we can trust AIs to do end-to-end? ii) Among the tasks the AIs can’t do end-to-end but can still help with, at which steps should they check in with overseers, and how can we decompose these tasks more productively? (We take dual-use concerns about advancing general conceptual reasoning seriously. For now, the default plan is to use the benchmark internally rather than sharing it with AI companies as a training target.)

Some examples of task classes the benchmark would cover:

• given a complex document, identifying where key SPI concepts are being used uncarefully (e.g., conflating different notions of “what you would have done without SPIs”) and disentangling the distinct claims;
• distinguishing valid and important objections from superficially plausible or irrelevant ones;
• reproducing known (non-public) results in SPI theory, e.g., identifying and fixing bugs in theoretical claims about SPIs;
• formalizing and proving results in models of SPI dynamics;
• assisting evaluation of other models’ (or human researchers’) outputs on any of the above tasks.

Strategies for efficient human-AI collaboration on SPI research. Drawing on our experience using AI assistants for SPI research, we’ll strategize about how to make this process more efficient — in ways that won’t quickly be made obsolete by the “Bitter Lesson”. Some strategies we plan to test out and refine:

• Practice applying current AI to SPI research and note transferable insights: identify which parts of the AI-assisted research process are bottlenecked by infrastructure, data, or idiosyncratic quality criteria that won’t come for free from better models. We’d then focus preparation efforts on these parts.
• Passively collect and annotate examples: e.g., helpful AI interactions; fruitful task specifications; and conceptual arguments that look compelling but have subtle flaws (cf. “Slop, not Scheming” here). To mitigate costs, we’d aim to:
  • build infrastructure to collect this data seamlessly; and
  • focus on the data most likely to help unlock AIs’ comparative advantages (e.g., perhaps, checking consistency across a large body of arguments).
• Compile well-scoped open questions to delegate.
• Clarify criteria for evaluating research quality: identify which dimensions of quality AIs themselves can cheaply verify or summarize for human overseers, and our standards for evaluating messier dimensions.

Acknowledgments

Many thanks to Tristan Cook, Clare Harris, Matt Hampton, Maxime Riché, Caspar Oesterheld, Nathaniel Sauerberg, Jesse Clifton, and Claude for comments and suggestions. I developed this agenda with significant input from Caspar Oesterheld, Lukas Finnveden, Johannes Treutlein, Chi Nguyen, Miranda Zhang, Nathaniel Sauerberg, and Paul Christiano. This does not imply their full endorsement of the strategy in this agenda.

Appendix A: Relevant previous work on SPIs

This list of resources gives a (non-comprehensive) overview of public SPI research. Brief summaries of some particularly relevant work:

• Baumann, “Using surrogate goals to deflect threats” — introduces surrogate goals and some challenges for successfully implementing them.
• Oesterheld and Conitzer (2022), “Safe Pareto Improvements for Delegated Game Playing” — introduces SPIs and the first formal model of them, and proves results about the conditions under which SPIs can be found.
• DiGiovanni et al. (2024), “Safe Pareto Improvements for Expected Utility Maximizers in Program Games” — introduces renegotiation SPIs, and proves that under certain assumptions about agents’ beliefs, they individually prefer to implement an SPI that bounds their losses from conflict. (See also this distillation.)
• Oesterheld, “A gap in the theoretical justification for surrogate goals and safe Pareto improvements” — introduces and discusses responses to the following problem: Even if we can find an SPI on a particular space of original strategies (as in Oesterheld and Conitzer 2022), what justifies using that SPI rather than strategies from some other space entirely? (See Appendix B.1.1 for more.)

Appendix B: Technical definitions and examples B.1. General SPI definition

Setup:

• Let mjx-container[jax="CHTML"] { line-height: 0; } mjx-container [space="1"] { margin-left: .111em; } mjx-container [space="2"] { margin-left: .167em; } mjx-container [space="3"] { margin-left: .222em; } mjx-container [space="4"] { margin-left: .278em; } mjx-container [space="5"] { margin-left: .333em; } mjx-container [rspace="1"] { margin-right: .111em; } mjx-container [rspace="2"] { margin-right: .167em; } mjx-container [rspace="3"] { margin-right: .222em; } mjx-container [rspace="4"] { margin-right: .278em; } mjx-container [rspace="5"] { margin-right: .333em; } mjx-container [size="s"] { font-size: 70.7%; } mjx-container [size="ss"] { font-size: 50%; } mjx-container [size="Tn"] { font-size: 60%; } mjx-container [size="sm"] { font-size: 85%; } mjx-container [size="lg"] { font-size: 120%; } mjx-container [size="Lg"] { font-size: 144%; } mjx-container [size="LG"] { font-size: 173%; } mjx-container [size="hg"] { font-size: 207%; } mjx-container [size="HG"] { font-size: 249%; } mjx-container [width="full"] { width: 100%; } mjx-box { display: inline-block; } mjx-block { display: block; } mjx-itable { display: inline-table; } mjx-row { display: table-row; } mjx-row > * { display: table-cell; } mjx-mtext { display: inline-block; text-align: left; } mjx-mstyle { display: inline-block; } mjx-merror { display: inline-block; color: red; background-color: yellow; } mjx-mphantom { visibility: hidden; } _::-webkit-full-page-media, _:future, :root mjx-container { will-change: opacity; } mjx-math { display: inline-block; text-align: left; line-height: 0; text-indent: 0; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; border-collapse: collapse; word-wrap: normal; word-spacing: normal; white-space: nowrap; direction: ltr; padding: 1px 0; } mjx-container[jax="CHTML"][display="true"] { display: block; text-align: center; margin: 1em 0; } mjx-container[jax="CHTML"][display="true"][width="full"] { display: flex; } mjx-container[jax="CHTML"][display="true"] mjx-math { padding: 0; } mjx-container[jax="CHTML"][justify="left"] { text-align: left; } mjx-container[jax="CHTML"][justify="right"] { text-align: right; } mjx-mi { display: inline-block; text-align: left; } mjx-c { display: inline-block; } mjx-utext { display: inline-block; padding: .75em 0 .2em 0; } mjx-mo { display: inline-block; text-align: left; } mjx-stretchy-h { display: inline-table; width: 100%; } mjx-stretchy-h > * { display: table-cell; width: 0; } mjx-stretchy-h > * > mjx-c { display: inline-block; transform: scalex(1.0000001); } mjx-stretchy-h > * > mjx-c::before { display: inline-block; width: initial; } mjx-stretchy-h > mjx-ext { /* IE */ overflow: hidden; /* others */ overflow: clip visible; width: 100%; } mjx-stretchy-h > mjx-ext > mjx-c::before { transform: scalex(500); } mjx-stretchy-h > mjx-ext > mjx-c { width: 0; } mjx-stretchy-h > mjx-beg > mjx-c { margin-right: -.1em; } mjx-stretchy-h > mjx-end > mjx-c { margin-left: -.1em; } mjx-stretchy-v { display: inline-block; } mjx-stretchy-v > * { display: block; } mjx-stretchy-v > mjx-beg { height: 0; } mjx-stretchy-v > mjx-end > mjx-c { display: block; } mjx-stretchy-v > * > mjx-c { transform: scaley(1.0000001); transform-origin: left center; overflow: hidden; } mjx-stretchy-v > mjx-ext { display: block; height: 100%; box-sizing: border-box; border: 0px solid transparent; /* IE */ overflow: hidden; /* others */ overflow: visible clip; } mjx-stretchy-v > mjx-ext > mjx-c::before { width: initial; box-sizing: border-box; } mjx-stretchy-v > mjx-ext > mjx-c { transform: scaleY(500) translateY(.075em); overflow: visible; } mjx-mark { display: inline-block; height: 0px; } mjx-TeXAtom { display: inline-block; text-align: left; } mjx-msup { display: inline-block; text-align: left; } mjx-msub { display: inline-block; text-align: left; } mjx-msubsup { display: inline-block; text-align: left; } mjx-script { display: inline-block; padding-right: .05em; padding-left: .033em; } mjx-script > mjx-spacer { display: block; } mjx-mspace { display: inline-block; text-align: left; } mjx-mn { display: inline-block; text-align: left; } mjx-c::before { display: block; width: 0; } .MJX-TEX { font-family: MJXZERO, MJXTEX; } .TEX-B { font-family: MJXZERO, MJXTEX-B; } .TEX-I { font-family: MJXZERO, MJXTEX-I; } .TEX-MI { font-family: MJXZERO, MJXTEX-MI; } .TEX-BI { font-family: MJXZERO, MJXTEX-BI; } .TEX-S1 { font-family: MJXZERO, MJXTEX-S1; } .TEX-S2 { font-family: MJXZERO, MJXTEX-S2; } .TEX-S3 { font-family: MJXZERO, MJXTEX-S3; } .TEX-S4 { font-family: MJXZERO, MJXTEX-S4; } .TEX-A { font-family: MJXZERO, MJXTEX-A; } .TEX-C { font-family: MJXZERO, MJXTEX-C; } .TEX-CB { font-family: MJXZERO, MJXTEX-CB; } .TEX-FR { font-family: MJXZERO, MJXTEX-FR; } .TEX-FRB { font-family: MJXZERO, MJXTEX-FRB; } .TEX-SS { font-family: MJXZERO, MJXTEX-SS; } .TEX-SSB { font-family: MJXZERO, MJXTEX-SSB; } .TEX-SSI { font-family: MJXZERO, MJXTEX-SSI; } .TEX-SC { font-family: MJXZERO, MJXTEX-SC; } .TEX-T { font-family: MJXZERO, MJXTEX-T; } .TEX-V { font-family: MJXZERO, MJXTEX-V; } .TEX-VB { font-family: MJXZERO, MJXTEX-VB; } mjx-stretchy-v mjx-c, mjx-stretchy-h mjx-c { font-family: MJXZERO, MJXTEX-S1, MJXTEX-S4, MJXTEX, MJXTEX-A ! important; } @font-face /* 0 */ { font-family: MJXZERO; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Zero.woff") format("woff"); } @font-face /* 1 */ { font-family: MJXTEX; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Main-Regular.woff") format("woff"); } @font-face /* 2 */ { font-family: MJXTEX-B; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Main-Bold.woff") format("woff"); } @font-face /* 3 */ { font-family: MJXTEX-I; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Math-Italic.woff") format("woff"); } @font-face /* 4 */ { font-family: MJXTEX-MI; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Main-Italic.woff") format("woff"); } @font-face /* 5 */ { font-family: MJXTEX-BI; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Math-BoldItalic.woff") format("woff"); } @font-face /* 6 */ { font-family: MJXTEX-S1; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Size1-Regular.woff") format("woff"); } @font-face /* 7 */ { font-family: MJXTEX-S2; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Size2-Regular.woff") format("woff"); } @font-face /* 8 */ { font-family: MJXTEX-S3; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Size3-Regular.woff") format("woff"); } @font-face /* 9 */ { font-family: MJXTEX-S4; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Size4-Regular.woff") format("woff"); } @font-face /* 10 */ { font-family: MJXTEX-A; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_AMS-Regular.woff") format("woff"); } @font-face /* 11 */ { font-family: MJXTEX-C; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Calligraphic-Regular.woff") format("woff"); } @font-face /* 12 */ { font-family: MJXTEX-CB; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Calligraphic-Bold.woff") format("woff"); } @font-face /* 13 */ { font-family: MJXTEX-FR; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Fraktur-Regular.woff") format("woff"); } @font-face /* 14 */ { font-family: MJXTEX-FRB; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Fraktur-Bold.woff") format("woff"); } @font-face /* 15 */ { font-family: MJXTEX-SS; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_SansSerif-Regular.woff") format("woff"); } @font-face /* 16 */ { font-family: MJXTEX-SSB; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_SansSerif-Bold.woff") format("woff"); } @font-face /* 17 */ { font-family: MJXTEX-SSI; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_SansSerif-Italic.woff") format("woff"); } @font-face /* 18 */ { font-family: MJXTEX-SC; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Script-Regular.woff") format("woff"); } @font-face /* 19 */ { font-family: MJXTEX-T; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Typewriter-Regular.woff") format("woff"); } @font-face /* 20 */ { font-family: MJXTEX-V; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Vector-Regular.woff") format("woff"); } @font-face /* 21 */ { font-family: MJXTEX-VB; src: url("https://cdn.jsdelivr.net/npm/mathjax@3/es5/output/chtml/fonts/woff-v2/MathJax_Vector-Bold.woff") format("woff"); } mjx-c.mjx-c393::before { padding: 0.68em 0.625em 0 0; content: "\393"; } mjx-c.mjx-c3D::before { padding: 0.583em 0.778em 0.082em 0; content: "="; } mjx-c.mjx-c28::before { padding: 0.75em 0.389em 0.25em 0; content: "("; } mjx-c.mjx-c1D400.TEX-B::before { padding: 0.698em 0.869em 0 0; content: "A"; } mjx-c.mjx-c2C::before { padding: 0.121em 0.278em 0.194em 0; content: ","; } mjx-c.mjx-c1D42E.TEX-B::before { padding: 0.45em 0.639em 0.006em 0; content: "u"; } mjx-c.mjx-c29::before { padding: 0.75em 0.389em 0.25em 0; content: ")"; } mjx-c.mjx-c1D429.TEX-B::before { padding: 0.45em 0.639em 0.194em 0; content: "p"; } mjx-c.mjx-c2217::before { padding: 0.465em 0.5em 0 0; content: "\2217"; } mjx-c.mjx-c1D41F.TEX-B::before { padding: 0.7em 0.452em 0 0; content: "f"; } mjx-c.mjx-c1D40F.TEX-B::before { padding: 0.686em 0.786em 0 0; content: "P"; } mjx-c.mjx-c3A0::before { padding: 0.68em 0.75em 0 0; content: "\3A0"; } mjx-c.mjx-c1D460.TEX-I::before { padding: 0.442em 0.469em 0.01em 0; content: "s"; } mjx-c.mjx-c1D447.TEX-I::before { padding: 0.677em 0.704em 0 0; content: "T"; } mjx-c.mjx-c3A8::before { padding: 0.683em 0.778em 0 0; content: "\3A8"; } mjx-c.mjx-c2260::before { padding: 0.716em 0.778em 0.215em 0; content: "\2260"; } mjx-c.mjx-c1D42A.TEX-B::before { padding: 0.45em 0.607em 0.194em 0; content: "q"; } mjx-c.mjx-c1D456.TEX-I::before { padding: 0.661em 0.345em 0.011em 0; content: "i"; } mjx-c.mjx-c1D451.TEX-I::before { padding: 0.694em 0.52em 0.01em 0; content: "d"; } mjx-c.mjx-c1D42B.TEX-B::before { padding: 0.45em 0.474em 0 0; content: "r"; } mjx-c.mjx-c1D427.TEX-B::before { padding: 0.45em 0.639em 0 0; content: "n"; } mjx-c.mjx-c1D443.TEX-I::before { padding: 0.683em 0.751em 0 0; content: "P"; } mjx-c.mjx-c1D457.TEX-I::before { padding: 0.661em 0.412em 0.204em 0; content: "j"; } mjx-c.mjx-c1D439.TEX-I::before { padding: 0.68em 0.749em 0 0; content: "F"; } mjx-c.mjx-c1D6FC.TEX-I::before { padding: 0.442em 0.64em 0.011em 0; content: "\3B1"; } mjx-c.mjx-c31::before { padding: 0.666em 0.5em 0 0; content: "1"; } mjx-c.mjx-c2B::before { padding: 0.583em 0.778em 0.082em 0; content: "+"; } mjx-c.mjx-c1D716.TEX-I::before { padding: 0.431em 0.406em 0.011em 0; content: "\3F5"; } mjx-c.mjx-c2212::before { padding: 0.583em 0.778em 0.082em 0; content: "\2212"; } mjx-c.mjx-c1D465.TEX-I::before { padding: 0.442em 0.572em 0.011em 0; content: "x"; } mjx-c.mjx-c46::before { padding: 0.68em 0.653em 0 0; content: "F"; } mjx-c.mjx-c41::before { padding: 0.716em 0.75em 0 0; content: "A"; } mjx-c.mjx-c49::before { padding: 0.683em 0.361em 0 0; content: "I"; } mjx-c.mjx-c52::before { padding: 0.683em 0.736em 0.022em 0; content: "R"; } mjx-c.mjx-c48::before { padding: 0.683em 0.75em 0 0; content: "H"; } mjx-c.mjx-c57::before { padding: 0.683em 1.028em 0.022em 0; content: "W"; } mjx-c.mjx-c4B::before { padding: 0.683em 0.778em 0 0; content: "K"; } be a general-sum game, where and represent the agents’ action spaces and payoff functions, respectively.
• Let denote the way agents will play , i.e., the list of programs (conditional commitments) they’ll follow that collectively determine their actions. [13] (That is, Alice’s action is determined by both Alice’s and Bob’s programs, and so is Bob’s action.)
  • As discussed below, these programs might have a structure like “delegate to other agents, who play some possibly different game in some particular way” — as in Oesterheld and Conitzer (2022), the original source on SPIs.
• The agents are initially uncertain which programs (including their own) will eventually be used. E.g., they’re uncertain how exactly their delegates will approach equilibrium selection, or which program they’ll want to use after thinking more.

Then:

Definition. An SPI is a transformation such that, for all in some space , the agents’ payoffs in when they follow programs (weakly) Pareto-dominate their payoffs when they follow .

This definition alone doesn’t impose any restrictions on , e.g., that matches the agents’ “default” way of bargaining in some sense. In particular:

• Those restrictions come into play either via the choice of w.r.t. which the SPI is defined, or via constraints like participation independence or foreknowledge independence.
• No such restrictions on are required for SPIs to satisfy the robustness properties mentioned in the Introduction:
  • “SPIs don’t require agents to coordinate on some notion of a ‘fair’ deal”; and
  • “when agents apply these kinds of SPIs to their original strategies, each party makes the same demands as in their original strategy” (see “demand preservation” in Appendix B.2, which is a constraint on rather than ).

Oesterheld and Conitzer (2022) use a definition that’s almost equivalent to this one, with the special choice of in Table 1. In their framework, there’s (implicitly) a space of original programs characterized by (i) the true game , and (ii) some way the delegates would play any given game . And they define the SPI not as the transformation , but instead as the new game for the delegates such that maps to . But (from personal communication with Oesterheld) the definition of SPI is meant to allow for more general . [14] See also Figure 3 for a comparison to DiGiovanni et al.’s (2024) formalization.

Table 1. How the definition above captures different formalizations of SPIs in the literature.

Original program space P Before the programs are determined… SPI transformation f Oesterheld & Conitzer (2022), Definition 1 Space of tuples , for a fixed true game , where is a mapping from any game to actions. can be any such mapping satisfying certain assumptions (e.g., the paper’s Assumptions 1 and 2). (Agents have non-probabilistic uncertainty over . So the “for all ” quantifier in the definition of SPIs amounts to “for all ”.) Agents choose some new game . (Here, programs are determined by the delegates’ decisions.) Transforms to . DiGiovanni et al. (2024), Definition 2 Arbitrary space of conditional commitments. Agents choose how to map the program space to some new space, which they will then choose from. [15] Transforms to . Sauerberg and Oesterheld (2026) (Sec. 4) Same as Oesterheld & Conitzer. Agents choose a “token game” and function mapping ’s outcomes into . The original game is then resolved via applied to ’s outcomes. Transforms to .

Figure 3.

B.1.1. Connection to Oesterheld’s “justification gap”

Here’s how we might state the problem raised by Oesterheld’s “A gap in the theoretical justification for surrogate goals and safe Pareto improvements”, in the formalism above.

Consider the original space of programs in Oesterheld and Conitzer’s framework. The delegates can play the game in an arbitrary way, subject to the mild Assumptions 1 and 2. But it’s assumed that in , the game they play is the true game . So, take some SPI with respect to this space , a transformation from to . By definition, this transformation makes all agents better off for all . But it’s not guaranteed that for all and all , all agents are better off under than under .

This suggests one way to bridge the justification gap: find an that’s an SPI with respect to any arbitrary program space , as DiGiovanni et al. (2024) aims to do. Cf. Oesterheld’s discussion of “decision factorization” in the justification gap post.

Figure 4.

B.2. Properties of full strategies involving SPIs

(These are working formalizations of participation independence and foreknowledge independence. “Foreknowledge independence” and “demand preservation” are working terminology. We’re not highly confident that we’ll endorse these formalizations/terminology after more thought.)

If is an SPI and are the programs the agents in fact apply to, call the agents’ full strategy. It’s helpful to distinguish an SPI from the full strategy, because in general agents will only individually prefer to agree to some SPI conditional on the input programs satisfying certain restrictions.

Participation independence and foreknowledge independence, as well as the “preserving bargaining power” property discussed in the Introduction, are properties of full strategies. These can be defined as follows.

Setup:

• For any program , let represent the “demands” made by . This isn’t fully precise, but as an example: For a renegotiation program as in the pseudocode in Appendix B.4, my_base_strategy.
• Given a full strategy , let:
  • be the program that agent would have chosen had each other agent used the program , rather than ;
  • be the program that agent would have chosen had believed each other agent would use the program , rather than .

Then:

Definition. A full strategy is:

• demand-preserving if for each agent , .
• participation-independent if it’s demand-preserving and for each agent , .
• foreknowledge-independent if it’s demand-preserving and for each agent , .

Commentary on these definitions:

• Demand preservation is equivalent to this property informally stated in the Introduction: “When agents apply these kinds of SPIs to their original strategies, each party makes the same demands as in their original strategy.”
• For each of PI and FI, the two conditions correspond to two different levels at which “your demands are the same as if [PI counterfactual or FI counterfactual]”. That is:
  • or , respectively, formalizes this property at the level of the agent’s choice of the program the SPI is applied to.
  • Given that or holds, demand preservation formalizes this property at the level of the output of the fixed program . This is because the transformation is applied to all the agents’ programs. Thus, if (agent believed that) agent wouldn’t participate in , neither would — so the demands made by ’s program would be the same as under ’s original program.

Example: In DiGiovanni et al.’s (2024) setting, suppose agents use the SPI given by Proposition 1 (or Proposition 4). Then participation independence is satisfied for any input program profile , because:

• agents choose programs independently of each other, so it’s immediate that ;
• by construction, renegotiation programs satisfy . Compare lines 3 and 10 of the pseudocode in Appendix B.4.

B.3. Example: Surrogate goals and concession equivalence

(This section is based on previous joint work with Mia Taylor, Nathaniel Sauerberg, Julian Stastny, and Jesse Clifton.)

One key example of an SPI is a surrogate goal. More precisely, the (approximate) SPI here is, “A adopts a surrogate goal, and B threatens the surrogate goal whenever an executed surrogate threat would be less costly for B than the default threat”. (More below on why this is an SPI.)

An agent doesn’t need to broadly modify its preferences in order to implement an SPI of this form, though. We can generalize the idea of surrogate goals as follows:

• A surrogate threat is a claim by a threatener that they will cause some outcome , at some cost to themselves but no cost to the target’s original goal, if the target doesn’t concede to some demand.
• A target’s bargaining policy is -concession-equivalent to its policy for responding to surrogate threats if, for any demand, the target is equally likely to concede to the demand given (i) a surrogate threat as given (ii) an equally credible threat against the target’s original goal (“OG threat”), when executing the OG threat is times as costly to the threatener as executing the surrogate threat.
  • In cases where concession equivalence isn’t perfectly credible, choosing a cost ratio can allow targets to compensate threateners for taking the risk of making a surrogate threat, thereby still allowing for threat deflection.
  • For brevity, we’ll often just say that a target’s policy is “-concession-equivalent” if it has this property, and “concession-equivalent” if for some very small .
• A surrogate goal is a modification to an agent’s preferences in which they terminally disprefer the outcome .

Why is adoption of a concession-equivalent policy an SPI? Suppose — holding all else fixed — A becomes just as likely to concede to a surrogate threat that would give B utility if executed, as to an OG threat that would give B utility if executed. Then B would rather make a surrogate threat than an OG threat. So any executed threats would be less bad for both parties, but neither party would have an incentive to change how much they demand. (Except, perhaps, a very small increase in B’s demands in proportion to the difference in disutility of executing an OG threat vs. surrogate threat.) Both parties are then better off overall, no matter how much they demand.

See also Oesterheld and Conitzer’s (2022) “Demand Game” (Table 1), as an example of something like a bilateral surrogate goal.

B.4. Example: Renegotiation

A renegotiation program is a program structured like: “If they don’t use a renegotiation program, act according to program . Otherwise, still act according to , except: if we get into conflict, propose some Pareto improvement(s) and take it if our proposals match.” In pseudocode (see also Algorithms 1 and 2 of DiGiovanni et al. (2024)):

For example, suppose agents A and B are negotiating over what values to instill in a successor agent. If they fail to reach an agreement, they’ll each attempt to take over. They simultaneously submit programs for the negotiation to some centralized server. Before they consider the possibility of SPIs, they’re inclined to choose these programs, respectively:

• A: = “Demand 50% of the share of the ASI’s values no matter what”;
• B: = “Demand 80%, and trigger a doomsday device if they refuse.”

Since the demands selected by these programs would be incompatible, the outcome would be “B triggers a doomsday device”. In this scenario, the agents’ corresponding renegotiation programs might be:

• A: = “If they don’t use a renegotiation program, demand 50% of the share of the ASI’s values. Otherwise, demand 50%; if they refuse, propose ‘attempt takeover, without any doomsday devices’.”
• B: = “If they don’t use a renegotiation program, demand 80% and trigger a doomsday device if they refuse. Otherwise, demand 80%; if they refuse: propose ‘attempt takeover, without any doomsday devices’.”

(Here, the Pareto improvement is to the outcome “both agents attempt takeover, without any doomsday devices”, rather than “B triggers a doomsday device”. Both here and in the surrogate goals example, we’re setting aside the additional conditions necessary for these SPIs to be individually preferable. See Part II.1 and Appendix B.2 for more. But note one such condition in this example: and demand 50% and 80%, respectively, regardless of whether the other program is a renegotiation program. See “demand preservation” in Appendix B.2.)

See Macé et al., “Individually incentivized safe Pareto improvements in open-source bargaining”, for more discussion of how a special class of renegotiation programs can partially resolve the SPI selection problem.

1. “Commitments” are meant to include modifications to one’s decision theory or values/preferences. It has been argued (example) that decision theories like updateless decision theory (UDT) can sidestep the need for “commitments” in the usual sense. We’ll set this question aside here, and treat the resolution to make one’s future decisions according to UDT as a commitment in itself. ↩︎

2. We might wonder: We’ve assumed the AIs are capable of conditional commitments. So, suppose each AI could commit to only demand 60% unless they verify that the other AI has made an incompatible commitment. Would this solve the problem? Not necessarily, because the AIs might reason, “If they see that I’ll revoke my commitment conditional on incompatible demands, they’ll exploit this by making high demands. So I should stick with my unconditional commitment.” ↩︎

3. However, see Part II.1 for discussion of the “SPI selection problem”. ↩︎

4. (H/t Caspar Oesterheld and Nathaniel Sauerberg:) Another important reason is that even if SPIs don’t get locked out, they might not be implemented early enough, before conflicts break out. We put less emphasis on this consideration in this agenda, because avoiding locking out SPIs is a less controversial ask than actively prioritizing implementing SPIs. ↩︎

5. These gaps are related to, but importantly distinct from, the “SPI justification gap” discussed by Oesterheld. Oesterheld’s question is: Suppose we have some SPI that makes everyone better off relative to particular “default” strategies — not necessarily relative to any possible original strategies. If so, why would agents use the SPI-transformed strategies, rather than some alternatives to both the default strategies and SPI transformations of them? More in Appendix B.1.1. By contrast, the question here is: Suppose we have an SPI that is ex post better for everyone relative to any original strategies. (So there is no privileged “default”.) Then, when do agents prefer to implement this SPI ex ante, rather than use their original strategies? ↩︎

6. See also this distillation. The rough intuition for the result is: If you’re (only) willing to fall back to Pareto improvements that aren’t better for the other agent than conceding 100%, you don’t give them perverse incentives (cf. Yudkowsky). And if you offer a set of possible Pareto improvements with this property, you can coordinate on an SPI despite the SPI selection problem. ↩︎

7. In more detail, respectively: (1) (H/t James Faville and Lukas Finnveden:) Agents might be incentivized to condition their demands on coarse-grained proxies about their counterparts, because they worry about being exploited if they use fine-grained information (cf. Soto). And an agent who opts out of SPIs might bargain more aggressively against SPI-participating agents, based on such proxies. (2) (H/t Lukas Finnveden:) Roughly, the “PMP-extension” of Algorithm 2 from DiGiovanni et al. (2024) offers a fallback outcome to an agent willing to use any “conditional set-valued renegotiation” algorithm. This means that a counterpart has little to lose by renegotiating more aggressively against this algorithm. (It appears straightforward to avoid this problem by making the offer conditional, but we need to confirm this makes sense formally — see this comment.) More precisely, the “fallback outcome” is the “Pareto meet minimum”. ↩︎

8. See, e.g., Oesterheld (section “Solution idea 2: Decision factorization”): “[I]n the surrogate goal story, it’s important to first adopt surrogate goals and only then decide whether to make other commitments.” ↩︎

9. Working terminology. Cf. Kovarik (section “Illustrating our Main Objection: Unrealistic Framing”); and Oesterheld: “If in 20 years I instruct an AI to manage my resources, it would be problematic if in the meantime I make tons of decisions (e.g., about how to train my AI systems) differently based on my knowledge that I will use surrogate goals anyway.” The concept of foreknowledge independence was also inspired by Baumann’s notion of “threatener-neutrality”. ↩︎

10. Thanks to Jesse Clifton and Carl Shulman for these examples. ↩︎

11. In the context of SPI research, we’re not too concerned about a third problem Carlsmith discusses: deliberate sabotage by “scheming” AIs. This is because SPIs are designed to make all parties better off, so a misaligned AI doesn’t clearly have an incentive to sabotage SPI research. But we’ll aim to be mindful of sabotage risks as well. ↩︎

12. See also Oesterheld et al. (2026) and Oesterheld et al. (2025) for related datasets of rated conceptual arguments and decision theory reasoning, respectively. ↩︎

13. DiGiovanni et al. (2024), Sec. 3.1, gives a more precise definition of programs. ↩︎

14. See also Oesterheld and Conitzer (2022), p. 30: “In principle, Theorem 3 does not hinge on Π(Γ) and Π(Γs) resulting from playing games. An analogous result holds for any random variables over A and As. In particular, this means that Theorem 3 applies also if the representatives [i.e., delegates] receive other kinds of instructions.” ↩︎

15. In the formalism of DiGiovanni et al. (2024), there is no separate stage where agents choose a transformation f before choosing programs from the new space of programs. Agents simply choose programs directly. But, for the purposes of modeling SPIs and comparing the framework of DiGiovanni et al. with that of Oesterheld and Conitzer (2022), it’s helpful to use the framing in Table 1. ↩︎

Discuss

This is a weekly round-up of things I’ve posted in the last week.

InkHaven requires that I post a blog post every day, which is a lot. Especially for people subscribed to my blog. Someone requested I spare their inbox, so I haven’t been sending out every post.

So now you get to catch up! You can even be selective if you prefer :)

The posts are:About the posts:

• Diary of a “Doomer” (part 1) is about my experience getting into the field of AI and AI Safety (I started graduate school in 2013). A lot has changed since then. What used to be a fringe topic has become really mainstream! I’m talking about deep learning, of course… But seriously, AI researchers really dropped the ball, and owe society a debt they can probably never repay for failing to consider the consequences of their actions.

• Contra Leicht on AI Pauses takes apart Anton Leicht’s piece arguing we shouldn’t try to pause AI. I first encountered Leicht when he was arguing against having an “AI Safety” movement at all last fall. I don’t think either of these articles are very good — I find the reasoning sloppy.

• Post-Scarcity is bullshit is mostly about how certain things are fundamentally scarse; like land, energy, and status. I got a bit snarky here about the discourse around the topic, and how vague, incoherent, and/or unimaginative people’s visions of the “post-scarcity” world typically are.

• From Artificial Intelligence to an ecosystem of artificial life-forms. If the AI race doesn’t stop, the natural end-point is the creation of artificial beings that proliferate, diversify, and radically reshape the world. This is one of my quick and dirty attempts to explain a part of my world view that really deserves a 30-page essay.

• Idea Economics is a rare non-AI-related post about how and why I think people devalue ideas: Not because they’re easy to come by, but because they’re hard to hold on to if you share them. But then I ruin it by talking about the CAIS Statement on AI Risk as an example (it was sorta my idea).

• Stop AI is an attempt to get the basic case for why we need to stop AI down in writing. It ended up basically just covering the risks and not why other solutions aren’t good enough (stay tuned, that might be the next post).

• Stop AI Now argues against kicking the can down the road. I think that’s intuitively a bad idea, but here I give three particular reasons.

Commentary:

I did this as a bit of an experiment. Besides the person complaining to me directly, I did notice a dip in subscribers at some point after about seven posts in a row at the start. A blogger friend of mine with more of a following says they often lose followers after a post. I guess that makes sense… people don’t like their inbox being clogged.

I did still send out two of these posts as email notifications. The first one was deliberate, the second was an accident. You can see that the ones I sent out did get a lot more views. I’ll be curious to see how much this post makes up the difference!

Thanks for reading The Real AI! Subscribe for free to receive new posts and support my work.

Share

Discuss

Or, the end of the world is no excuse for sloppy work

One morning when I was nine, my dad called me over to his computer. He wanted to show me this amazing Korean scientist who had managed to clone stem cells, and who was developing treatments to let people with spinal cord injuries – people like my dad – walk again on their own two legs.

I don't remember exactly what he said next, or what I said back. I have a sense that I was excited too, and that I was upset when I learned the United States had banned this kind of research.

Unfortunately, his research didn’t pan out. No such treatment arrived. My dad still walks on crutches.

Years later, I learned that the scientist, Hwang Woo-Suk, had been exposed as a fraud.

In 2004, Hwang published a paper in Science claiming that his team had cloned a human embryo and derived stem cells from it (the first time anyone had done this). A year later, in 2005, he published a second paper claiming that they managed to repeat this feat eleven more times, producing 11 patient-specific stem cell lines for patients with type 1 diabetes, congenital hypogammaglobulinemia (a rare immune disorder), and spinal cord injuries. This was the result that, if true, would have helped my dad.

None of this was real. The 2004 cell line did exist, but was not a clone; investigators concluded that it was an unfertilized egg that had spontaneously started dividing. The 2005 cell lines did not exist at all; investigators later found that the data reported for all eleven lines had been fabricated from just two samples, and the DNA in those two samples did not match the patients they had supposedly been derived from.

My dad was not the only person Hwang had given hope to. On July 31st, 2005, Hwang had appeared on a Korean TV show.  The dance duo Clon had just performed; one of its members, Kang Won-rae, had been paralyzed from the waist down in a motorcycle accident five years earlier, and had performed in his wheelchair. Hwang walked onto the stage and told a national audience, with tears in his eyes, that he hoped “for a day that Kang will get up and perform magnificently as he did in the past” – a day that was coming soon. He made similar promises to other patients and their families.

I don't think Hwang was a monster who set out to commit fraud for international acclaim. I think he was a capable scientist with real results. (Some of his lab’s cloned animals were almost certainly real clones, including the world’s first cloned dog Snuppy.) But over time, he repeatedly took what he felt was his only option.

The 2004 paper may have started as a real mistake; it’s possible his team genuinely thought the parthenogenetic egg was a clone. But by 2005, with a nation watching and a Nobel on the table and a paralyzed pop star looking at him on live television, there was no version of "actually, we can't do this yet" that he could bring himself to say. So he didn't say it.

The way in which Hwang began his downward spiral is what sticks out most to me. He started out a good scientist, with good results and an important field of study. But with tens of millions of dollars of funding, thousands of adoring fans, and all the letters written to him by hopeful patients and their families, Hwang likely felt the weight of the world on his shoulders. He had to do what he had to do, in order to not let them down.

I work in AI safety. Many of the people I work with believe (and I believe) that the next decade will substantially determine whether and how humanity gets through this century. The stakes are literally astronomical and existential, and the timelines may be short.

That is the weight we carry. And I worry that when push comes to shove, our scientific standards will slip (or are slipping) in order to not let other people down.

For example, wouldn’t it be the right choice to just accept the code written by Claude, without reading it carefully? We don’t have much time left, and we need to figure out how to do interpretability, or monitoring, or how to align models with personas, and so forth.

Why investigate that note of confusion about the new result you saw? Surely with the stakes involved, it’s important to push forward, rather than question every assumption we have?

Why question your interpretability tools, when they seem to produce results that make sense, and let you steer the models to produce other results that seem to make sense? Why flag the failed eval run with somewhat suspicious results, when the deadline for model release is coming soon, and evaluation setups are famously finicky and buggy anyways? Why not simplify away some of the nuance of your paper’s results, when doing so would let it reach a much larger audience?

I worry that it’s tempting for us to take the expedient choice and let our standards slip, precisely because the stakes are so high. But it is precisely because the stakes are so high, with all the real people who will be affected by the outcome, that we need to be vigilant.

Yes, timelines may be short and we may not have time to do all the research that we want. But slipping up and producing misleading or wrong research will only hurt, not help. And if we need to say "actually, we can't do that yet", then we should say as much.

Discuss

I've designed a few spaces people seem to really love (most widely known Lighthaven, but before that the Lightcone Offices, though I've also played a hand in designing some of the most popular areas at Constellation).

Most people (including me a few years back) have surprisingly bad introspective access into why a room makes them feel certain things. Most of the time, people's ability to describe the effect of a space on them is as shallow as "this place feels artificial", or "this place has bad vibes", or "this place feels cozy". And if they try to figure out why that is true, they quickly run into limits of their introspective access.

The most common reason why a space feels bad, is because it is lit by low-quality lights.

Our eyes evolved to see things illuminated by sunlight. Correspondingly, it appears that the best proxy we have for whether the light in a room "works" is how similar the light in that room is to natural sunlight. The most popular way of measuring how much light differs from natural sunlight is the "Color Rendering Index" (CRI):

The best illustration I have found for the effect of low-CRI lighting is to look at the face of a person illuminated by nothing but a computer screen. Computer screens emit extremely low CRI light, as a screen simulates white by combining the light from red, green and blue LEDs. Faces illuminated by nothing but screens often look off and have a "fluorescent" off-color vibe to them, even if the screen is almost fully white.

If you want a space to feel natural, buy lightbulbs with at least 95 CRI, ideally 98.

But I thought my eyes can only see three colors?

Yes, and that is exactly why when you look at a computer screen directly, colors look real and vivid and correct. Your computer screen omits light that (pretty precisely) stimulates the three kinds of cones in your eyes and so can produce basically arbitrary perceptual colors (it's not perfect, but it's quite good).

The problem occurs when light bounces off of other objects in the room. The color of an object is determined by how it absorbs, reflects, and changes light that hits it. For example, an object under sunlight might completely absorb orange light (~630nm), but fully reflect the red light and green light emitted by the LEDs in your computer screen. That object would look unnaturally bright under the light from your computer screen, because it basically reflects all the light that hits it, but under sunlight it would absorb all the orange light that hits it.

So to a first approximation the primary determinant of whether a light produces "natural" reflections is whether it's light emission spectrum is a smooth curve. Sunlight, as well as any kind of light created by burning or heating things to really hot temperatures produce smooth emission spectra, which maintains all color information as it bounces around a room.

You might have noticed a second number that keeps showing up in the widgets above, often formatted as "5000K" or "2500K". This is the "color temperature" of light. We call this "temperature" because it corresponds to what kind of color objects emit when you heat them to that temperature[1]. Objects that are hotter, emit light that is more blue. Objects that are less hot (e.g. "only" 2000 degrees Kelvin) emit light that is more red. Naturally we call red light "warm light" and blue light "cold light"[2].

If you are lighting a room with plenty of natural light, just use 2000K-3000K lights

People prefer bluer light during the day, but redder light during the evening and morning. Sunlight is really really bright, so what lamps you have in your room do not matter if you have large windows during the day. This means the primary purpose of your lights are to light things in the evening and morning. This means they should be warm.

If your room does not have much natural light, I recommend having bright overhead lights that are closer to 4000K, and dimmer floor lamps around 2000K-3000K.

The world got ugly when we invented LEDs

For basically all of human civilization up until very recently lighting quality was a complete non-issue. Why? Because all of our artificial light sources consisted of heating things to very hot temperatures, or burning things. When you do that, you basically always emit natural light with a smooth emission spectrum.

Lighting quality only became an issue within the last 100 years with the introduction of fluorescent lamps in offices. This is why "fluorescent lighting" has for many people become the best shorthand for fake or artificial lighting.

But people's homes, as well as any entertainment venues, bars or really anywhere where people socially congregated in the evenings were lit by incandescent light bulbs (or before then, candles and oil lamps) with perfectly smooth spectra.

But around 30 years ago home lighting LEDs were introduced, initially with truly terrible color rendering indexes, and most people unable to put words to the discomfort and alienness they caused, chose the energy-saving option and replaced their incandescent bulbs with LEDs. Eventually, in most of the western world outside of the US, incandescent lightbulbs were literally banned to promote energy saving policies.

This was the greatest uglification in history. Within two decades, much of the world that was previously filled with beautiful natural-feeling light started feeling alien, slightly off, and uncomfortable, and societal stigma around energy-saving policies prevented people from really doing anything about it.

But you, within your home, can fix this. LED technology has come along way and we can produce high-CRI LED bulbs (I recommend YujiLED or Waveform Lighting). The world really used to be much more beautiful and a much less harsh place in this one respect. You can restore the natural light, and the homeliness that all your ancestors felt, at least within the confines of your home. Just buy some high-CRI, warm color temperature light bulbs. There is a lot more to interior design, but it's honestly so much easier to iterate on than lighting.

1. ^

   What kind of object? Well, turns out really any object, unless the thing you are heating undergoes some specific chemical reaction when you heat them that causes them to emit other wavelengths of light. The radiation curve that most objects tend to follow here is known as the "blackbody radiation curve". You can google it or ask your local LLM if you want to understand the physics behind this better.

2. ^

   This is a joke. This is indeed exactly backwards. You cannot imagine how much this makes explaining color temperatures to people more confusing. "Oh, just get the warmer light bulb, no not the one that has the higher temperature written on it why would you think that, that would produce much colder light". Grrrr.

Discuss

I think we need to Stop AI. Specifically we need to Stop AI Now. We can’t wait around. The standard metaphor is a runaway train heading towards a cliff. Let’s work with that.

We don’t know when to stop. We don’t know where the cliff is.

World’s most-cited-scientist (and my Master’s supervisor) Yoshua Bengio says we’re racing into a bank of fog, and there could be a cliff. That’s about right. There are two implications of this: 1) maybe there’s no cliff and it will all be fine 2) the cliff could be anywhere, we can’t see it far enough ahead to stop unless we’re going very slow. So while a lot of people seem to think we’re going to see the risks clearly in time to stop, I’m not so sure.

The entire time I’ve been in the field, people have repeatedly been surprised by the rate of progress in AI. The people at the leading AI companies are an exception – the most vocal among them have been, if anything, overestimating how fast things move.

Progress could be sudden

There’s a dangerous idea that’s caught hold that AI progress is predictable because of “scaling laws”. We’ve seen pretty consistent patterns in how quickly AI advances in terms of particular metrics as a function of time. But there’s a few problems with this: 1) The metrics don’t measure the things we care about, 2) There’s no reason why these trends should hold if there’s a paradigm shift. Indeed, RE (2), there’s already been a major shift with the deep learning era where massively more resources are being put towards AI year-on-year than before. The rate of progress changed.

There’s no reason this can’t happen again. Indeed, I think we should expect it to happen again for at least two different reasons.

First, at some point, when AI R&D really kicks into gear, we could discover learning algorithms that work much better than today’s. I think the current AI paradigm leaves much to be desired, with major improvements, e.g. in long-term memory and efficiency. And those could arrive suddenly, and take an AI system from “really useful, but still needs a lot of hand-holding” to “we’re not sure we can stop this thing, maybe we should, um… shut off all the computers?”

Second, at some point, AI agents could really take off (we may be in the beginning of this, already), and get very good at effectively and efficiently causing things to happen in the physical realm, and could then start to rapidly and autonomously scale up the amount of physical resources (e.g. energy) directed by AI towards accelerating both AI R&D, and this process of acquiring resources and influence.

We don’t know what sort of behaviors/capabilities are dangerous.

Another dangerous trend is an increasing focus on capabilities that are obviously dangerous, such as bioweapons or cyber-attacks, to the exclusion of unknown risks.
*This RAND report is an exemplar. I previously wrote a detailed response, maybe I’ll post it soon.

The “unknown risks” argument is “When you play against a much better chess player, you know they will win, but you don’t know how”. The things you see coming, they also see coming. They do something else.

We should be worried about any system that is very smart posing a risk to us. Sometimes we can make a fairly strong case that a system lacks a particular capability, and that this makes it safe. For instance, an AI system that has only been trained to play games of Chess or Go is probably going to be safe, even if it’s an insanely good player.1

Arguments that might seem stronger than they are include:

• It’s stuck in a computer, we can just unplug the computer.

• Its memory is wiped after every interaction, it would struggle to make and execute coherent long-term plans.

The problem with both arguments is that they assume that the AI cannot use its influence over the external world to acquire new capabilities. For instance, a smart AI that notices it is limited by such things could pay people to help give it a robot body or a better external memory, or trick them into it.

For those familiar with it, Pickle Rick is a nice fictional example of an intelligent system using external resources to overcome its initial limitations.

In general, it’s hard to know what to make of a system that is clearly really smart, and not fully understood. A lot of experts (Yann Lecun, Gary Marcus, …) claim that current approaches to AI are fundamentally limited, but this is just them stating their opinion, which many other experts disagree with. The reality is we just don’t know.

But even if the system is fundamentally limited in some way, it could still cause massive risks. For instance, lacking a sense of smell probably wouldn’t stop an otherwise super intelligent AI from taking over the world if it wanted to.

It takes time to slow down. The train doesn’t stop when we slam on the brakes.

What needs to happen, once “we” decide to stop? A rough list I have in my head is:

• The US government decides to stop AI, and starts trying to broker an agreement with China and maybe a few other key players.

• The US, China, etc. reach an agreement on how to stop AI.

• The rest of the world gets on board with this agreement.

I expect these steps to take time, quite likely a lot of time. How do you actually stop AI? I have an answer, but there are still a lot of details to be worked out, and I don’t think we’ll really know the answer to this question until world powers actually start prioritizing this issue and are willing to make major sacrifices and compromises to achieve it.

A unilateral pause in the US could be implemented faster (but would still require navigating the politics of the thing, which could take arbitrarily long), and to be fair, I think this is what many people imagine a “pause” looking like: frontier AI companies suddenly cease their R&D operations; they send their researchers off on vacation, and stop their big training runs. And the US is ahead right now, so China wouldn’t immediately race ahead. How quickly might they catch up? Three considerations are: 1) How hard are they racing? 2) How far behind are they? 3) How reliant are they on copying US companies to make progress?

The problem with a unilateral pause is that it expires. You get a few months -- or a few years, if you’re lucky -- to figure things out, and then we’re off to the races again. But we can’t count on figuring things out in that amount of time! We don’t even know what we need to figure out. “Solving alignment” (as popularly conceived) may not be enough.

It’s getting harder to stop. The brakes are fraying.

It’s getting harder and harder every day to stop. Every day we wait, AI companies get richer. AI gets more embedded into society and infrastructure. AI gets smarter. AI research advances. More people get AI psychosis. More AI computer chips are built. More datacenters.

There’s a sense in which AI is already out-of-control.2 The very people building it have repeatedly expressed concern, fear, anxiety, dread, apprehension, etc. about the risks it brings. They say they would like to slow down, if they could. Elon Musk says he has AI nightmares. They don’t seem to feel like they’re in control. As easy as it might seem for AI CEOs to say “oh, damn, this model really is dangerous, we’d better pause”, it’s not clear they will be able to at the critical moment. Maybe a dangerous AI has already escaped human control. But more generally, I am concerned that we will increasingly lack a “nexus” of control at all.

“Racing to the cliff” is not a good strategy.

Despite the risks, a lot of people I talk to in the AI Safety community think that we should keep building more powerful AI until it’s more clear that it’s getting too dangerous. One argument is: Until AI literally kills everyone, it’s basically just great; let’s keep getting all the benefits for as long as we can. This is sort of mostly a vibes-y thing. These people are “pro-technology” and really, really don’t want to be mistaken for Luddites.

The more significant argument is this: the whole point of pausing is to do more safety/alignment research, and we can make the most use of wall-clock time when we have more advanced models to study, and to use to help us with the research. This is clearly a 12-D chess move. In order for it to work, we’d need to know where the precipice is, and we’d need to slam the breaks before we get there, and we’d need to make sure that nobody cuts the breaks in the meantime. I’m not optimistic about any of those things working out, but the plan requires all three of them to. I say: slam the breaks now, while we still can, and we just have to hope that we can stop soon enough.

Thanks for reading The Real AI! Subscribe for free to receive new posts and support my work.

Share

1

Even such a limited system might in principle be able to discover things about the world outside the game and want to gain influence over it -- this would be a bit like the plot of “The Matrix”, but in reverse. This is an area where we do have some uncertainty, but where I’m comfortable saying “I don’t think we need to worry about that yet”.

2

See number five of Ten different ways of thinking about Gradual Disempowerment.

Discuss

Epistemic Status: Abstract claims, but grounded in data science... though the data science is somewhat stale. I wrote this on March 5th of 2026 based on memories of work I did and methods applied circa 2017, and pushed it out for publication after realizing that maybe there is appetite for it, after I saw this post.

Suppose we took a snapshot of each person in the US, and made a list of their "skills", as one might do with a D&D character.

I would like to report on what I expect would happen if this was attempted in real life, and why (until I get to the point that you understand the title of the essay about "Budgeting" skills being Important a little ways in).

I haven't done this recently, with modern data, but I felt that this was likely to be something in my brain that most people don't know about, and worth an article.

At the end of the essay there will be a call to action! I want to start at least one study group in the SF bay area to "Level Up Budgeting" so I could attend somewhere face to face and talk about books or essays or tools, and I'd be happy if various meetup communities around the world formed their own local study groups, so there can be cross-pollination and transient diversity and so on.

Object Level Skill DiscussionFirst, The Boring Skills

With a skill list for every person in the US, we would see a lot of lists for people whose chief skill is their ability to use a QWERTY keyboard and use a suite of Office Software to create spreadsheets, presentations, and prose.

It turns out you would see a lot of people who have a CPR certification, and almost everyone who gets that in modern times gets an AED certification as well.

(When I think of this, it reminds me of how rationalist!Harry went straight for a magical first aid kit on his first day of being able to buy magical gear. But also, it is interesting how "skills" blur very very swiftly into "certification of those skills" in the actual speech of actual people.)

In real life, writing down a skill list for every person would probably lead to variation in how people write.

Some people would write "Skills: word, cpr, aed, ..." and others might write "Skills & Certifications: MSOffice, AED, CPR, ..." and in some platonically perfect realm these would parse down and be applied to generate the same basic inference of the same basic capacities in the author of the list.

Languages already lump. To really dig into a skill, you don't just know "the name of the nameable skill" but actually can DO things.

There are probably subskills!

Some of the subskills might not even have names but still be something that can be transmitted and learned by someone saying "hold your hand <like this>... see?" and then "no, your pinky is getting in the way... more <like this>". We are interested in all of it, of course, from the microskills to the macro level.

Second, The List Of Skills

Most people in real life can't cast a "healing spell" (because magic isn't real, so far as I know).

And "krav maga" is pretty rare, and not what people think of first.

Fighting is not central to real life in the modern world... mostly (cops and soldiers and bouncers and bail bondsmen do make up a non-trivial chunk of the jobs though).

Usually people think of the civilian job market, in "real life", when they actually make lists of skills and post them online.

Even in the military, they have complex software helping with the task of killing people effectively to serve diplomatic ends (or whatever), but in practice they end up needing to model and optimize the logistics of hiring 18 year olds and turning just the right number of them into "fuel-truck drivers-license test proctors" as fast as possible. In a deep sense: a lot of skills are pretty prosaic.

The "reality" that is "out there" is nebulous, of course.

Depending on how much lumping and splitting you do, so far as I could tell back when I had access to the data, there are roughly 50k to 200k total skills that people will think are distinct enough to write out as distinct phrases.

Some of the phrases will point to having a specific security clearance level, but often someone might just write "Abilities: office, cpr/aed, ... security clearance, ..." as if those two words "security clearance" were a skill, or pointed to skills, or something?

(Metonymy is the name for what happens when people hit the limits of their words, and just name a thing for whatever thing happens to be nearby that they already know the name of.)

One could probably get the number of skills down around 35k if you put serious effect into de-duping the phrases that people use, and ignore hapax legomena.

But if you only de-dedupe the top 100 most common skills by hand based on semantics, and de-dupe the skill phrases that show up in skill lists algorithmically (reducing plural and single forms into one form), and keep the skill phrases that only ONE PERSON EVER thought deserved to be in a skill list, and yet also hunted far and wide to gather all the English language resumes you could (from the 1980s to 2026) I bet you could get as many as 300k skills, without much trouble.

And if you took each one of these and then traced it to the person, and asked them to teach someone this "officially listed skill" I bet it would turn out to often have 10 subskills that can only be described with a phrase or sentence or paragraph... which means that there are plausibly ~3M skills that are super granular and would take a "paragraph of description" to point to, potentially? Things like "the best way to hold a nail in your other hand when hammering nails".

When I estimate ~3M such skills, I'm being pretty sketch and rough... I would be surprised if it was less than 500,000 of them, and I would also be surprised if it was more than 10 million.

Third, The Serious Lumping Begins

A very reasonable person might think that this is crazy.

They might think "computers" is really "just one skill" and that "git" and "svn" are NOT meaningfully different skills.

They might assert that there is ONE "computer" skill, and that if you are super great at that good at the one skill (but better than a normie!) you could apply that skill in a tech support job, but then you might level that skill up until you could apply it to work in Computer Research.

This would be a very reasonable and pragmatic perspective, but in that case, there's a lot of prior classification work!

The reasonable people at the Bureau of Labor Statistics have already lumped things down to 832 narrow job categories fitting into 116 medial grained categories and just 21 basic categories.

Following them, seeking VERY lumpy lumps, you might reasonably say that the total number of "skills"...

...is either ~21 (some examples being "legal", "protection", "production", "management", "sales", and so on)...

...or else perhaps ~116 (some examples being "legal support skills", "fire fighting skills", "printing skills", "operational management skills", or "wholesale sales skills").

The BLS... is almost certainly too narrow.

Those aren't really skills, you know?

Those are really more like job categories... right?

But it does seem to be true that job categories sort of wink and nudge at the sense in which some skills might be more convergently useful than other skills, or have prerequisite skills, and be useful mostly only in concert with other skills.

Most people have played the piano at least a little, when killing time and kinda bored while waiting somewhere that a piano existed, and quite a few people have practiced specific piano skills (and there are many distinct such skills to practice separately (like just sitting properly, even)) but if we bow down to the lumpiest lumpers and their traditions, even using the full 832 BLS categories, all(!) of the piano skills (and so much more) are lumped into "27-2042  Musicians and Singers".

This is NOT GOOD ENOUGH to tell us what skills to spend 10 minutes practicing every day, or how to hire people who will pay the piano in the way that we really want to hear it played in a bar from just what can be found about them via online scraping of their data.

If we are going to HIRE for jobs, based on a gears level understanding of what cognitive or physical performance or capacity goes on in that job then we need something more granular than jobs to point at, or reason about, that "are the gears".

Fourth, Lumping The Skills With Math?!?

Here's a thought: maybe we take all the skill lists as is to create ~100k skills from some corpus (with regional choices, and choice of era, and some amount of de-duping) that someone ever mentioned.

We treat each unique skill skill as a node in a network (or as mathematicians might call it a "graph theoretic graph")... and then for each skill list in each resume, we draw lines between every pair of skills that occur in the the same list.

And then for the next list we do the same (making some lines +1 in strength if they were paired in a previous list, so AED and CPR are decently likely to have a three-weight-strong link after processing maybe a dozen random resumes).

This gives us a weighted graph, which comes up in a LOT of optimization problems like shortest path discovery and pagerank and so on.

I have done this before! That's why I am writing this essay ;-)

Lots of "clusters" of skills fall out of such an analysis in the form of dense cliques where many members of the clique are strongly linked to other members of the clique, and weakly linked to anything else. These cliques represent tasks that are demanded in the same job, or abilities taught in very common forms of general education, or sometimes tasks that show up in "standard career progresses" (where a drafting student becomes an engineering tech and eventually a licensed surveyor, and their skill list has things from that whole history).

Every so often you'll find skills that are very common, and that many cliques of skills ALL point to.

For example, there's a bunch of skills that teachers, in the field of education, are proud of, where, by mid career, almost everyone is bragging on their resume that they can do "curriculum design" (no matter whether they're a math teacher or a piano teacher or a kindergarten teacher).

There are other skills like "security clearance" that might show up in the protection area..

But then "cpr" is linked to both of these (and many others)!

This strongly suggests a high degree of convergent instrumental utility exists in skills like this, across a wide variety of fields, even though the skill is "narrowly a skill" that is more like "git reflogging" that can be taught and practiced and tested, but also "broadly a skill" in the sense that it comes up for practically everyone.

Fifth, Seeking "Betweenness" Centrality You'll Find "Budgeting"

There are many measures of "centrality" in graph theory.

One way to arrange the math of it might pinch out the thing at the very center of the very biggest clique (which itself is perhaps at the center of the biggest macro-clique and so on)... but that won't give us these insanely broadly applicable skills that can be taught and learned!

The thing we want, if we're looking for very very broadly valuable skills in almost any domain or any job is betweenness centrality.

The way this works (roughly) is that we pick a lot of pairs of nodes, and spend compute to find the shortest path between them... over and over... and every time we do this we add a point to all the nodes that were on this "shortest journey" from node to node.

At the end, we find the chokepoint... the master node, the node from which you can go ALMOST ANYWHERE very very quickly.

"Curriculum design" is more central than "math curriculum" because it is invoked by more kinds of people with more diversity of skills. And "CPR" has still more betweenness centrality than "curriculum design" (because CPR is useful for cops with cop skills, and life guards with swim instructor skills, and fire fighters, but also elementary school teachers, and summer camp instructors and so on).

Here is the punchline: the thing with the most betweenneess centrality out of all skills is "budgeting".

Which... uh... you know... makes sense? Maybe? <3

Sixth, Meditations On The Betweenness Centrality Of "Budgeting"

Here are some thoughts:

FIRST, consider the holy grail of "rationality" (in the sense of "verbally transmissible cognitive practices that conduce to higher chances of success at nearly any goal" is skill transfer).

Finding skills with high skill transfer and broad applicability lets you spend the least amount of time leveling up, and gives you the most benefit from them.

The skill of "budgeting" makes sense here because like... finitude is everywhere? Tradeoffs are everywhere? Also time is real and ubiquitous. And "making tradeoffs over time in the face of scarcity" is basically the essence of budgeting.

"Budgeting" arguably deserves some halo, because it makes sense, from first principles, that almost every agent would need this skill, if you actually think about agency itself in a first principles way.

SECOND, this comes up in practice in business because the "mangle" of organizations often leads to a handful of people on like..  "the budgeting committee" (which no one wants to be on because it SOUNDS SO BORING) wielding enormous organizational power, and needing to be able to justify their use of power when budgets cause weekly, monthly, quarterly, yearly adjustments in what the organization can and will choose to do.

This will show up, predictably, at the HEART of many moral mazes. We would kind of expect, then, like with the "security clearance" skill, that "budgeting" is a pointer to skills related to the use of language in clarifying AND politically obfuscating ways, depending on who a budgeter is talking to.

The skill might be anti-inductively hard to master because it is woven into bizlore (not academia) and it is very political by default.

THIRD, the lists of skills from which this kind of conclusion can be drawn often come from a wide variety of people, posting resumes online and such. Some of them are older, and more advanced in their careers.

There is a decent chance that if you start in almost any profession or organization, and are good at anything, eventually you'll rise to a role of teaching, guidance, planning, and in a word "leadership". And almost all leaders of non-trivial organizations have accounting and bookkeeping to handle... and also this idea that "the procedures of accounting" could be used to map and understand and control and generally guide a team?

So maybe the betweenness centrality of budgeting comes from a real pinch point, that shows up over and over again in MANY careers, where "budgeting" (about your expertise, on behalf of a group that collectively wields that expertise) becomes essential.

FOURTH, we can compare Biden's semi-famous quote "Don’t tell me what you value. Show me your budget—and I’ll tell you what you value" with a lot of common topics on LW like money itself (the unit of caring!) and kelly betting (essentially iterated bet budgeting), and the relation of money to VNM rational utility functions and bayesian beliefs.

If the skill of "budgeting" turned out to be a super power wouldn't actually be that surprising, if you have a sense that more is possible.

A Sort Of A Methods Section

This essay so far has been very high level. I haven't really substantiated it, except to casually suggest what MIGHT be observed, if someone hypothetically looked at certain ways of describing the world in an intensely operational way. Maybe you just don't think it makes sense?

I generally don't like talking about myself, but... uh...

I applied to work at Google in 2013 when I believed that they had the Mandate Of Heaven with respect to the Singularity.

(For several reasons. One big one was that I had previously thought The Singularity Institute had the Mandate Of Heaven earlier than that, but then they sold their banner and trademark and social outreach apparatus to Kurzweil in 2012, who worked at Google. This was before OpenAI was founded in 2014. If I had waited another year, I might have joined OpenAI, instead.)

By 2018 I had given up on Google, because it is a moral maze, and was clearly not going to save the world, and I moved on to working on a blockchain project with friends that was not much related to the Singularity or to AGI in any obvious ways, but it gave me an opportunity to design decentralized utopias and dystopias for decent pay, and that was just super super hard to turn down.

In the intervening time, from 2013 to 2017, I first worked inside the Google Brain "cultural silo" (because I had been told I could work on deep learning aimed at comprehensive proactive benevolence) but after about a year of working on bullshit (like trying to get people to watch longer YouTube content that ads could fit in the middle of) I changed to work on something that actually fucking matters.

So I hopped over into the "PeopleOps" silo by 2015, and instead of increasing the size of YouTube's "ad inventory" I helped build the job search engine that is one of many topic-focused engines existing inside the overall Google Machine for that overall technological macro object to wield according to high level metrics.

(Job search seemed like the closest thing I would be able to find to a sword of good, that The Google Machine would have a hard time abusing too egregiously. And it seemed like it could plausibly produce tens of thousands of dollars in consumer surplus for people who used "free job search" (supported merely by ad money, that is) to find better jobs.)

For about 18 months, between 2015-2017, I was doing data science on the theory and practice of searching for jobs specifically through the lens of "skills and education".

I didn't have the ability to contact people in the outside world about it (because we didn't want to leak that we were working on job search and so on) but my team had access to "all the resumes" and "all the job postings"... in English anyway (and it turns out that the language used in different job markets around the US was regionally dialectical, and so we optimized at the beginning for just a handful of cities, and when we tried to use machine translation to generalize it to Telugu or Japanese the relevance numbers just totally cratered because there is a lot of nuance in the euphemisms people use to talk about hiring and firing and such).

We could play with ways to parse and analyze them using now-old-fashioned pre-LLM NLP techniques. So for a bit more than a year, I could, in fact, do the kind of analyses I described above, but with real data.

And "budgeting" did, in fact, fall out of that actual data, back in ~2017, in the way described <3

A Call To Action: Study Groups

Does anyone else want to do this?

I hope that people in the comments chime in in lots of ways because I still feel like "budgeting" is something I'm STILL learning to do well. I would love to hear good textbooks. I would love to hear war stories. I would love to hear ways to spend an hour a day practicing something for a week and become really good at "budgeting".

Does anyone know how to hire the best CFO who is a total wizard at the "budgeting" skill? I don't! I think that's an important thing to be able to do, and I can't reliably do it.

I would be open to driving to Berkeley once a week to talk about some shared reading, or report on homework we assigned ourselves, and hear about other people's challenges and growth at "budgeting, the most central of skills" <3

Discuss

I've been working on a mini-book on RLVR for the past few weekends, sharing the v0 now: https://rlvrbook.com

Please check it out!

Discuss

Introduction: Solving for Sleep

I have had some moderate sleep issues from Autumn 2025 onwards. But in April 2026 as I'm spending the month living at the Lighthaven campus in Berkeley, CA, my sleep has improved significantly despite theoretically worse sleeping conditions. This post is my attempt to break down what my sleep issues are/were, what my sleep conditions were like before and after, what changed, and what actionables I can take next.

I'm at the Inkhaven Writers' Residency at Lighthaven, so I figured at the very least this would be sufficiently ratty LessWrong content.

The symptoms of my sleep issues can be best described as follows:

1. I tend to wake up 1-2 times per night most nights. Sometimes this is me waking up to urinate (nocturia), but lately it's just been random wakeups for no reason. I can usually get back to sleep, but I think these wakeups have disrupted REM, which is very important to avoid.
2. I find it very difficult to sleep in past 7:30 AM or so. This isn't the worst thing in the world, but even when I stay up late my body seems to want to wake up between 6:30 and 7:30. Turning off my alarm is useless. I'm not able to make up lost sleep by sleeping in on the weekends.[1]
3. I have a hard time falling asleep at night more often than not. I hadn't had this issue in many years but it resurged in late 2025.

Having spent many nights thinking about it, I think the causes of my sleep issues are as follows:

1. I do not practice good sleep hygiene. My office and my bedroom are the same room. Though I resolved in 2026 to quit, I still sometimes find myself on Twitter or YouTube in bed in the middle of the day.
2. I do not exercise enough. In the summer I get a fair amount of cardio (in summer 2025 it was unusual for me to take less than 13,000 steps per day).
3. I live in close quarters with a roommate. My apartment is ~625 square feet and my roommate's waking hours tend to vary. I think this is more psychological than anything (we will return to this).
4. I drink alcohol majority of evenings. Moderate alcohol consumption is bad for sleep, though I don't think this is as big of a detriment to my sleep as I thought (we will return to this).
5. I stress too much about not sleeping. The fact that I worry about how little sleep I get is probably doing more harm than good.

I have suspected these were causes for some time, but I haven't been able to test them until this month when my lodging situation changed radically (and my sleep with it).

How I slept before

My bedroom in Toronto is a fairly small space containing a double bed, a large desk, a dresser, and a bookshelf (not pictured):

Not a bad setup. But this pano from 2021 makes it look larger than it actually is. The bed is wedged against the corner (as are the beds of all single men in their 20s), and there is nothing demarcating the sleeping zone, the work zone, and the play/leisure zone. For me, places often mix.

My wakeup routine is highly varied. Sometimes I lie in bed and do nothing for an hour or longer, sometimes I get up immediately. I shower in the morning like 2/3rds of days but sometimes I will shower in the afternoon or even the evening (I work from home so I have some flexibility there).

"Going to bed" for me in Toronto looks like wandering home or switching off my computer, getting changed, and climbing into bed anywhere between 10:00 PM and 1:00 AM. If my morning ritual is weak, my evening ritual is nonexistent.

How I sleep currently

Barring the Bodega Bay Beach Episode of Inkhaven (which our Residents and Team are all enjoying and taking very well), I sleep in this shared dormitory in a guest house a few blocks offsite of the Lighthaven Campus in fabulous Berkeley:

My bed is the bottom bunk on the left hand side.

I currently have a very strict morning routine at Inkhaven. I aim to arrive at Lighthaven between 7:30 and 8:00, which means I need to awake around 7:00, shower, dress (I usually lay my clothes out the night before), and slip out within a small amount of time.

I tend to go to bed between 11:30 PM and 12:45 AM, depending on how silly I got in the Winner's Lounge the evening before. I am often first into bed in the dorm (some of these animals stay awake till 3 or 4 writing about AI safety). I never hear them come in.

I always sleep all the way through the night. I only medicated myself on the evening of 1 April, my second night there, but since then I've been able to get to sleep as soon as I climbed into bed. I usually sleep for 6.0 to 7.5 hours. Maybe 25% of days at Inkhaven I've felt sleepy and woozy during the early afternoon, but never anywhere severe as I've felt on my worst insomnia days in Toronto.

What changed exactly?

• I exercise more, at least more than I did in Toronto in Fall/Winter. This is almost certainly the strongest effect. I take many steps per day, both milling about Lighthaven and in my explorations of Berkeley and Oakland and San Francisco. The Bay is very hilly compared to glacier-flattened Toronto, so 10,000 daily steps is a lot more demanding upon me in Berkeley than the Annex.
• My work, play, and sleep are very separated. I work exclusively at Lighthaven. I hang out/scroll Twitter exclusively at Lighthaven. If I want to lie down for a quick catnap (which I've never done), I'd do it at Lighthaven. I never go on my phone/laptop in bed, not even in the mornings/nights. This is very good sleep hygiene, the best I've ever practiced.
• Going to bed is high friction. There are so many lovely people at this Residency, and I want to stay up talking with them late at night. Going to bed requires me to pack up my briefcase and walk 4 minutes offsite. That's enough friction that I'm not going to head to bed until I'm so sleepy that I have no choice.
• I'm not stressed about sleep/mornings. Not only do I not lie awake thinking about "oh no, I'm gonna have an insomniatic night", I'm also not stressed about fighting over the shower/kitchen in the morning. In my dormitory I'm always first awake and first to the shower; through all of April I've never had to wait to bathe or make coffee or get breakfast.
• I am sleeping somewhere different. Novelty seems to matter to my brain when I need to sleep. Even when I am tired, I no longer associate my bed with sleep, so I don't fall asleep easily in bed. But if I'm tired on a couch or a subway or a zimzilbang spa, the drowsies kick in almost immediately.

What didn't change (and didn't seem to matter)?

• I still use screens before bed. Except for the 10 minutes it takes me to walk to my accomodations, get changed, and get into bed, I do not do any kind of screen detox before going to sleep. My sister once did a nightly one hour screen cleanse before bed. Maybe that worked for her, but I was always suspicious of it. Indeed it seems like I can stay up late writing or watching a movie and still get a decent sleep the next day.
• I still drink alcohol at night. I've never gotten roaringly drunk at Lighthaven, but I'll have some beer or wine most nights after a hard day's work. Sometimes it's in a social context, sometimes it's as I sit in Rat Park performatively reading poetry or blogposts. I've been drinking a lot less in the last 4 months or so— Dry January in particular seemed to curb my nocturia— but it's interesting that my habits haven't changed much at Inkhaven yet I feel like my sleep is much better.
• I still naturally wake up around 7:00 AM. This isn't a short-term frustration of mine (I'm glad I get up early because I want to spend more time at Lighthaven!) but it's weird that I still have literally zero pressure to sleep in. I have no idea why this is.

What actionables can I take?

I am more convinced than ever that sleep hygiene is my current biggest life impediment. I am more productive, more pleasant, and I feel better when I sleep well.

Mentally I feel really stuck in my tiny Toronto bedroom. I feel as though there is no solution; as though when I return home I will fall back into my terrible sleep habits. But this is an emotional response, not a rational one. I think there are some concrete actionables I can take.

• Exercise more. There's nothing like May in Toronto. Now that I've had my boot soles serviced and I've figured out how to treat blisters (the trick is to put a bandage on the blister), I have zero excuse not to hit 10, 15, or 20 thousand daily steps for the next 5-6 months. I'd do well to start strength training again after a 4.5 year hiatus. I've also learned that I really like boxing, so maybe I should look into that too.
• Work more outside of the house. My current work does require me to use my big fancy PC at home sometimes (and answer video calls at a moment's notice sometimes, which is really irritating) but a lot of my work can be done from libraries and cafes. (Honestly getting a new job would probably help with this in a number of ways, which is something I was already considering.)
• Recreate more outside of the house. Again, Toronto in Spring and Summer is wonderful. I gotta get out of the house more and create more barriers between unwinding after work and going to sleep. I should watch more movies or go to tboy oil wrestling shows or something.
• Create physical dividers between my bed and the rest of my life. My living room and kitchen area is shared with one roommate, but I don't think he would have an issue with me doing more of my work on the dining room. I should also watch YouTube and films on our TV rather than my computer monitor, but I have a weird neurosis about other people seeing what videos/films I am watching, and I much prefer to keep it private. That's a problem for another post. Also, maybe I should get some kind of divider to separate my bed from the office side of my bedroom.
• Travel more? Maybe the sea air of California has been disproportionately helpful to me in a way I've not considered. My brain seems to reward novelty above all else, but that's also a problem for another post.

Maintenance of good sleep is important to Not Being Miserable. Not Being Miserable is a cornerstone of my long-term survival strategy, and I ought to investigate any small adjustments I can make to decrease my misery. The ROI is probably going to be worth it.

1. ^

   My mom has a worse version of this problem where it's impossible for her to stay asleep longer than 5:30 or 6:00 or so. I hope it never gets this bad.

Discuss