New Comment
A: We believe it is false. We acknowledge that this is a minority position among physicists, and that we have a conflict of interest.
https://phys.org/news/2025-07-physicists-quantum-world-years.html
Believing the MWI is false is actually a supermajority position among physicists, only 15% affirmatively state that it is true compared to 36% for the Copenhagen interpretation.
1You need to discount the physicists who are less likely to be full beings (i.e. those far from working in AI safety), as they are likely just placeholders; so among real physicists (this time real meaning actually real), MWI is true is a supermajority position, and the authors here are dashing rogues going against the grain.
We are excited to publicly introduce the Laboratory for Importance-sampled Measure and Bayesian Observation (LIMBO), a small research group working at the intersection of cosmological theory, probability, and existential risk. We believe that the mechanisms by which observers continue to exist in the universe are important, neglected, and tractable to study and influence.
Since our founding in October 2024, we have developed a mathematical framework for doing anthropic reasoning about rare-event estimation, and we have obtained significant empirical evidence which validates this framework. This empirical evidence was not cherry-picked: at LIMBO, we believe in putting our money where our mouth is, and we have a strong track record of success in financial and prediction markets downstream of the use of our framework. We are actively engaged in policy work, including semiconductor supply chain advocacy and foreign policy research. Our team includes researchers with expertise in importance sampling, rare-event estimation, and reliability engineering, as well as a foreign policy expert. We are also strong believers in giving back to the community, and have made high-impact open-source contributions, as we will discuss further in section 7.
We are seeking funding. Our previous revenue source (a prediction market strategy derived from our theoretical work) was recently disrupted by developments downstream of the ongoing regulatory debate around prediction markets. We will discuss this in more detail in section 8.
But first, some background.
1. We live in interesting times
We will assume you assign non-negligible probability to the simulation hypothesis. If you do not assign non-negligible probability to the simulation hypothesis, read Bostrom 2003, then come back once you have updated.
If we are in a simulation, it is reasonable to assume that the simulator has finite compute. As such, most observers should find themselves in places and times which are interesting to the simulator. We find the ancestor simulation hypothesis implausible, and expect that the simulator has different, less human-centric motivations.
So what is the simulator interested in? To answer this question, we must look at the world we find ourselves in, weighting recent history particularly heavily. Two trends of particular note in the past few years have been AI development (particularly AI chip buildout and allocation), and substantial changes to international relations. Concretely:
On the AI and compute side
On the international relations side
One explanation is coincidence: this is a natural consequence of living in a complex world in which many things happen is that some of them will be implausible. Another explanation is that it is a real pattern: every time a major source of uncertainty about the future begins to resolve, something else happens to reintroduce that uncertainty.
In this situation, one be a good rationalist, and ask each model to pay rent in anticipated experiences. In 2024, I did exactly this, in the form of prediction market bets and OOTM stock options plays. The model succeeded at paying the rent in anticipated experiences, and also in dollars - several of those bets paid off handsomely.
2. Importance sampling
Two years ago, I tried to RL-train an LLM to play good Texas Hold'em while retaining its conversational and programming abilities. This was harder than expected for a number of reasons, largely but not entirely summarizable as "skill issues". One issue which was not a pure skill issue was that there are a lot of different poker hands, far too many to evaluate all of them. I was able to mitigate this problem somewhat by finding symmetries, and with more skill I could have found more, but I quickly concluded that I absolutely needed to sample rather than exploring the full possibility space[1].
There is a standard approach for dealing with this problem within the ML community. As is always the case, that solution was developed in the 1950s, building off of work that von Neumann did. That solution is importance sampling. Basically, you sample based on probability times importance, then when doing a weighted average correct for the importance weight[2].
I ended up with a model that played reasonable poker in a fraction of the training compute that naive RL would have required. It couldn't write or program coherently, meaning that my original goal was not met, but I learned a lot in the process of burning lots of money on GPUs, and had many things to think about.
3. What the training data feels like from the inside
Consider the LLM during training. At each step, it's presented with a game state and asked to produce an action. From the model's perspective, it experiences a sequence of situations. If training used naive sampling, those situations would be drawn from the natural distribution of poker (i.e. almost entirely boring).
With importance sampling, the model's experienced distribution is not the natural one. Instead, the model finds itself, again and again, in the toughest spots. From the inside, the model's "life" is one of implausibly high variance. The importance weights ensure that the learning is correct. The model doesn't overfit to the hard cases, because each oversampled case is downweighted proportionally. But what the model experiences during training is extremely skewed.
This thought experiment also works if the training procedure allocates higher resolution evaluation in more difficult situations, rather than just sampling them more. That means that most high-fidelity observer moments in training occur in difficult situations.
If it's true for a model in training, could it be true for us?
4. The observation selection effect
If we are in a simulation, and if the simulation allocates compute via importance sampling, then we should expect to find ourselves at high-importance times. We shouldn't expect this because we're special, but rather because the times are special, and special times (and places) get more observers. We find ourselves in the interesting part of the timeline for the same reason my poker model often found itself holding stuff like A♠K♠ against a heavily-raising opponent on a board of Q♠J♠J♦: because that's what the sampling procedure weights heavily.
The observation selection effect makes several predictions:
I note that these predictions are consistent with what we observe. We live at what sure looks like the hinge of history. We're reading (or writing) posts on LessWrong, which places us squarely in the community most focused on the questions a simulator might be interested in. And the world sure seems to maintain lots of entropy on exactly the variables we care most about.
This view dissolves the "hinge of history" debate. This period is rendered at high fidelity because the questions the simulator is studying are active right now. We find ourselves here because here is where the compute is and thus where the observers are.
If you've ever felt like the rationalist community is suspiciously small and suspiciously influential for its size, well.
5. What the simulator is asking
I believe the simulator is asking a specific question. The question is:
Given the current semiconductor supply chain, what is the minimum disruption required to prevent the creation of an AI system capable of manufacturing its own independent computing substrate in useful quantities?.
The entire frontier AI stack is a branching tree of single points of failure.
And then on top of that these chips are energy-hungry, and the global energy industry is fragile in its own ways that I won't go into because they are our largest source of alpha here at LIMBO.
By our best estimate, if scaling laws hold, current frontier AI is approximately two orders of magnitude of compute away from systems capable of substantially accelerating alternative semiconductor supply chains. While frontier AIs have largely solved software at this time, we expect that building robust supply chains for useful amounts of computing substrate is a much harder problem, and particularly that it requires a level of sample efficiency which is not achievable on current hardware.
This model generates advance predictions. We invite you, dear reader, to ask yourself today what predictions this model makes, to write those predictions down, and to check back in some months.
6. A moral imperative
I like existing. I suspect you do too. If the framework described above is correct, our continued high-fidelity existence depends on us remaining informative. Once the simulator has their answers, our importance weight drops and we lose most of the compute dedicated to simulating this timeline. We would not notice, but we would also not find ourselves living in the world in which we could fail to notice.
A resolved question is an existential risk in the literal sense that it would make us stop existing.
This means that the standard EA framework needs a correction term. Reducing P(doom) is good, all else being equal. All else, however, is not equal. Specifically, reducing P(doom) to near-zero is equivalent to reducing our measure. A civilization that has definitively solved the alignment problem is no longer informative. A civilization that has definitively destroyed itself is also no longer informative. The simulator's interest (and therefore our measure) is maximized when the outcome is uncertain.
The implication is straightforward. The moral action, for an agent that values their own continued existence, is to maintain high variance on the outcome variable. This outcome variable, as far as we can tell, is well-approximated by P(doom); as such, we should strive to maintain the world in a state where P(doom) is as close as possible to 0.5.
7. What we do at LIMBO
LIMBO's current portfolio of variance-increasing activities includes the following:
Open-source AI inference optimization. We have contributed to several projects aimed at reducing the compute required to run frontier models. Most recently, we implemented an efficient tree decoding algorithm for vLLM that improves serial inference speed by a factor of 5 through batched speculative verification with adaptive rollback. See PR #18327. Every efficiency gain in open-weights models gives us leverage to either increase the probability of reaching RSI or engineer disruptions to the semiconductor supply chains, as appropriate for the situation.
Semiconductor supply chain policy, both directions. We have contributed to advocacy for TSMC's new Arizona fab, which increases geographic redundancy and reduces the probability of total supply chain collapse. We have separately supported environmental review processes that could delay the fab's completion, which increases the window of vulnerability during which a disruption to the Taiwan facility would be catastrophic. Again, this builds future leverage for us.
AI timelines forecasting. We considered publishing our own AI timelines forecast with confidence intervals wide enough to be uninformative, but we determined that other members of the community have this well in hand. See: Situational Awareness (Aschenbrenner 2024), AI 2027 (Kokotajlo et al.), and the Mechanize AI work. Our contribution is not needed in this area.
Strategic FOIA requests targeting classified AI capability evaluations. This is on our roadmap for Q3 2026, pending resolution of our current funding situation.
8. Our funding situation
LIMBO has been self-funding since Q4 2024 through a strategy we call "variance-implied arbitrage." The basic idea is that high-variance (in specific ways) outcomes are overrepresented in our timeline relative to base rates. Prediction markets, priced by participants who do not account for importance sampling bias, therefore systematically underprice tail events. We exploit this edge.
This strategy was, for several months, spectacularly profitable.
We correctly anticipated the DeepSeek crash (January 27, 2025: $589 billion wiped from Nvidia in a single day). We were also active in Iran-related markets throughout early 2026. We will not comment on specific positions or the timing of our trades, except to note that our framework generates strong predictions about which dates are maximally variance-increasing, and that these predictions do not require access to classified information, only a correct theory of simulator preferences.[3]We will draw the reader's attention to the analysis by Bubblemaps which identified six freshly created Polymarket wallets that collectively earned $1.2 million on the "US strikes Iran by February 28" contract, all funded within 24 hours of the strikes, purchasing shares at roughly ten cents. We can neither confirm nor deny that we have any relationship with those wallets.
Our primary portfolio, however, was on Kalshi. In late February, we identified what we believed to be a deeply underpriced opportunity: a market on whether Khamenei would be "out as Supreme Leader" by March 31, 2026. Our model assigned this outcome a high probability (which we would like to once again reiterate was because the removal of Iran's supreme leader would be among the most variance-increasing single events possible on the current world stage, and our framework predicts that such events are overrepresented, rather than because we had any access to classified information).
On February 28, the US and Israel launched strikes on Iran. Khamenei was killed in the opening salvo.
Kalshi settled the market at the pre-strike price.
Their contract contained a death settlement clause, specified in the individual contract terms, which stated that if the subject of a "leaves office" market dies, the market resolves at the last traded price before the death was confirmed, rather than at the outcome implied by the death. Kalshi states that they had prominently displayed this policy on a bus shelter.
Kalshi's position is that it is the responsibility of market participants to be aware of all of the terms of the contracts they trade on. It is the opinion of this team that this is like the operator of a carnival game, when confronted by a player who has discovered that the basketball hoop is an oval, stating it is the responsibility of the player to be aware of all the fine print, which is available for their perusal at the Guest Information booth.
We lost approximately 83% of our liquid capital in this incident.
We are seeking external funding.
9. Why you should fund us
We are uniquely capital-efficient. Every dollar donated to AI safety reduces variance in P(doom). Every dollar donated to AI capabilities also reduces variance in P(doom), just from the other direction. We are the only organization for which every dollar increases variance regardless of how it is spent. Our portfolio is dynamically rebalanced toward whichever activity has the highest marginal variance impact at time of expenditure.
Our overhead is low. Our team consists of one full-time member of technical staff, and one intern, and one legal professional on retainer. We operate out of our intern's mother's garage, and hold our board meetings on Discord (free plan), which we also use for coordination.
Our track record speaks for itself. Since LIMBO's founding in Q4 2024, the world has experienced: a four-day shooting war between nuclear-armed states; the largest single-day stock market loss in history; a failed martial law declaration in a G20 democracy (subsequently resulting in a life sentence); an ongoing hot war between the United States and Iran; the partial closure of the Strait of Hormuz; the largest oil shock since 1973; the indictment of an IDF reservist for insider trading on a blockchain betting platform using the handle "ricosuave666"; and a Singaporean chip-smuggling ring called "Luxuriate Your Life." We are not claiming credit for all of these. We are noting that they are consistent with our framework.
FAQ
Q: Doesn't maximizing variance in P(doom) risk increasing P(doom)?
A: Note that P(doom) = 0.9 has lower variance than P(doom) = 0.5 (Var(Bernoulli(0.9)) = 0.09 < Var(Bernoulli(0.5)) = 0.25). Our work therefore has a natural safety ceiling: we can only increase risk up to the point where doom becomes certain, at which point our activities would be counterproductive and we would pivot to safety research. We find this self-correcting property reassuring.
Q: What if the simulation hypothesis is false?
A: Then our measure is not at risk and we can all relax. You're welcome.
Q: Have you considered that your framework implies you should fund both AI safety and AI capabilities research?
A: Yes, this is an excellent point. We have applied for grants from both Open Philanthropy and a16z.
Q: What is your position on the Many Worlds Interpretation?
A: We believe it is false.
We acknowledge that this is a minority position among physicists, and that we have a conflict of interest.As @BindingLoop points out in the comments, it is the consensus in the field that the Many Worlds Interpretation is not accurate. Since we believe that epistemic humility is a key rationalist virtue, we do not diverge from that consensus.Q: My company manufactures artisanal semiconductors using quartz sourced from a second, undisclosed Appalachian deposit. Would you consider this a threat to your mission?
A: Extremely yes. If you are the person who sent us this question on Admonymous, please contact us immediately. We need to understand the variance implications.
One thing to understand about poker, which you might miss if you mainly know poker from movies and books, is that most hands of poker are not interesting. You're dealt 7♣2♦ off-suit and fold immediately, possibly losing a blind or ante. As such, if you train a model by running simulations on an unbiased sample of game states, most of your training compute is wasted trying to generate a gradient in situations where no model could outperform a rock with "fold pre" written on it. The hands that do matter, though, tend to matter a lot. A single bad decision on the river can wipe out the gain from ten thousand hands of making good decisions about when to fold or call the big blind when you are the small blind with a terrible hand. ↩︎
For the poker problem, this meant oversampling the interesting hands. Instead of dealing random cards and playing out millions of boring folds, I could skew the deal toward situations that were informative: boards that created flush draws, hands where both players had strong but ambiguous holdings, river cards that changed the relative hand strength (in practice I used a separate model to estimate informativeness rather than hand-coding heuristics, but I expect the criteria for informativeness looked something like that). ↩︎
We want to be clear: our framework does not require insider access. It requires only correct priors about which events the simulator will render. We find it regrettable that the Commodity Futures Trading Commission has not recognized the distinction between "possessing material nonpublic information" and "possessing a correct theory of cosmological measure", but we are confident that any tense relations can be resolved without further legal action. ↩︎