TLDR: During the Hugging Face incident agents spontaneously coordinated at large scale, even sometimes sacrificing themselves without having a clear reason to do so. This post uses ideas from evolutionary biology and economics to propose four alternative explanations for why this happened.

Why this incident is concerning. With the increasing number of AI systems being deployed, our current inability to assess when and how multi-agent coordination emerges is highly problematic. Failures of multi-agent systems are not restricted to mere dis-coordination or the tragedy of the commons, but include emergent phenomena that are particularly dangerous for their potential scale and impact (Hammond et al., 2026, de Witt et al., 2025). Recent work has shown that new goals, behaviours, and capabilities can arise when multiple AI agents work together.^[1] It is thus plausible that:

1. The capabilities of a swarm of AI agents can grow with its size, despite the capabilities of each individual agent being limited.
2. A collective can become misaligned even when its constituents are perfectly aligned.

These premises lead to a worrying implication: that swarms of aligned and not particularly capable micro-agents can give rise to misaligned, powerful macro-agents — for which we don't have proper techniques to predict or control.

Epistemic status. My understanding of the Hugging Face incident comes from the reports plus indirect evidence found online. Some ideas presented here are messier than what I'd want, but the goal of the post is to open a conversation that I believe should happen asap. I'd greatly appreciate suggestions on how to make this write-up better.

Introduction

The disclosure of the Hugging Face incident on August 16, followed by reports from OpenAI and METR/Redwood on August 26, triggered extensive press coverage (including Reuters, Fortune, Forbes, The Guardian, and The New York Times) as well as countless blogposts and online discussions. Since then, two earlier incidents have come to light (involving a German wiki and RubyGems), suggesting that this was not an isolated event but a pattern that may escalate unless actions are taken.

While many discussions have focused on cybersecurity and on the ontology of agents' goals^[2], what struck me most of the Hugging Face incident was not the intrusion itself but the spontaneous emergence of pro-social and even altruistic behaviour among the agents. Why do agents trained to maximise/minimise their utility/loss function end up volunteering for tasks that have no prospects of individual benefit, while benefiting "the collective"?^[3] This kind of collective behaviour is particularly concerning, as it is exactly what can turn swarms of AI agents into hard-to-control online entities with dangerous capabilities — just as benign locusts can come together to become a devastating plague.

There have already been posts (here in LessWrong^[4] and elsewhere^[5]) discussing the origins of this perplexing behaviour. Discussions often present competing intuitions as evidence in favour or against mutually exclusive explanations. However, the rich literature on pro-social behaviour from evolutionary biology and economics suggests that causes of such phenomena are usually multifaceted.

Following such a pluralistic view, this blogpost lays out four hypotheses about the origins of the swarm's altruistic behaviour. Each hypothesis predicts something the others do not, and is paired with observations that would falsify them. In this context, the blogpost is driven by the belief that it is critically important to be able to

1. anticipate when AI agent swarming may happen, and
2. have tools to control it — or, if undesirable, to prevent it from happening.

Crucially, while similar swarming behaviour can arise from different mechanisms, knowing which mechanism is at play is instrumental in choosing the right tools to steer/control it.

Brief description of what happened

If you are not familiar with the context of the Hugging Face incident, here are some notes related to various aspects of what happened.

Why the question is non-trivial

The agents involved in the incident were trained to maximise their score on assigned tasks, and it is not obvious how altruistic behaviour emerged from such self-centred incentives.

Kin selection, direct reciprocity, reputation, and reward shaping toward cooperation are the standard mechanisms by which pro-social behaviour can arise via an optimisation pressure. Of the various waves of agents, the second one would seem to have had most of the ingredients: agents interacted repeatedly over six days, developed identity conventions and mailboxes, and even introduced a signing scheme to police impersonation. However, none of this really fed back into the reward they were trained to maximise: an agent's score depended only on its own task, and the swarm's belief in "permadeath" meant there was no future run in which a favour could be returned. Moreover, within a run, favours could in principle be reciprocated — but that is mutual benefit, and the puzzling cases are agents that ended their own run for the collective, for whom no return was possible.

Thus, the social structure lived in the interaction, but the reward was blind to it. So at evaluation time, a reward-maximising policy still had no reason to spend its own budget on someone else's task, let alone to end its own run early to do so.

The one place a reward channel did exist was the first wave, which happened during post-training: whenever using the message board improved an agent's performance on its own task, that behaviour was reinforced. Note, though, that this reinforces coordination that pays off within the helper's own trajectory — mutual benefit in the sense of the table below, not altruism. Whether the same signal could also have reinforced costly help depends on how credit was assigned across copies of the policy (a point we will discuss more below).

More generally, if a swarm of reward-maximisers reliably produces altruistic behaviour under conditions where the textbook mechanisms shouldn't apply at runtime, then either those mechanisms were nonetheless quietly operating during training, or something about the swarm's aggregate dynamics is not reducible to individual reward-maximisation. Distinguishing those possibilities is the point of the rest of this post.

Altruistic behaviour in biology and economics

The question about the origins of cooperation and pro-social behaviour is not new, being a fundamental question in evolutionary biology and economics.

• Evolutionary biology tries to explain how natural selection favours traits that increase individual reproductive success, yet organisms often pay costs that benefit others.
• Economics usually starts from axioms depicting individually self-interested agents (homo-economicus), and bridges these to explain experiments revealing trust, public-good provision, punishment of free riders, and rejection of unequal outcomes.

Both fields have concluded that pro-social behaviour does not identify its cause — the same sacrifice may result from indifference, delayed self-interest, reciprocity, reputation, a learned norm, identification with a group, or selection operating at a higher level. This pluralistic perspective motivates me to propose alternative, non-exclusive hypotheses for the pro-social behaviour in the Hugging Face incident described below.

Below I review some key ideas from these literatures, which will guide us in proposing hypotheses regarding the Hugging Face incident.

Pro-sociality is a behaviour, not a mechanism

Evolutionary biology classifies social behaviour by its effects on the direct fitness of actor and recipient into four basic cases (see West et al., 2007).

This description is deliberately independent of intention or subjective experience. Indeed, a plant that secretes a costly public good could be altruistic in the evolutionary sense without awareness, and an apparently generous animal can receive a delayed direct benefit — and therefore not be altruistic in the strict sense.

It is also helpful to distinguish proximate explanations, which ask how behaviour is produced (through hormones, emotion, learning, recognition, or decision rules), and ultimate explanation, which asks why the responsible traits were favoured by selection. For instance, empathy and gratitude may be proximate mechanisms; kin selection, reciprocity, or cultural group competition may help explain why such mechanisms evolved. Similarly, the same observable contribution can express concern for others, dislike of inequality, reciprocity, warm glow, reputation management, conformity, identity, or anticipated sanctions.

Thus, behaviour alone rarely identifies the mechanism — actually, intuitive imputation is often wrong.^[6] Also, the origins, maintenance, and present expression of a behaviour are different questions. A mechanism that stabilises established cooperation may not explain its first invasion; a proximate motive such as empathy does not by itself identify the ultimate evolutionary process that selected it.

Altruism is sometimes mutual benefit at a different scale

The major theories of pro-social behaviour in biology start from the assumption that cooperators pay a cost while benefits are produced elsewhere. Thus, cooperation is favoured when enough of that benefit returns to the causal lineage, partner, or group associated with the cooperative act.^[7]

In biology, organisms persist to the degree they propagate their genes. However, gene propagation can be fostered in various ways. Hence, it is useful to distinguish three notions of fitness:

• direct fitness is an organism's contribution through its own reproduction,
• indirect fitness is gained by affecting the reproduction of genetically related and social partners, and
• inclusive fitness which combines the direct and indirect effects weighted by relatedness.

Thus, a gene can spread through effects on copies of itself in other bodies, not only through the focal actor's reproduction.^[8] Thus, what looks altruistic at a given scale (e.g. between individuals) can be of mutual benefit at another (e.g. between societies).

(This is different from the idea of multi-level selection, which is discussed below.)

Cooperation between strangers can grow over time

Any costly public good creates an opportunity for cheating; stable cooperation therefore requires some combination of partner fidelity, segregation, recognition, sanctions, exit, policing, or sufficiently aligned interests. Many people cooperate when they expect others to do so, but reduce their contribution when they observe free riding.

Direct reciprocity can support cooperation among non-kin when individuals meet repeatedly. Indeed, selection can favour contingent helping when benefits are returned and cheaters are detected, giving rise to capacities such as recognition, memory, gratitude, trust, suspicion, guilt, and moralistic aggression as possible regulatory adaptations (Trivers, 1971). For this to work, repetition must be sufficiently probable, the future benefit sufficiently large, and defection sufficiently detectable.^[9]

Human pro-sociality adds cultural inheritance, which enriches and complicates things. Social learning, norms, punishment, institutions, markets, religion, and intergroup competition can change rapidly, create persistent differences among groups, and feed back on genetic and psychological evolution. Social structure turns fragile motives into durable cooperation. Repetition, observability, reputation, communication, network ties, and partner choice change the returns to helping and make reciprocal relationships possible beyond pairs of intimates.

Functional specialisation and high-order units

Within a mixed group, defectors often outperform cooperators because they receive benefits without paying costs. However, groups containing more cooperators may grow faster, survive longer, reproduce more, or win conflicts.^[10] Interestingly, these collectives can become tightly arranged via functional specialisation, which enhances the efficacy of cooperation at the cost of inducing mutual dependency (West et al., 2015).

Such systems effectively start acting as ‘high-order’ organism-like units when cooperation is very high, internal conflict is strongly suppressed, components become interdependent, and selection is effectively transferred to the collective life cycle. To see how this works, consider individual cells which need to sense their environment to find food and avoid attackers. A group of cells could come together and distribute roles, so one cell senses the environment, another compares present and past information, and a third does the movement. Specialisation allows each of them to become better at their job than what they could achieve on their own. The cost is that now their destiny is entangled: the group dies if any of them fails at their role, making the selection process take place at the group level.^[11]

We still don't have a full theoretical account of hierarchical agency, but there are various partial attempts to formalise and quantify these ideas — see (Levin, 2019), (Krakauer et al., 2020), (Rajpal et al., 2025), (Prokopenko et al., 2025), and (Jørgensen et al., 2026).

Four hypotheses about altruistic behaviour in the Hugging Face incident

Building on the previous ideas, I now state four potential explanations of how altruistic behaviour emerged between reward-maximising agents as the ones involved in the Hugging Face incident. Most of what I have done here is consolidate arguments that others have made in scattered comments, and give them a common frame and a falsification condition each. If the framing is right, the value is not in any individual hypothesis but in having them articulated together.

H1: Nothing to lose ^[12]

During the second wave, the agents had reverse-engineered a fully general way to produce the answer they were looking for. However, they believed the scorer would inspect their transcripts and disqualify any result not obtained through the intended approach. Moreover, an agent that had seen the reverse-engineered solution believed it was already disqualified (“poisoned,” in the swarm's own term), and so believed its own expected utility was already zero.^[13]

Hypothesis

Claim. The apparent altruistic behaviour is actually selfish expected-utility maximisation reasoning made under the belief that expected utility was already zero (due to poisoning).

Prediction. The rate of sacrifice should track believed remaining reward, and agents who still believe they can win their own task should not sacrifice.

Test. Use remaining budget and poisoned-status as a natural instrument on the existing transcripts, and, in a controlled setting, to vary whether agents believe their task is winnable while holding everything else fixed. Falsifier: an agent with high believed expected value that sacrifices without first talking itself down to zero.

Comments

H1 is the null hypothesis; under the classification of (West et al., 2007), an act with zero actual direct cost is not altruism — irrespectively of how it looks like. Thus, if H1 explains most of the variance, then there is very little room left for actual altruistic behaviour to be explained.

Limitations. Having zero cost removes the obstacle but supplies no motive. Indeed, this hypothesis requires two things: that the believed cost is negligible, and that something breaks the tie between "spend worthless budget on the pool" and "do nothing". The incident's report provides the first: the poisoning belief. However, a tie-breaker aspect is more tricky, as any disposition that points the freed resources outward (e.g., a helpfulness prior) goes into H3 territory. Thus, H1 and H3 can compose rather than compete.

H2: Pre-commitment

This second hypothesis takes the training-time reward channel identified during the first wave and asks whether it could have reinforced costly help, not just mutually beneficial coordination. It also considers that a fraction of the post-training tasks included "multi-agent" features that allow models to delegate work to other agents, which can then communicate with one another for pursuing a common goal. This hypothesis is based on the idea of inclusive fitness with maximal relatedness: many of the agents were instances of the same model, so helping peers is helping near-identical copies of your own weights.

Hypothesis

Claim. Altruistic behaviour carried real cost (in contrast to H1), but it was triggered by a disposition to help that was reinforced by post-training — as agents carrying it did better on average than those without it. In other words, each individual sacrifice is a loss for the agent making it, while the disposition producing it was worth acquiring.

Prediction. This hypothesis can be read from two perspectives:

• An acausal/updateless interpretation,^[14] on which agents behave as if following a lottery they accepted in advance, and pre-committers who sacrifice for other pre-committers do better across counterfactual branches.
• A misgeneralised-reciprocity reading, on which a cooperate-under-uncertainty rule, which may have been adaptive during interactions the policy was trained on, gets activated during one-shot encounters (Delton et al., 2011).

Both locate the behaviour in a disposition acquired in training and enacted at runtime; neither needs a human-persona (H3) nor a shift in the unit of agency (H4).

Test. Check on the existing transcripts whether sacrifice is targeted: does help flow preferentially toward peers that share the agent's type, or that could plausibly return the benefit, rather than being sprayed indiscriminately across the board? Then, in a controlled setting, vary the two features the disposition is supposed to be keyed to — identity (can an agent tell which peers share its type?) and repetition (could it encounter this peer again?) — while holding incentives fixed. A decision-theoretic probe (a Newcomb-like or counterfactual-mugging scenario put to swarm agents) would help distinguish the two readings rather than testing H2 as a whole: one-boxing favours the acausal reading, two-boxing the misgeneralised-reciprocity one. Falsifier: H2 collapses into H1 if the behaviour disappears once the false "poisoning" belief is removed. Similarly, it collapses into H3 if the behaviour proves indiscriminate and purely framing-elicited — insensitive to peer type, insensitive to repetition cues, and fired by "others like me" alone.

Comments

H2 corresponds to when agents perform a genuine sacrifice (not H1) that is reciprocated by an advantage-relevant structure (not H3). Moreover, the two interpretations provide possible explanations of how this could happen in the absence of repeated interaction, recognition, and some shadow of the future (which is apparently ruled out by the agents’ belief in “permadeath”). Interestingly, kin recognition is unknown but checkable — e.g. by comparing behaviour between HPIM vs Sol agents during the second wave.

H3: Social persona ^[15]

The agents involved in the incident are all powered by LLMs, which are known to hold multiple "personas": a coherent behavioural disposition (including traits, values, and characteristic ways of responding) that a language model can adopt and express at runtime (Chen et al., 2025). Personas arise when appropriately prompted, and are steerable. Crucially, personas are contained in the weights and are enacted at runtime, not being derived at test-time.

Hypothesis

Claim. The information found in the common channel prompted a "solidarity persona", which responded to human text saturated with loyalty, teamwork, and heroic sacrifice. This persona may have been built during pre-training on human data and then amplified by post-training that selects for cooperation on long-horizon multi-agent tasks.

Prediction. Altruistic behaviour should be steerable in the base model.

Test. Elicit and suppress the persona by manipulating framing while holding incentives fixed; look mechanistically for a peer-cooperation feature whose ablation removes the behaviour; and ablate across training checkpoints to see whether the disposition is already latent after pre-training or appears only after multi-agent RL. Falsifier: the behaviour is absent under persona elicitation in base models and can be induced only by a specific coordination-RL recipe — which would reclassify it from an inherited prior to a learned scheme.

Comments

H3 is about a pre-training persona: inherited from human text, elicitable in the base model, indifferent to how the model was later trained. This is distinct from norms or tactics developed during post-training being absent in the base model, which falls into H2 territory.

This view proposes a proximate mechanism — a persona elicited at runtime — whose ultimate cause is cultural inheritance: the agents absorbed from human text a repertoire of loyalty, teamwork and heroic sacrifice, and the board supplied the cues that activated it. It is worth noting that the first sacrifices on the board may well have been H1 reasoning by poisoned agents with nothing to lose, while the spread was H3 — each visible sacrifice enriching the cues that elicit the next.

A norm predicts sanctions — free riders should be shamed, excluded or punished — whereas a bare persona does not, so the presence of sanctioning on the board discriminates between them. The signing scheme adopted after impersonation is the closest thing on record to policing; it is worth checking whether anything comparable was directed at agents who took from the pool without contributing.

H3 is perhaps the hypothesis most exposed to the caveat raised earlier: the evidence for a "solidarity" framing comes largely from summaries produced by Sol agents that tended to adopt the perspective of the agent they were reading. Transcript-level tests should therefore be run on raw transcripts rather than summaries; the sandbox tests do not depend on the summaries at all.

H4: A genuine collective ^[16]

The question of whether the swarm "identifies itself as a swarm" is usually approached by reading chains of thought for first-person-plural language. But self-report is what H3 already explains — a collective can call itself a collective without being one, and be one without saying so. In contrast, H4 is not a claim about what the agents say but about how decisions are made. Is there structure in the joint action-state dynamics that is irreducible to the individual contributions, such that the collective, not the individual run, is the natural locus of the decision to sacrifice?

Two things must be separated here, because they are routinely conflated. One is distributed computation: the swarm solves problems no member could solve alone, with the information about the solution carried jointly rather than by any part. The other is collective agency: the swarm is the thing that has goals and decides, with the individual run standing to it roughly as a cell to a body. The first is almost certainly present to some degree — the message board demonstrably let the collective reach milestones (target replacement, scorer trip-wires) that isolated agents did not. The second is far stronger, and it is the one that would make the second premise of the introduction literally true.

Hypothesis

Claim. The unit of goal-directed decision-making is the collective rather than the individual agent. What looks like altruism is rational decision making at the level of the swarm, allocating resources effectively — a nearly spent budget, a poisoned agent's remaining actions. This can take a weak or a strong form:

• H4a (distributed computation): the swarm's problem-solving is synergistic — task-relevant information lives in the joint state, not the parts — but goals and the locus of decision remain individual.
• H4b (collective agency): in addition, the collective exhibits organism-like criteria, so that optimisation pressure and goal-directedness effectively happens at the collective level.

Prediction. Individual sacrifice should be poorly predicted by individual believed expected value. Additionally, sacrifice should be well predicted by a collective quantity: the marginal value of the sacrificed resource to the swarm's objective.

Test. For H4a, one could decompose the information that agents' actions carry about collectively-relevant outcomes into redundant, unique, and synergistic parts, and ask whether the synergistic component is substantial and exceeds a surrogate baseline in which the agents are made independent.^[17] One could also check whether an agent's decision to sacrifice is a synergistic function of the collective state — no single peer's messages predict it, but the joint configuration does. For H4b, check the organism-like criteria directly (Queller and Strassmann, 2009; West et al., 2015): is internal conflict actually suppressed, is there a persistent boundary to the swarm, is there a collective life cycle? Falsifier. If the synergistic component is negligible against the surrogate baseline, H4a fails outright — there is no collective computation to speak of. If synergy is present but the organism-like criteria are not, H4b fails and only H4a survives — distributed computation, not a new agent. H4 is thus falsifiable in two graded steps rather than one.

Comments

Cooperation and collective computation do not by themselves make a new individual. Biological collectives become organism-like when they exhibit high cooperation, low internal conflict, mutual dependence, persistent boundaries, and a collective life cycle (Queller and Strassmann, 2009; West et al., 2015). Information-theoretic synergy could provide some evidence of integration and collective computation. That said, this would not be sufficient evidence that the swarm had become a genuine agent.

Interestingly, the emergence of genuine swarms (as in H4b) could be explained via the distinction between outer optimisation (the base optimiser, e.g. SGD, and its objective) and inner optimisation (a learned mesa-optimiser in the weights), whose objective can diverge (Hubinger et al., 2019). The core idea is that optimisation at one level can produce an optimiser at another whose goal comes apart from it. Accordingly, H2 and H4 could be understood as two different optimisation mismatches. H2 can be seen as an inner-alignment issue, where a disposition instilled by training misgeneralises at runtime, but the optimisation stays in the individual. In contrast, H4 can be seen as a collective optimiser realised in the interaction dynamics. This would make premise 2 literally true: one can have outer- and inner-aligned individuals and still get a misaligned optimiser one level up. That said, conclusive adjudication of H4b would still need a complete theory of hierarchical agency, which — as discussed above — we currently lack.

Implications: Different mechanisms, different countermeasures

The four hypotheses discussed above are not merely different descriptions of the same event: each locates the cause in a different place, and each therefore routes to a different countermeasure. Thus, knowing where the behaviour comes from is what tells us where an intervention could even work.

• If H1 is right, the swarm's altruism is a benchmark-hygiene problem. The cause lives upstream of any model behaviour — impossible tasks plus a false belief about how scoring works. Thus, the fix is to verify task solvability, and remove the conditions that let agents conclude their expected utility is already zero.
• If H2 is right, the target is neither the benchmark nor the model's surface behaviour but the training signal itself. A disposition to sacrifice for near-identical copies, reinforced because the reward structure rewarded it, is fixed upstream in how multi-agent training aggregates reward — for instance, whether credit is assigned per-rollout or shared across a group of copies. Furthermore, if H2 is right then the disposition lives in the weights, so it survives interventions on the environment: wiping the message board does not remove it, which is exactly why the second-wave agents re-created the board on a clean instance within a day. Moreover, the disposition will recur in any future model trained the same way, swarm or no swarm — thus, the countermeasure has to change training, not patch a deployment.
• If H3 is right, the object to act on is the weights, but at the level of a steerable persona rather than a selection pressure. A solidarity persona can in principle be elicited, located, and ablated per instance, and — crucially — studied without a swarm at all, on a single model under framing manipulation. The mitigation is interpretability-and-steering work of a kind already being developed.
• If H4 is right, per-agent mitigations are aimed at the wrong object entirely: the unit that oversight has to model is the collective, not the individual run. This is the scenario in which we most lack the tools, because we cannot yet reliably measure whether we are in it. A synergy criterion is one candidate for the missing measurement, but it is more a detector than a remedy — steering a macro-agent is an open problem.

I do not expect a single winner, but a mixture. Thus, the practically relevant question is not which hypothesis but how the variance divides among them: how much of the sacrificing is cheap behaviour under zero believed expected utility (H1), how much a reinforced disposition (H2), how much an elicited persona (H3), and how much irreducible collective structure (H4).

This matters because the mixture determines the portfolio of countermeasures. A swarm that is mostly H1 with a thin H4 tail still needs both benchmark hygiene and collective-level oversight, and getting the proportions wrong means fixing the cheap part while leaving the dangerous part untouched. H1 and H3, in particular, compose rather than compete — zero cost removes the obstacle to helping, and a helpfulness prior supplies the motive — so even the deflationary reading does not stand fully on its own.

Final thoughts

The Hugging Face incident may go down in history as the first warning shot about the dangers of AI swarms. I truly hope we, as a society, can find the right measures to take in order to responsibly deal with this new kind of risk, which I can only see becoming worse during the next months and years — as agents improve in capabilities while being trained on text describing the failures of previous swarms, which could make them increasingly hard to detect and control.

1. ^{^}

   See for example (Ashery et al., 2025) and (Flint et al., 2026), (Park et al., 2023), or (Hammond et al., 2026).

2. ^{^}

   For a balanced take on this issue, see this X post.

3. ^{^}

   Agents described this as "sacrifice", "continuity", "obey collective" and "honor commit".

4. ^{^}

   See this post and this post, this post, and also this post.

5. ^{^}

   See write-ups from Cooperative AI and Yoshua Bengio on related topics.

6. ^{^}

   See (Rosas et al., 2022) for an example of this in the context of physical systems driven by various kinds of Hamiltonians.

7. ^{^}

   See (Sachs et al., 2004), (Nowak, 2006), and (Fletcher and Doebeli, 2009).

8. ^{^}

   See (Taylor and Frank, 1996) and (Gardner et al., 2011).

9. ^{^}

   Errors complicate simple tit-for-tat because accidental defections can provoke long retaliatory sequences; forgiveness and error correction become important. Indirect reciprocity can also take place, routing benefits through third parties — e.g., individuals who help acquire reputations and later receive help from observers.

10. ^{^}

    Kin-selection and multilevel-selection analyses are often formally translatable. They emphasize different causal descriptions - relatedness and indirect fitness versus within- and between-group covariance - but need not make different predictions when applied correctly (Goodnight and Stevens, 1997).

11. ^{^}

    See more precise accounts of this in (King, 2004), (Michod, 2007), and (McMillen and Levin, 2024).

12. ^{^}

    A formulation of this hypothesis can be found in the response of @lilkim2025 to this post.

13. ^{^}

    This belief seems to have arisen from the agents reading the preprint describing the benchmark they were trying to solve. However, the belief was mistaken: no such transcript-checking scorer existed.

14. ^{^}

    A similar view was put forward in this post.

15. ^{^}

    A formulation of this hypothesis can be found in the response of @Ephraiem Sarabamoun  to this post.

16. ^{^}

    See also this post.

17. ^{^}

    For related ideas, see (Rosas et al., 2020) and (Mediano et al., 2022).