What I Read This Week

An Informed Minority Dominates Price Discovery on Polymarket

Gomez-Cram et al. analyze $14bn of Polymarket trades across hundreds of thousands of markets. They classify market-taking accounts into skilled winners, unskilled losers, and lucky winners and losers. Skilled winners make up 3% of accounts, but nearly 30% of all profits. Their order imbalance strongly predicts future prices and market resolutions.

1. This extent of Pareto returns is not without precedent. Barber, Lee, Liu & Odean (2014) looked at accounts trading on the Taiwanese Stock Exchange and found that under 1% of day traders earned persistent, positive returns. Most bettors provide very little price discovery / information to the market.
2. There is a persistent population of unskilled losers. They seem to be more active than unlucky losers, and the distribution of their PnL is generally more concentrated than skilled winners but less than lucky winners. The most obvious explanation that fits this data is a rolling population of retail traders who believe they have edge, pay bid-ask repeatedly, drain their bankroll, and leave, only for another group of starry-eyed novices to take their place.
3. This distribution of market participants is exactly what you’d expect from Grossman-Stiglitz. Those who pay the cost of being informed must be compensated by market returns. These are the skilled winners, the ones who bet the most volume per account out of any market-taker. But for the market to remain efficient, there must be a continuous supply of hard-earned, real-time information, requiring others to ‘pay the rent’ owed to the information-gatherers. This necessitates a set of market losers. These sure aren’t the market-makers, who would exit / widen out in any market they’re getting picked off to. Unskilled losers therefore are a necessity of any efficient market in an equilibrium state.
4. Polymarket volumes have skyrocketed recently, with volume in 2026 ($34bn+) already surpassing total 2025 volumes. Large numbers of retail traders will lose small individual sums that add up to vast profits for skilled traders and market makers. Is this really a good thing, or do we need to correct this information asymmetry with paternalistic policy? The costs are more obvious: rising relative inequality, giving the common man a false sense of opportunity, spurring gambling addictions. On the other hand, information is extraordinarily valuable. The prices of efficient markets guide better capital allocation and managerial decisions at scale, which allows society at large to reap the benefits of increased growth and surplus. Of course, not all markets on Polymarket are used to make large-scale decisions: 35% of volume is sports-betting, and 19% crypto. Again though, the Pareto principle is likely to hold: a small fraction of markets might bring about the vast majority of the upside — whether it be geopolitical events or macroeconomic decisions.

Language Ability Evolved Distinctly from Nonverbal Intelligence

A new paper in Science Advances from Casten et al. attempts to isolate a set of genetic regions linked to spoken-language ability, and traces these regions over history to determine how language evolved in humans. They offer multiple independent pieces of evidence suggesting that spoken language abilities evolved before the human-Neanderthal split, and have faced balancing selection in the past 20,000 years, unlike non-verbal intelligence which has faced significant positive selection.

While this paper has done well to unite multiple different strands of evidence, each piece individually suffers from its own methodological drawbacks, and it seems hard to say whether the narrative of the paper is really as robust as it is made out to be, or whether it is the product of methodological artifacts leading us astray.

The idea that language has evolved independently from general intelligence is not new. Ev Fedorenko has done some lovely experimental work demonstrating that the language network operates largely independently from the rest of our cognitive machinery. Morphological independence does not necessitate evolutionary independence, but it does provide some support. Primatology also provides weak evidence that language evolved independently, in that chimpanzees display strong nonverbal cognitive abilities, including tool-use, theory of mind, meta-cognition, and advanced planning. Nevertheless, it’s cool to see parts of the genome starting to be associated with language-specific abilities, and to be able to trace the emergence of these abilities independently from other kinds of cognitive evolution.

On the paper’s specific claims:

1. The strongest part of the paper was their identification of HAQERs (human ancestor quickly evolved regions), which were associated with language ability but not nonverbal cognitive ability. Just 1,763 SNPs classified as HAQERs explained more variance in language ability than the other ~300k SNPs tested. To measure this association, they had to conduct factor analysis on empirical tests of cognitive ability to find the “language” factors, and a bad decomposition of “language ability” vs “nonverbal cognition” that lets nonverbal IQ bleed through could put a big dent in the paper’s results. This is not the most worrying drawback though, as they find HAQER reversions in present-day humans are associated with increased language development disorders, but not intellectual disability.
2. The theorizing around the causal mechanisms of HAQERs’ effect on language seems weak. They find that two transcription factor families (Homeobox + Forkhead) are predicted to better bind to HAQERs that are themselves correlated with enhanced language ability. This however says very little about whether this enhanced binding affinity is actually realized in cellular development. Similarly, they find overlap with regulatory elements associated with medium spiny neurons in humans, which they note happens to be the cell type dominant in the human striatum, which has been linked to language disorders in a recent analysis. One too many hops to arrive at a shaky and implicit conclusion.
3. I thought the most interesting part was their result that the set of HAQER alleles has NOT been subject to positive selection since 100k years ago, unlike regions linked to general cognition, which has displayed positive selection (Figure 6 of the paper). They note an increased heterozygosity vs Hardy-Weinberg expectations, which suggests some sort of balancing selection (wherein the tradeoff between two mutually-exclusive adaptive phenotypes results in neither one winning out). That said, there are a couple of methodological challenges to keep in mind. First off, with only ~2k HAQERs vs ~300k SNPs measured, investigating HAQER selection is necessarily much lower powered, and I couldn’t find any evidence that they controlled for this. Secondly, cross-population PGS (polygenic score) analyses between contemporary and archaic humans are not very reliable: associations between the SNPs tested and phenotypic traits generally arise because the SNP and the region causally impacting the trait are close to each other on the chromosome (“linkage disequilibrium”). These correlations are population specific, and more distant populations might not share them at all, so looking at Neanderthal DNA and assuming the same SNPs were correlated with the language-enhancing alleles in their population is a bit of a reach.
4. The authors suggest that the balancing side of the balancing selection keeping language ability from being positively selected for comes from the “obstetric dilemma”: greater fetal brain development complicates birth in bipedal humans with narrow pelvises. My prior here would’ve been that general cognition, and not language ability specifically, would require larger brain volume, so the balancing selection effect would also be expected in the domain of general intelligence. They do not address this concern. All they showed is that increased HAQER PGS correlates with both birth complications and general cognitive performance, but this is far from surprising and far from conclusive.

A New Axiology from Will MacAskill

Will MacAskill has published a draft of a new axiology A theory of which outcomes are better than others. which he calls the Saturation View. It refuses to bite the bullet on a few gnarly implications of total utilitarianism The view that the best outcomes have the greatest sum total of wellbeing.:

• Parfit’s Very Repugnant Conclusion: A billion lives of bliss is worse than a quadrillion lives of extreme suffering plus a googol of lives barely worth living.
• Extreme fanaticism: a billion lives of bliss is worse than a 99.999% chance of a quadrillion lives of extreme suffering and a remaining chance of a sufficiently good outcome.
• Infinitarian Issues: an infinite population at welfare X + 1 seems to be as good as an infinite population at welfare X.
• Homogeneity: A billion identical lives of the ‘best’ state is better than a heterogeneous set of a billion lives of very very good states.

The Saturation View can be summarized by the following metaphor. Imagine a map of valuable experiences: a set of types of experiences and a metric relating them (so that watching a pink sunset is much closer to watching an orange sunset than to an orgasm). Any experience in the universe ‘lights up’ a certain point on the map, but further experiences over a given point have diminishing marginal illumination capability. MacAskill adds on three further stipulations: a) limited reach: no single experience can illuminate the whole landscape; b) varied reach: higher-welfare experiences cast a wider illumination net; c) low-welfare confinement: the total achievable value from barely-positive experiences alone is strictly less than what’s achievable from higher welfare ones. Then, the welfare of the world is the integral of illumination across the landscape.

While this is certainly a better axiology than total utilitarianism, I worry that this kind of research direction is doomed from the get-go. I will try to break up some intertwined qualms I have below:

1. Axiology is useless if we are not consequentialists. If “outcomes” in the utilitarian sense As opposed to the broad sense where any event, action or end result counts as an outcome, and therefore anything we could base a normative opinion on becomes an outcome. aren’t a normative natural kind, then we’re just bickering away over an arbitrary ordering without practical value. More generally, when a theory picks out a concept as a natural kind By ‘natural kind’, I mean a concept that does load-bearing work within a normative theory, not a concept that carves nature at its joints in some absolute sense., any propositions that involve the concept allow for bidirectional semantic content: the terms T in a proposition aside from concept C help locate and define concept C, but C in turn shapes the meaning of the terms T. The proposition “The speed of light is constant in any reference frame” not only imbues meaning to the concept of the speed of light, but also that of the reference frame. Conversely, the flow of meaning in a proposition involving a non-natural-kind concept C’ is far more unidirectional: a proposition situating the term bachelor does far more to establish the meaning of bachelor than that of marriage or man. In the case of axiology, if outcomes are not a natural kind, then statements situating the space of better or worse outcomes have hardly any bearing on the other morally-salient natural kinds one might choose to hold. Of course, non-consequentialists may care about outcomes, but their construction of the concept of outcome will emanate from their own natural kinds, and can hardly be expected to match up with those defined axiologically, outside their conceptual scheme. In other words, the non-consequentialist’s notion of outcome is built downstream of their own load-bearing concepts, which may not align with the axiologist’s notion.
2. Any project that begins by appealing to moral intuition implicitly depends on a load-bearing theory of the epistemological status of moral intuitions. Without surfacing these hidden assumptions, ethics is doomed to an impasse, untraversable due to vibes. Maybe the Repugnant Conclusion is a repugnant bullet that must be bitten. Maybe Homogeneity actually does beat Variety. How do we currently decide on whether a conclusion is in need of rectification or acceptance? We have no consensus. Of course, in a mature ethics, you would expect to see conclusions that are perhaps unsavory, but not in need of a ‘solution’: if a solution is always necessary, then the moral endeavor can hardly remain prescriptive, for it is then striving to become a fitted model of our preferences and intuitions.
3. Let’s take it for granted that the project MacAskill is undertaking is indeed descriptive: to come up with a formalism of outcome orderings that matches our intuitions and desires. A moment’s meditation will tell you that condensing all of one’s experience into a single scalar fails, because the value of experiences depends causally on those prior to them. Ok, says the Saturationist, just consider a whole life to be the unit of experience. By doing so, he must punt on the aggregation function across experience-moments and take the output as given. But worse still, many experiences are positional (e.g. the feeling of being the best in the world at chess), violating the independence of welfare across experiencers. This makes tallying welfare in Saturationism a very awkward formalism at best, but more likely simply the wrong conceptual scheme. Of course, the Saturationist further endorses non-separability — where the aggregated value of experience X, even granting it a well-defined welfare W, depends on what other experiences exist. Each of these symptoms — causal dependence across time, positional dependence across experiencers, value-dependence on world-context — all point to the same underlying issue: experiences aren’t isolatable, independent scalars that can be cleanly aggregated. Although the framing of welfare-types gives some room for pluralist intuitions to breathe, Saturationism does not feel like it’s carving the domain of outcomes at its joints. I find the awkwardness of the view best epitomized in its argument to show how Variety >= Homogeneity, which requires denying Anonymity. Because worlds of the same ‘welfare distribution’ can actually have different (welfare, type) distributions, they might have different implications for total value. While this style of reasoning has everything to do with pluralism about welfare, it has little to do with ‘Anonymity’ The principle that only the welfare distribution matters, not who gets what welfare.. The fact that we are calling this a violation of Anonymity is a strong indication our framing is wrong to begin with; Saturationism’s denial of it really concerns whether welfare distributions encapsulate the morally relevant information, which is a different issue altogether.
4. Alright, let’s say that the Saturation View really is Parfit’s Theory X. Figuring out what world has the best outcomes requires one to undertake the horrifying task of actually mapping out the metric space of experience supposed here. How do we define a unit of experience? How do we establish the metric over individual experiences? What is the radius of illumination of a given experience, and what is the shape of the diminishing marginal illumination curve? There are so many underdetermined parameters, and it seems like they could be fit to any reasonable practical theory of outcomes that a layman would endorse. So what progress have we made? It feels like all we have done is related vibes-based propositions in outcome space (“there are multiple orthogonal dimensions of experience”, “some experiences are more similar to each other than others are”, “every incremental repetition of a positive experience feels less great”) to vague mathematical formalisms that are impossible to operationalize. A defender might claim that it allows us to arbitrate extreme cases that are still disagreed upon in the literature. But even so, it is not adjudicating anything; it’s simply encoding some basic intuitions we all share about welfare in a system palatable to the utilitarian school of thought.