In particular, public drug consumption is so rife that I witnessed a couple smoking a joint outside Bradford's only shopping mall, in front of crowds of customers (some families with kids!). Rather than refer them to the police, the security guards allowed the smoking, and watched them as they smoked. An open secret is that drug possession for personal use is de-facto decriminalised in Britain. Virtually no one goes to jail for possession alone, unless involved in its supply. This begs the question though, why then bother legalising drugs? Surely we've hit a tolerable equilibrium, where we adopt the libertarian approach of not dictating what we put in our bodies, yet use criminalisation as an excuse (for private agents) to police the externalities associated with public consumption?

One case for legalisation is that quality improves. There exists asymmetric information in drug markets, with consumers not knowing the potency of their product. This was a ubiquitous feature of cocaine markets for a long time. Suppliers converged to a monopolistic competition with a markup set in part by diluting the product. Yet today, the average purity levels of street-seizures of cocaine now surpass 80%. The dark web is easy to use to get the pure product at similar prices, so street-level suppliers must compete with that. In other words, black markets already follow a structure of Bertrand competition. Arguably, government legalisation and regulation will actually reduce consumer surplus. Indeed, most leftists back legalisation just to get another market to tax and regulate to oblivion1. With cocaine now increasingly becoming a staple of British social and cultural life, it’s vital that the state is kept as far away from it as possible.

Hence, as a libertarian, I’m surprisingly ambivalent about legalisation. Perhaps in countries such as America, where there is a nonzero chance of going to jail just for personal use (particularly for black men), and their jails are overflowing with nonviolent drug users, ending the war on drugs makes sense. I find it particularly absurd that you can face years in prison just for being convicted of possessing paraphernalia, then violating probation on a positive drug test. Yet why should we import the American-themed discourse on this issue? An unregulated black market is surely a libertarian dream?

Not so fast. We spend millions fighting drug dealers. When we already face a looming debt crisis, we need to consider as many options for savings as we can get. Hence, the fiscally conservative case for legalisation is compelling. We can even divert some of the funds we spend trying to catch dealers towards driving the smackheads as far out of public sight as possible. To the extent that the government or local authorities own land, they must enforce their property rights on that land, therefore the state has a moral obligation to eradicate public drug use on its streets. Any other solution erodes the very idea of property rights, setting a bad precedent. This is before we get to the externalities surrounding antisocial behaviour, used needles being littered, and even the pungent smell of cannabis.

UPDATE (02/06/2026): Perhaps the most interesting question relating to legalisation is how norms will shift. Yes norms will change. I think most of those arguing for legalisation are in denial - drug consumption will be expected to increase, and it’s important to consider the tradeoffs here. We know that some particularly weak individuals whom struggle to resist temptation or “peer pressure” are harmed: think of those migrating to harder drugs, or those whose lives have been ruined by addiction. Our current abstinence norms (outside a few subcultures) raise the search costs to acquisition, yet this isn't always individually-rational or incentive-compatible, so to enforce this equilibrium we impose stigma as a penalty for deviation.

Of course, drug use tends to be correlated with “loserdom” (the question as to what proportion of this is causal vs selection remains open), so in part this stigma reflects a real signal. However, some drug consumption amongst some individuals can be expected to complement their cultural consumption and hence actually increase their human capital. I suspect this is disproportionately true for high-IQ individuals. Plus for those prone to criminality (whom are also more impulsive than average), some drugs present a socially valuable outside option in that they substitute for violent crime. Violent crime has plummeted throughout the last few decades whilst opioids have increased in popularity. I don't think this is a coincidence: to the extent that opioids are substitutes for alcohol, and present early mortality risk for those who might otherwise be committing crime, they should be expected to promote crime reduction by sedating the antisocial. Acquisitive crime rises in the short-run, but addicts are disproportionately antisocial so many are likely to engage in theft in the absence of drugs anyhow, and again increased mortality risk mitigates this rise. Therefore, we cannot state for certain that the socially or individually optimal level of drug consumption is zero.

What we can state with high confidence is that this stigma is uncorrelated with total social or present-valued lifetime private costs, as I explained in my earlier post on rational addiction. Then I attributed such to the stigma instead reflecting adjacent complementarities, which makes sense when we consider it a tax for deviating from an abstinence equilibrium; more likely to be socially optimal the more addictive a drug is. However, consider the public choice element of prohibition. Alcohol has been present in Western society for centuries. Cannabis, cocaine, and opiates originated in different climates to those of Western Europe and North America, so arrived to the party only after we started exploring the New World. Hence the alcohol industry had more time to invest in lobbying, and this in tandem with cultural norms on drinking already being mature and well entrenched contributed to a relatively swift reversal of prohibition. Arguably, one can consider the war on drugs to be protection for the alcohol industry.

By ending this protection and establishing a free market in intoxication, society will converge to more efficient norms and cross-subsidisation of cultural life. This to me is the strongest case for legalising drugs. Alcohol consistently ranks as one of the most dangerous in its social and private costs, and there is no compelling reason why other drugs shouldn't be able to compete on a level-playing field.

Here we have Looking at the Flowers (25th June 2022). What is remarkable in this painting is that everything but the flowers themselves are represented in 3D - a unique feature of Hockney's innovative “photographic drawings” technique. He pioneered this method whereby he walks around the object, takes many photos from different angles, and uses a software model to aggregate them into one 3D image.

I can attest that it works! As I approach the landscape, the chairs become almost tangible, as if they are figures in a toy set. The stool with the blue vase projects forward. The two Hockneys complement this immersive experience, and make you feel as if you are joining him (or should I say them!?) in this living room. In this sense, what we have is not a photograph, as a photograph merely captures an object at a single point in space-time, but rather a painting depicting our actual subjective experience of perspective which involves motion through space-time. Yet this experience arises from a combination of photographs depicting fixed objects.

Stare into this phenomenal invention long enough, and you soon realise that multiple perspectives are compatible with the same observed reality and underlying structure (an observation analogous to special relativity?)1. In this sense, the final product mirrors the very best Mondrian paintings. The value of the artwork is derived from the synthetic whole, yet that synthetic piece arises from a combination of smaller interchangeable pieces. Think of art conducted in this manner as a language. In neoplasticism, the letters or syllables are the lines, grids, and colours. Here it's the photographs of the same object from different angles themselves. The former sought precisely a universal language of art, hence a universal beauty, in the quest for a universal and immutable truth. The latter seeks to demonstrate (successfully) that the same universal reality can produce or be consistent with multiple perspectives. Nonetheless, the core theme, of universality and objectivity in beauty and truth, is looming as the elephant in the gallery.

A very Platonic vision of aesthetics arises. It's not enough to decompose the painting: the whole is the immutable object that beauty is derived from. Beauty is simply indivisible, as are all the forms. Beauty is a function of universal and objective reality, and so is itself universal and objective. From order, represented by a decomposible lattice of symbols (lines and grids in neoplasticism, photos in Hockney's work), you get the indivisible whole, and so increasing disorder is represented. In my view, the very best artworks encapsulate this increasing disorder after the final output. The second law of thermodynamics applies to art and beauty too, hence another reason for the Platonic vision of beauty!

By the way, here are some of the other exhibits. These same twenty flowers also appeared as single paintings, each framed individually:

What I find particularly striking is the red background with the brownish hue. Or is it the reverse? Again this confirms Hockney's overrarching point on the same reality producing multiple perspectives.

Likewise, he has done a good job emphasising the flowers, vase, cloth, and even the shadow, as distinct objects. Each spring out at you, independently of the other. Yet without the other, the paintings are simply incomplete. He achieves this via his signature bright colours, which are again a necessary ingredient in generating the whole. Change the brightness or the colour even slightly, and the painting is discordant and loses harmony.

This is my favourite piece, inspired by Monet's lilies (below Hockney's), entitled Water Lilies in the Pond with Pots of Flowers (10th-22nd June 2021):

Hockney spent almost a fortnight creating this single piece, as six distinct drawings on an iPad. Who knew that such a resonating picture can be formed from a few strokes on a screen2? In his quintessential style, he attempts to replicate the perception of viewing the natural landscapes in real life. Although I find the frames clunky, they highlight that we're looking at six objects combined together rather than a single photograph captured instantaneously.

Of course, software algorithms were deployed to create the highlight of the exhibition, and that work was produced before the advent of LLMs or diffusion models. As in many fields, AI can be expected to complement rather than substitute human talent. AI-generated works per se tend to display an overtly “artificial” aura behind them; compromising their beauty. However, as artists on the frontier grapple with the expanded possibilities afforded to them (as Hockney did), we will find that (as with all technological advances) culture benefits as a result.

Ultimately, the genius of Hockney is that he deliberately shuns photographs, yet generates the same experience of perception, yet in real time and motion as opposed to a fixed capture. Everything, including the ubiquitous bright colours, is constructed immaculately to produce this phenomenon. As you gaze and ponder, you derive a sense of what it feels to see. The utility of art comes not just from the object itself, but in the entire act of viewing. Indeed, this is why you should visit more art galleries. You need to see these works in person to gain a full appreciation of their aesthetic qualities and beauty.

If you are in the West Yorkshire area, or anywhere in England3, you should come up to visit. Whilst Hockney's death adds poignancy, you get a unique first-hand experience of viewing the work of one of Britain's greatest artists of all time. What's more, all of this is free!

UPDATE (03/07/2026): I was thinking the other day that Hockney was inspired by Monet's impressionism, as one can see via the heavy emphasis on nature and bright colours. Yet does this somewhat undermine my Platonic argument? Impressionism was a seismic change to the art scene at the time, away from the grandeur of Renaissance themed works such as Raphael and DaVinci, and defenders of that old paradigm could have made the same argument regarding the Platonic objectivity of beauty. Here's Fable to tell you why this artistic innovation is actually more consistent with the forms than earlier movements.

One notable explanation is tractability. Consider the French port vs porte, or point vs pointe. These words differ only in gender, yet carry distinct semantic meanings. Gender allows you to apply constant scaling to your lexicon, thereby reduces the costs of creating additional words. You suddenly have a lot more options to play with. In addition, noun classification can assist with the general process of categorisation. For instance, Swahili has a noun class for abstract nouns. The map of semantics then becomes a lot easier to navigate.

Likewise, gender facilitates comprehension. If the noun for uncle is masculine, and the noun for aunt is feminine, then this classification eases learning. You remember that the word for uncle or aunt must be masculine or feminine respectively, which reduces the pool of words to recall. The noun possessing a gender also provides additional information for reference, as clues if you struggle to recall, if you’re just acquiring language or learning a foreign language.

Sounds plausible enough, but then why do many languages possess a neutral gender? For the same reason that all English pronouns aside from the third-person singular do not possess gender. Sometimes we may want to aggregate the genders together, or sometimes the semantics of a word possesses a meaning that corresponds to both of the genders to some degree. This is contrary to the postmodernist claim that gender is an arbitary classification that we impose on the world, rather than us using the simplest tools possible to assist our understanding via establishing linguistic correspondance to reality.1 In economics pedagogy, the most common metaphor to illustrate the purpose of theoretical models is the Tube map. We judge a formal system, or at least a matrix of associative relations, via how easily using such allows us to understand the world accurately. Language is no different.

Nonetheless, the gender classification itself can appear somewhat arbitrary. Why for instance do only the third-person pronouns in English possess gender and the others do not? If we are referring to multiple third-persons, that we want to refer to as individuals rather than a group, then the gender of our objects is an intuitive means of distinction. This can be consistent with Derrida's perspective that language creates and reinforces social hierarchies. It can also be consistent with a non blank-slatist interpretation of human nature where the division of the sexes is one of the key phenomena that drives natural selection. However, the fact there is always a semantic logic behind the use of gender distinctions suggests this classification is not solely arbitrary, and therefore the direction of causality is likely from human nature to language as opposed to the reverse.

What do we make of the claim that gender classification is inherently sexist? Is it even related to reinforcing the patriarchy anyhow? Many large families, such as Turkic or the East Asian families, lack gender. Are they egalitarian paradises? Why do we not have classification based on race, nation, or any other tribal identity2? There does not appear to be a correlation between discrimination vs noun classification, which again undermines the postmodernist view of epistemology.

Garicano, Aghion, and Bergeaud emphasise that America holds a clear advantage on productivity growth, the frontiers of output such as the tech industry, and business dynamism. This generates meaningful and substantial spillovers to Europe via falling prices and gains to consumer surplus (I, as a European, am using a smartphone to write this on Substack at zero fixed or marginal cost3 after all!). However, the labour-market gains in wages and employment in these industries are disproportionately concentrated at the headquarters and labs of these frontier industries, owing to agglomeration effects (which themselves occur at the domestic nation). These benefits are non-tradeable, and so America disproportionately gains relative to the rest of the world. To argue otherwise would be tantamount to claiming that growth in the US tech industry benefits the Bay Area and the Rust Belt equally4!

The other day on X, I stumbled across this paper. As Romer emphasised in his work on endogenous growth theory, technology diffusion across countries (and more recent work examining diffusion within countries between ‘frontier’ vs ‘laggard’ industries5) is not instantaneous. Gains to TFP in America take time to spillover into Europe. This causes TFP growth to be higher in America at any given point in time, and these differences compound over time. Krugman highlights a convergence that tends to dissipate once accounting for endogenous growth dynamics: a point that he's all too-aware of, of course…

On a more technical note, we shouldn't necessarily rely on current PPP comparisons as they emphasise purchasing power today. This bias towards the present neglects the productivity growth and price changes accruing over time. Therefore current measures tend to exclude much of the gains arising from (relative) growth. Using a base year to hold prices constant is a better means of capturing these dynamics, despite treating a non-constant variable as constant at a somewhat arbitrary moment in time. Yet no matter which base year you choose, the divergence tends to arise, therefore constant PPP is a more robust measure too.

Of course, international measurement only gets harder from here. When estimating inflation domestically, you require changes to consumption bundles and the quality of the goods and services within. For PPP, you must repeat this exercise for these differences in representative bundles across countries at the same point in time. Simply put, a reasonable prior is that you lose precision via switching from national deflators (which constant PPP is equivalent to) to current PPP. Product sampling and quality adjustment techniques of course differ cross-country. In Britain, the HMRC payroll data vs the ONS’ surveys imply entirely different dynamics for GDP per hour growth, so even statistical agencies within a country disagree!

However, these deflators tend to neglect convergence in tradeables prices. PPP eliminates international price differences for tradeables as a result of exchange rates: best emphasised by The Economist's Big Mac Index. Nonetheless this adjustment is only really valid for tradeables, and non-tradeables form a large proportion of our economies. The Krugman camp is adjusting by using a measurement ill-equipped to answer the questions they're asking. Nonetheless, we can disaggregate the national accounts, and this does not seem too empirically relevant overall. PPP falls faster than the prices of tradeables and this pattern reverses for non-tradeables, and these roughly offset. Again, whilst Krugman's point may be valid for the tradeable sector, Garicano et al are correct with respect to non-tradeables, which form a larger share of both our economies.

In any case, the fact that America leads Europe in terms of productivity growth is not really in dispute6. Ultimately, it’s this variable that determines growth and hence differences in living standards across countries. Indeed, this advantage is the most robust point emphasised throughout this debate.

Our model

Suppose that the rate at which matches occur is described by an exogenous real-valued φ and match quality (and surplus) an exogenous real-valued q. The number of matches M is a Poisson process of rate φ. For simplicity, q is normally distributed i.i.d with mean zero and variance of one4. Individuals will form matches if q is weakly preferred to their reservation utility, which is a function of their outside options and the expected NPV of all possible matches preferred to this match:

where b is the outside option and F(q) is the C.D.F. of q.

Phones increase the value of the outside option to b’=b0+ε, whilst φ increases. The number of matches is M = φ(1-F(R))N for N being the total number of those seeking a romantic relationship5. So Δφ>-Δ(1-F(R)) for phones to increase matching via my proposed channel of reduced matching barriers, and vice versa for the smartphone theory to hold. Can we introduce some numbers to arrive at a precise conclusion? Yes we can. Let's now turn to a calibration exercise.

Calibration

For over a decade now, online dating has been the most common channel via which relationships form. About 30% of US adults have used online dating apps, and 12% are in romantic relationships as a direct result of online dating. Considering all the online platforms via which matching can occur, and those platonic friendships online that eventually develop into romantic relationships, these numbers understate the extent to which the internet facilitates matching. On the other hand, these figures are heavily US-centric, whilst fertility has fallen worldwide. Nonetheless, I think the overall effect has been to increase the matching rate, and so therefore I will use a rise from φ=1 to φ=1.3 as my choice in parameters for this calibration6.

There is also some evidence that people prefer matches formed online relative to offline, and if anything revealed preferences from dating behaviour shifting to online imply that Ariely et al (2010) likely understate the improved matching efficiency. However this reduces acceptance rates in the model, as people become more selective given that reservation utility is higher. Perhaps a change in 1-F(q) from 0.3 to 0.2 is plausible.

The notion that phones are a substantial gain in the outside option value is uncontroversial, and perhaps this rise is up to around half a standard-deviation point of average match surplus. I will use b0=0.2 and b’=0.5.

From the model, using our chosen parameters, these are the values of reservation utility prior to smartphones. R is 0.494 at a discount rate of 2%, and 0.485 at a 5% discount rate. After smartphones, if β=0.98 then R=0.7548 and if β=0.95 then R=0.747.

Why my priors imply null effects with high uncertainty

With these numbers, the cheaper matching effect roughly outweighs the effect of the gains to the outside option, albeit this result is highly uncertain and sensitive to the choice of parameters involved. Adjusting for this, the model implies that the effects of smartphones on number of matches is close to zero yet almost certainly non-significant and not robust.7

One significant limitation of this exercise however is that I have not estimated a regression structurally to inform these parameter values of match rates, changes to cross-country match rates after smartphone adoption, and an outside-option proxy (aggregating a weighted mix of app usage intensity, time on social media or dating apps, survey data on relationship quality, and so on). If there exists high-quality datasets for this empirical exercise, perhaps a staggered DiD is possible, and one can then feed the estimates into this model. More research on this topic is certainly necessary, otherwise model-free and empirically slack commentary will prevail in this discourse.

Share

So what explains the baby bust?

In general8, I believe that factors such as high housing costs (a result of persistent NIMBY policies across the West) and declining maternal and child mortality (meaning people that previously internalised the risk of death are adjusting their planned births downwards) are the most compelling explanations. Less kids amongst families already giving birth seems to dwarf the impact of rising childlessness.

The cross-country data clearly rejects the thesis that feminism (in particular increased female labour-force participation) is driving the decline, as fertility is collapsing in highly conservative societies such as Italy and China, and generally is declining regardless of how feminist a nation is9. To the extent that the motherhood penalty is higher in such societies, this should actually reduce TFR in equilibrium.

However there are some cultures where TFR is stable and often high. The Amish and Haredi Jewish communities, and to a lesser extent Central Asians, are the most commonly cited examples. Although fertility is declining in sub-Saharan Africa and the Muslim world too, those TFRs tend to be higher than those in Western or East Asian societies. Hence “cultural” factors likely play some role (with spillovers observed), yet at this stage features as nothing more than a residual lacking precise mechanisms or estimates - instead being a kitchen-sink for our (often ideologically motivated) priors.

“Here is my take on AI writing. I don’t personally like it, but that’s taste, and I’m sure it’s possible to post train the model to write “well” (conditional on some notion of taste). But that’s besides the point for me: AI writing makes all writing sound the same. If you use AI enough it takes about 30 seconds into a piece to realize it, and then my brain kind of shuts off because that’s the 20th time that day I’ve read the same cadence and tone.

For code it’s fine since the only goal is to get it to run. But if you’re trying to write, eg a substack, part of the goal is to keep the reader engaged and connected to you—the writer—and for that you need to have your own “voice”."

AI for writing is a shortcut. And for some purposes it makes a lot of sense. But I think for writing it is in the interest of the writer to differentiate themselves by using their own voice. Especially as more and more people start using AI.”

This is the result:

"My view on AI writing is perhaps less "AI is bad" and more "AI compresses variance".

I don't personally enjoy reading most AI-generated writing, although that is largely a matter of taste. I assume it is entirely possible to post-train models into writing "well", conditional on whatever objective function one has in mind.

But that isn't really my concern.

The issue is that AI writing increasingly converges toward a similar cadence and tone. After enough exposure, one develops a kind of detector. Roughly thirty seconds into reading something, you realise: I've already read this nineteen times today.

And then my brain quietly switches off.

Code feels different. The objective function for code is fairly clear: does it run? Does it solve the problem? Elegant code is nice, but ultimately secondary.

Writing serves a different purpose. If you're writing a Substack, an essay, or even a long X post, part of the output is not merely information transfer. The writer themselves is part of the product. You want readers to remain engaged, but also to feel some connection with the person behind the words.

That requires a voice.

AI is obviously useful as a shortcut. For some tasks it is a very good shortcut. But I increasingly suspect that for writing, as more people adopt AI, preserving your own style becomes more valuable rather than less.

Differentiation compounds."

I'm not impressed. The first time I've been dissappointed with Chat, which is a large signal on the quality of AI prose.

Jobs as task bundles

I will start by introducing the Coasian notion of jobs as a continuous bundle of distinct and discrete tasks: inspired by Garicano et al (2026). The idea is that exposure measures can be misleading as jobs aggregate multiple tasks rather than being uniform in composition. Some of these tasks are complementary, and some substitutes. Different inputs can complement some tasks and substitute others. Bottlenecks in enacting one task can delay the entire bundle operation1. Therefore exposure measures assume a linearity that simply does not exist.

Crucially, the effects of AI in this model depend on how tight the bundle is. If the bundle is weaker, so it’s cheaper to break the bundling, AI automation of some tasks are more likely to narrow the job boundary and pose greater disemployment effects, and vice versa. In fact, if the bundling is sufficiently strong and bottlenecks tight, then AI automation improving productivity in some tasks may enhance the overall occupational productivity - leading to positive employment and wage effects.

Consider a job as a production function of two tasks:

y = (q1^α)(q2^(1-α)).

Task one can be automated, whilst two cannot. There exists a time constraint, t1+t2<=1 for nonnegative times devoted to tasks one and two respectively.

Two scenarios are now introduced. If bottlenecks are sufficiently tight such that task one can only be partially automated, then output is given by:

1. y = ([{(μ(m))^(1-η)}{t1s1}^η]^α)[t2s2]^(1-α)

where μ(m) is an increasing function of frontier AI capability m. Si denotes the share of the job represented by task i. The shares are equivalent to the comparative advantage of the worker with respect to different tasks; in other words his abilities (relevant later). η gives the “human share” of output within a job. Notice that we have simply represented qi as a function of ti, si, and μ(m). Output remains Cobb-Douglas, with AI featuring as labour-augmenting technology.

If bundling is weak, then:

2. y = (1-c)[{μ(m)}^α]s2^(1-α).

Our employee focuses exclusively on task two. Task one is fully automated at cost c, which incorporates the bottlenecks and Coasian frictions discussed. Here, AI is not labour-augmenting but a distinct input in itself.

As Garicano has noted, when m is sufficiently high and c sufficiently low, more jobs converge to the weak bundling case. Eventually, this is how AI could replace most human labour. For simplicity, let’s call this AGI. Here we treat both m and c as exogenous, yet in reality, as m rises, c falls2. Nonetheless, social and political-economy factors alone (as well as Hayek's famous coordination problem) will ensure that c remains strictly positive, as well as Moravec's paradox (although robotics capabilities are advancing too, albeit at a slower rate). Therefore even with AGI, human employment remains positive. Indeed, this is a recurring theme across all the models covered here. In other words, to maintain a future career, look to sectors and invest in skills where humans still maintain a comparative advantage3.

Share

If jobs are task bundles, what does this mean for wages?

In scenario 1, our worker aims to maximise output4 subject to his binding time constraint5:

t*1 = αη/(αη+1-α), t*2 = (1-α)/(αη+1-α).

Both the output and human shares determine task allocation within a job. Raising either raises the time devoted to the corresponding tasks. This is attenuated by the cross-task implications: both tasks are essential for the job, yet allocating more time to one task reduces the time available on the other. By substituting into our production function, you get equilibrium output.

Wages are a function of revenue, Py. Let P=p1+p2, with pi simply being the share of value that task i contributes to production. p1q1 = αPy and p2q2 = (1-α)Py. In scenario 1, the worker gets the human share of revenue for task one.

We now arrive at one of the key results. The wage effects differ according to whether AI is complementary (the labour-augmenting scenario 1) or a substitute (scenario 2). ω = (αη+1-α)Py or ω = (1-α)Py for scenarios 1 and 2 respectively. The relevant shares determine the allocation of the wage across tasks and inputs, as is standard in perfect competition. Crucially, in scenario 1, as long as η is positive, then αη+1-α>1-α, so wages are higher when AI is labour-augmenting relative to when AI is a substitute.

There exists heterogeneity however in how workers will allocate across occupations depending on their abilities. Let s*1 divide workers into low task one ability and high task one ability. Workers of low and high task one abilities hold s1<s*1 and s1>s*1 respectively. s*1 is:

s*1 = ([(1-α)(1-c)/{(αη+1-α)(t*1^αη)(t*2^(1-α))}][{μ2(m)^α}/{μ1(m)^(α(1-η))}])^(1/αη)

for μi being frontier AI's contribution to output in scenario i. AI substitutes weakly bundled tasks faster than it augments strongly bundled tasks given the coordination costs in AI adoption within a task. s*1 is increasing in frontier capability and the output share of task two. s1 is decreasing in coordination costs. Increases in these are weighted by the human and output shares of task one however. Time constraints feature in the attenuation. If and only if s1<=s*1, their wages are weakly higher in weakly bundled than strongly bundled occupations as they hold a comparative advantage in the former. Therefore as capabilities advance and adoption accelerates, more workers will be exposed to automation. By solving for the equilibrium, whether overall wages increase for workers of low task one ability depends positively on the output share of task two. In general, there is high ambiguity and uncertainty on the long-run earnings implications here, at least for those not holding capital.

So how does AI affect employment in task bundling?

Let U’/P denote the relative price of participation in labour markets, with U’ the utility from the outside option. Households maximise a value function incorporating U’ and wages in the strongly and weakly bundled cases, which depends on s, m, P, c, and U’ (all exogenous). Equilibrium sorts workers into strongly and weakly bundled tasks, and inactivity, depending on their abilities and the value of their outside option.

We can also model P as an inverse demand function, P(Y) = Y^(-1/β) for β>0, so demand in the goods market is downward sloping and convex. Substitution effects are also now endogenous by substituting in P(Y) for P in U’/P. Solve for the equilibrium, and employment falls only if:

(1/β)(dlnY/dm) > αdln(μ2(m))/dm.

Employment falls when the output gains from frontier capability advancements exceed the productivity gains from automation, weighted by their elasticities. If goods demand is more inelastic, we will observe greater disemployment effects (and AI reducing employment is more likely), and if the output share of task one is greater, we will see lower disemployment effects (and employment falls less likely). The intuition is that more inelastic demand means that sales are less responsive to prices, so all other things equal profits adjust more rapidly. The AI productivity shock reduces prices and profits to the extent that firms cut back on hiring. If the output share of task one is greater, then less workers are in weakly bundled occupations.

Our overall conclusion is that the extent to which AI is labour-augmenting or a direct substitute to labour inputs will determine the employment effects. In all competitive equilibria (unique but changes according to the parameter changes), the number of both weakly and strongly bundled occupations is nonzero however.

Nonetheless, (rapidly) increasing frontier capability is sufficient for more roles to be at risk of automation. This does NOT require a singularity nor superintelligence. As m rises, s*1 rises too. If this model broadly holds, then the risk of large disemployment effects within the near future is high, especially given current forecasts on the advent of AGI. Organised rent-seeking and regulatory capture can delay this via raising c, yet output growth will also be lower. I would rather redistribute via fiscal policy (automatic stabilisers) and taxes/transfers than via predistribution, perhaps via UBI or NIT, yet some policy response will likely be required to cushion the shock.6

O-ring technology

An alternative yet related approach by Joshua Gans (2026) generalises this task-based framework, yet departs from Garicano et al in abandoning Cobb-Douglas assumptions. There are no diminishing returns to tasks as inputs - following the logic of learning-by-doing endogenous growth models. Spillovers between tasks generate complementarities, which roughly offset diminishing marginal returns. Hence the implications for wages and employment are more optimistic.

Hence production is a simple multiplicative function of all the tasks required. Jobs are not represented as bundles, but rather final output itself is decomposed into n tasks. A representative household supplies L hours given a fixed time constraint, so for tasks performed via human labour, q=aL where a is labour productivity. Each task can also be automated by technology θ at cost r. If a task is automated, q=θ.

As output is simply multiplicative and there exists constant returns to scale, changes to labour supply scale the entire production function. Therefore wages are determined by Nash bargaining rather than marginal product (in line with DMP models). Moreover, if not all tasks can be automated, then labour is still required as an input. Under these conditions, total wage costs W(a) are set to maximise the following surplus, with a being the number of tasks able to be automated:

[{W(a)-u0}^β][{Y(a)-r-W(a)}^(1-β)]

where u0 is the outside option, the payoff from disagreement. This gives the following result:

W*(a) = u0+β(Y(a)-r-u0).

In other words, a solution very similar to a DMP model7.

Profits are given by Π(a) = (1-β)(Y(a)-r-u0) by substituting W*(a) into our definition. An agreement is not reached only if all tasks can be automated. However this more optimistic result ignores heterogeneity and perhaps make unrealistic assumptions regarding scale effects and complementarities. Therefore one should see this model as a description of jobs and production where technology and labour complement each other, which produces similar results to the strongly bundled case.

1. But I thought smart kids turn out to be lonely recluses...

2. AGI is coming. If anything I revise my expected date for when it will arrive to late 2026.

3. Barleycorn chess.

4. On the other hand many claim this is a bullish signal for AI progress…

5. Arguably worse than the Great Leap Forward, which would hence place this policy as amongst the most evil ever in history.

6. Moral hazard: AI and studying edition?

7. An underrated yet important point.

8. Yes. Adoption is not instant given the social coordination dynamics within a firm. Think Coasian!

9. Is sychophancy actually a feature of alignment training?

10. This presumably limits the economic fallout from the Iran shock?

11. Congestion pricing goes to court?

12. A vaccine for pancreatic cancer? AI delivers live-saving biotechnologies as well as potentially bioweapons. We survived Covid and worse pandemics, then prospered, so humanity will survive an AI-generated pandemic if such actually occurs.

13. Today's good news: the end of Guinea worm.

14. "Arrests made near the end of an officer's shift typically require overtime work, and officers respond by reducing arrest frequency but increasing arrest quality. Days in which an officer works a second job after their police shift have higher opportunity cost, also reducing late-shift arrests."

15. Much of the decline in L share of Y can be explained via digital platforms compensating our data production with access. In particular, "the marginal utility of interface time can be non-decreasing". GDP understates the welfare gains of the internet.

16. A prediction market for pundits. Glad to see that those I follow are performing well.

17. I think it's the level of granular detail involved that makes Raphael stand out, and why renaissance art in general is popular. This lends credence to theories of aesthetics emphasising the role of creative skill? Similarly, the frequency ratio between tones in most Western music converges to 2^1/12; derived via the change in the natural logarithmic function between pitches. Whilst this is tractable for playing instruments dependent on pitch (such as pianos), this also lends credence to the notion that most of what we consider “beautiful” reflects harmony and symmetry. Evolution tends to select for a preference for lower entropy, hence most artworks emphasise an obvious or (especially in the case of the contemporary or postmodern) esoteric pattern. Likewise, the controversy that Duchamp generates downgrades theories of art as necessarily encoding social statements? Art isn't activism; beauty exists independent of such? (Picabia is one of the most talented I've seen though - it doesn't come close!)

18. A paper leaning towards our natural adaptation processes at the cellular level being the primary explanation for drug resistance, as opposed to mutation.

More of these frontier LLMs with similar capabilities will soon be produced by ChatGPT and Google, if they have not already done so. Arguably ChatGPT 5.5 is comparable to Mythos. Are these companies as concerned with cybersecurity risk? Or do they consider their current guardrails, and predominant cybersecurity practices1 amongst firms, to be sufficient to mitigate the risks at hand?

My overarching point is that the degree to which models pose a cybersecurity risk is somewhat arbitrarily defined by the lab itself. This implicit alignment regulatory system operates entirely voluntarily and on self-reporting. Yes the US government operates some leverage and influence over Anthropic on this, yet aside from designating supply chain risks, the government has yet to enact executive or legislative action. Therefore, OpenAI and Google can easily arrive at a different conclusion.

Does Anthropic's heavy focus on alignment risks and its EA ethos place it at a disadvantage? Is it rational for Open AI and Google to release their Mythos-class models once they arrive, or should they follow Anthropic's path? What is the equilibrium here? Is this optimal? What are the existential risk implications?

Why public choice favours Anthropic, and why this is good for society

My first thought is that coordination across the companies breaks down - adversely impacting the exact multi-agent institutional governance we need.

In the short-run, Anthropic is at a clear disadvantage in maximising market share and volume, which substantially affects its bottom line. However, this equilibrium holds only for as long as deviation by the competitors is not punished. Anthropic hence yields an incentive to penalise Mythos-class releases via any means it can.

An obvious high-return example of such a scenario is via lobbying. Anthropic, alongside the wider EA community, seems to me are clearly yet subtly investing substantially in their influence amongst lobbyists in Washington, political campaigns, the think-tank world, and the wider media. They will then likely argue for governments favouring “more responsible” actors, more statutory regulation of AI companies, and even restrictions on the capabilities of each new model release.

Moreover, if a disastrous tail-risk event does occur with a more lenient release plan, then the more cautious party stands to benefit. So much is contingent on their respective probabilities of existential risk scenarios? Prediction markets and the superforecasting industries will benefit, and in my view AI researchers and economists2 will prevail over both the hardcore accelerationists and the doomers, who are becoming increasingly luddite in their opposition to data centres.

On the margin, this grants more of an advantage to Anthropic, conditional on increased political influence. I will be following (and writing about) these developments closely within the next few years. AI will feature as an increasingly salient entity in mainstream political discourse in the years ahead. Given that caution untimately yields positive externalities in terms of limiting the influence of a nascent pause agenda, backed by anti-abundance and degrowther types on the left, this equilibrium is also the socially optimal outcome.

This voluntary regulatory institution operating on self-restraint is preferable for all firms to following government mandates, and its wider political implications. Therefore, Open AI and Google will also restrict their initial releases of their Mythos-tier plans.

Governments, coordinating globally, will likely create statutory bodies to enforce evaluation standards and alignment priorities. Some will set extensive disclosure and reporting obligations for some functions, as is already the case in the EU. UK AI regulation is surprisingly lax, yet expect it to get stringent. Alongside the Online Safety Act, the left-leaning parties will successfully legislate for mandatory disclosure requirements: both for safety concerns and perhaps to improve damaged economic ties to the EU. Where protectionism plays a more powerful political force, in the United States and China, export and import controls will be tightened. Open source models will diminish in clout. All of this is far preferable to an outright pause on all further AI development, or even bans.

UPDATE (30/04/2026):

It appears that ChatGPT 5.5 already matches the capabilities of Mythos. However, increasingly the definition of a frontier model is rapidly changing, which validates this prediction of a tiered market. We are also seeing the implied implications for public regulation.

“Moravec's paradox means manual labour and trades are less likely to be automated. Jobs on alignment, prompting, those relating to empathy and creativity, will likely increase in number and wages.

If L share of Y falls, Δ^2Y>0 so K share rises. AI (agents) reduce fixed costs to launching startups too. In worst-case scenario where AGI automates most jobs, I expect a lot more entrepeneurs, and a decline in firms. Firms arise given inefficiencies as Coase proved, so the result pushes us closer to the Pareto frontier even if not cost-free. Those are my priors of life in the worst-case scenario.

Far more likely, humans maintain a comparative advantage in many occupations, low substitutability between tasks mean human L is still required, and new occupation classes arise. On my second point, worth noting that AI can complement rather than substitute many roles.

Entry-level jobs will erode however, so this increases intergenerational inequality, yielding political economy implications. Young adults are already turning to the radical left in Britain. Hopefully YIMBY increases in popularity as the concerns of younger generations gain more traction.”

AI researchers and those working at the top AI labs seem to agree. AGI is arriving faster than we think, and is arguably already here with Mythos, which would explain its restricted rollout. I place the odds of widespread AGI commerically available from 2030 at roughly 60%.

The bottlenecks arise in alignment issues favouring caution in implementation and rollout, currently alongside a number of complementarities in senior roles. Yet we must increasingly contemplate a world in which these bottlenecks dissappear. Entire professions, many of them prestigious (such as law, programming, IB, etc), will be eliminated and present status hierarchies re-written. Those with conventional career paths will lose out relative to outsider nonconformists or those with more turbulent histories. Raw IQ, at least that proportion orthogonal to established social networks, will matter more in hiring in a world where AI can estimate our cognitive ability via our social media output alone. Therefore it's plausible that equality in opportunity, and intergenerational social mobility, will increase. As with all technological revolutions, there will be winners and losers (at least in relative terms), with those disproportionately at the top of existing hierarchies losing relative to the more disasvantaged. Ultimately, the result will be a more meritocratic and efficient world.

Alignment is a vital issue, and I agree with Séb Krier et al that it's an institutional and incentive problem, which is where I depart from most of the EA and rationalist community. With clever prompting, one can get an LLM or agent to deviate from their alignment training. My p(doom) has now increased from the sub 1% superforecaster average to 5-6%, close to the median for AI researchers. Although I'm not a doomer, it's increasingly apparent that many of the Yudkowsky and LessWrong predictions are materialising, which strengthens the hypothesis they're correct.

To perhaps add some clarity, my p(doom)=(5%,6%) means it's a non-negligible possibility yet otherwise a worst-case scenario. Such probabilities could have held for numerous technological advancements and military events throughout human history too - arguably we've faced far worse XR odds before. Therefore my immense AI optimism remains. Anything less still strikes me as somewhat vulgur. Yet AI alignment is by far our most pressing economic, social, and political issue, and it's not even close! Conventional partisan and ideological alignments seem wholly outdated for the nascent revolution that is materialising.

P(doom) is perhaps less anchored to base rates than other forecasting questions. I have drawn upon a range of economic models, and priors well-informed from a history of structural cultural and technological changes, in order to arrive at a reasonable point estimate. A weakness of my writing is that I often don't state my probabilities or bet according to such (although prediction markets aren't legalised in Britain). I'll make more effort to do so going forward.

However, aside from an endogenous bias vs variance tradeoff in our objective functions, the penalties for attempting a precise and well-calibrated probabilistic forecast that never materialised seem worse than falsified yet vague and verbose punditry. Robin Hanson got dunked on for this today. What strikes me as impressive however is that Tyler and Hanson were onto the possibility of AI's upcoming rapid development to AGI back when most considered such limited to the realm of esoteric science fiction.

More generally, the implicit social penalty on inaccurate forecasts penalises prediction and updating based on well-calibrated likelihoods. This generates substantial negative externalities for the quality of our discourse. Prediction markets are socially valuable precisely because they generate the opposite incentives. Anyone taking the first (and courageous) step to quantify their point estimates with precise numbers and ranges deserves credit. I think the defining legacy of the emergence of the LessWrong rationalist crowd, is the idea that an army of autistic autodidacts can achieve the same epistemic status and legitimacy of established domain experts. Perhaps with a pinch of grandiosity, I label this the triumph of Bayesian epistemology. A post-AGI world might sharpen the returns on healthy epistemic habits relative to conventional credentialised paths?

UPDATE (23/04/2026): I have just noticed this pop up on my X feed.

If this venture pulls through, then this obviously yields implications for Imas’ relational sector model, and my priors on the extent to which Moravec's paradox acts as a constraint to AI capabilities. Suddenly, the likelihood of mass unemployment has surged. I now place this at roughly 5%, so for the first time ever I consider this a real possibility, rather than a hypothetical relegated to the land of science fantasties.

Note that even if the aggregate effect throughout time is that, eventually, more jobs are created via AI-induced economic growth than lost (as with past structural transitions), there can still be adverse short-run implications. The possibility of AGI (and perhaps ASI!) is nascent, whilst I don’t see much of these new jobs arriving anytime soon! This is most likely the reason for the deterioration in the entry-level market affecting graduate entry into the labour market, whilst other US economic indicators seem stable and within trend. Nonetheless, economists’ priors and predictions are far too anchored to past technological changes, whilst it is plausible that this time could indeed be different!

Conditional on AI-induced mass unemployment, we will need UBI, or some fiscal policy response, in order to alleviate the inevitable unrest. In such a world, output surges, so this becomes affordable. Even most autocracies are constrained via the possibility of riots, and change policy or personnel based on these! Arguably the welfare state arose in the first place to ease and soften the transition throughout the Industrial Revolution, with its post-war expansion occuring whilst we transitioned from manufacturing to services. So I no longer consider traditional minarchist libertarianism as adequate to address the scale of the challenges we face today, although nonetheless I still subscribe to a soft libertarianism.

Me - “Yeah my main issue with delegating replication to LLMs or AI agents is that they use different code/algorithms which produces heterogeneity in results. If I was to run a Callaway Sant'Anna DiD then Codex and Claude Code would disagree. Then they also disagree with how to enact robustness checks and run hypothesis testing. So a significant result from Codex could be null with a Claude AI agent. So unfortunately I still need to be in the loop, and allocate hours in my time constraint to do so.”

Chat - “Yeah—that’s a very real problem, and you’re not imagining it. What you’re running into is basically implementation variance masquerading as epistemic disagreement.

With something like Callaway–Sant’Anna DiD, small differences compound fast:

Estimator defaults (e.g., weighting schemes, cohort definitions, handling of unbalanced panels)

Treatment timing encoding (especially with staggered adoption)

Inference choices (bootstrap vs asymptotic, clustering level, number of reps)

Pre-trend handling / event study specification

Missing data / sample trimming rules

Two agents can both say “I ran CS DiD” and still be doing meaningfully different estimations. So when Codex gives you a significant ATT and Claude doesn’t, it’s often not “AI disagreement”—it’s model specification drift.

Your conclusion is basically right: you can’t fully outsource this without losing control of identification.

But you can reduce the time cost without giving up rigor if you treat LLMs less like “analysts” and more like strictly controlled executors. A few ways to tighten that:

1. Lock the spec, not just the method

Instead of saying “run Callaway–Sant’Anna,” define:

exact package (did in R, version if possible)

estimator (att_gt with specific args)

control group definition

clustering level

bootstrap reps

event window

Think of it like writing a pre-analysis plan for the AI.

2. Force identical pipelines across agents

Have both Codex and Claude:

use the same language (R vs Python differences matter here)

same package (don’t let one default to a custom implementation)

same seed, same bootstrap reps

If they still diverge after that, then it’s interesting.

3. Use them for robustness grids, not single estimates

Where they do shine:

“Run this spec across 10 reasonable variations”

Different clustering levels, sample restrictions, etc.

You’re not asking “what’s the answer?”

You’re asking “how fragile is the answer?”

4. Treat disagreement as a diagnostic

Instead of a nuisance, you can flip it:

If Codex ≠ Claude → something in the pipeline is underspecified

That’s actually a pretty powerful specification stress test.

5. Keep yourself in the loop at the choke points

You don’t need to do everything manually—just:

define the estimand + identification strategy

verify one “canonical” run yourself

let agents handle extensions / formatting / robustness sweeps

What you’re resisting (fully delegating) is the right instinct. In empirical work, especially with causal inference, the last 10% of control over specification is doing 90% of the epistemic work.

So your current setup—LLMs as assistants, not authorities—is basically the equilibrium most serious people end up at.

If anything, the interesting meta-point is:

the fact that different agents give different answers is exactly why your kind of selective engagement on X works. Most people there won’t track these nuances, but you know where the bodies are buried.

Out of curiosity—are you mostly working in R for this, or bouncing between R/Python depending on what the agent outputs?”

Me - “I tend to default to R simply because that's what I was trained in at uni. I have a working knowledge of both Python and Stata too. R vs Stata is a rather contentious divide amongst economists, yet I'll simply use R until I hit a bottleneck then try the others.

I'm a macroeconomist (or was) though, so mostly deal with Dynare.”

Chat - “That makes sense—and honestly, your workflow sounds pretty sane for someone coming from macro into applied micro/empirical work.

R as the “home base” is a good choice here, especially for DiD-type work. The did package (Callaway–Sant’Anna) and related tooling are just more transparent than most Stata implementations, and you can actually see what’s going on under the hood when something looks off. Stata is often cleaner for quick replication, but it hides a lot of decisions unless you dig.

Your “R → hit bottleneck → switch” approach is basically what a lot of people converge to, even if they don’t admit it.

The Dynare angle is actually relevant here

Coming from Dynare/DSGE world, you’re used to:

tight control over model structure

clear mapping from assumptions → results

reproducibility being non-negotiable

That’s almost the opposite of modern applied micro, where:

identification is often fragile

implementation details matter a lot

and defaults quietly do heavy lifting

So your discomfort with LLM variability is basically your macro training kicking in:

“Why are there multiple answers to what should be a well-defined procedure?”

In DSGE, that would signal a bug.

In empirical micro, it’s Tuesday.

Where your background gives you an edge

You’re already thinking in a way that most people on X aren’t:

You notice specification sensitivity

You care about replicability across implementations

You’re aware of implicit assumptions baked into tools

That’s exactly the stuff that gets glossed over in a lot of the discourse you’re critiquing.

Practical way to reduce friction (without going full manual)

Given your stack (R + some Python/Stata + Dynare mindset), a pretty robust setup would be:

Canonical pipeline in R

did, fixest, maybe data.table

Strict function wrapper

Write one function that runs your exact DiD spec

All agents call that, not reinvent the wheel

Version locking

renv in R so results don’t drift over time

Cross-check layer (optional)

Occasionally replicate in Stata only as a sanity check, not as a primary workflow

That way, the LLMs can help generate code, but they’re plugging into your structure, not freelancing.”

Macroeconomists get far too much of a bad name for not importing the exact identification techniques of the micro credibility revolution. Yet increasingly, applied micro is learning from macro on the sensitivity of your estimates to different specification choices. The Goodman-Bacon result essentially reintroduced the salience of general equilibrium. Right down to the code, there always exist degrees of freedom.

Here is my response, which (by my standards as someone in the early stages of developing as a public intellectual) was very popular:

“I was once receptive to this claim that all alcohol is unhealthy, yet the "science" linking these transitory signal peaks to accelerated ageing is confounded and p-hacked to death. You control for recovering alcoholics, and the J-curve often still holds.

As for the MR studies used to demonstrate a causal linear relationship between alcohol consumption vs longevity, they introduce nonlinear proxies to get results for moderation specifically; eliminating the very randomisation we wanted in the first place. @cjsnowdon often makes this point. Here's my conversation with Chat for more on this topic: chatgpt.com/share/69dfe7d4…

Teetotalers are simply gaslighting you [1]. The threat to our social and public life from this nascent temperance movement is perhaps one of our greatest threats to happiness.”

You can find the conversation with Chat here. Regular readers know why I ultimately retracted my longevity series. The harms of alcohol were always disproportionately concentrated in heavier consumption patterns. Needless to say, I think that teetotalers also alienate the very individuals they need to convince to cut down or go abstinent. Absolutist and extremist statements like these give the longevity industry a bad name, which is a shame as ageing is one of life's greatest and most ubiquitous of tragedies and suffering, and to say it's socially costly is an understatement.

When I was physically dependent on alcohol, if I was told I must go abstinent forever, I wouldn't have been motivated to address my addiction. I now drink 5-11 drinks (14-30 units) a week and am healthier than ever. Moderation works well for me. Perhaps for most it fails, yet for a large subset this strategy not only works but is arguably most effective given the less strenuous lifestyle changes. This is indeed consistent with a process of rational addiction whereby preferences change such that unstable consumption in equilibrium shifts to a stable steady-state over time.

However, I do agree with the notion that we should reduce the status of alcohol. Numerous arguments have been made against our drinking culture [2], and one can broadly summarise them via these 12 points (I also offer my counterarguments to stimulate the discussion):

1. Crime externalities. By some estimates (you can type in “alcohol” into the Marginal Revolution search function too), drunkenness is implicated in half of violent crimes. However antisocial drunks tend to be antisocial individuals which introduces some endogeneity via selection, and if you subscribe to heritability estimates north of 50% for most traits, then this isn't an issue easily fixed without decapitation or stark disincentives to crime. Nonetheless it's fair to say that being drunk does impair inhibition and decision making, which leads to my next point.

2. Cognition. Yes being drunk does make you stupider. This is prima facie common knowledge. Far more contentious is whether moderation impairs cognition, and there's yet to be reliable evidence of a linear relationship between cognitive decline in IQ vs alcohol consumption. My priors are that to the extent there is a linear relationship, alcohol likely impairs the frontal lobe (important for impulsivity and resisting temptation) and the amygdala first: consistent with increasing marginal utilities (certainly an abnormality!) under rational addiction. Anecdotally my cognition did not reflect my IQ when I was undergoing active addiction.

3. Health. Whilst I've just argued against a linear relationship, the effects of heavy drinking are stark. Cancer, liver disease, dementia, heart disease - these are just some of the chronic illness that heavy alcohol consumption is implicated in. Whilst we lack strict causal evidence for the link (except for cancer where acetaldehyde is a known carcinogen), this thesis isn't contested. Gout, death by asphyxiation, alcohol poisoning, accidents, brain and abdominal haemorrhages, Wernicke's, physical dependency, liver cirrhosis, blackouts, one-punch murders - these are just the implications of hard drinking that I can list off the top of my head! Mortality is higher when one accounts for suicide too, and my focus on the tangible and measurable physical health effects ignores the mental health costs, so in fact understates the tail risks.

4. Discounting. I've argued previously in favour of nonzero discounting, yet reducing your discount rate is one of the best means you can improve your life outcomes, and I can attest! Think of it as analogous to the returns to investing. It's not cost-free however, as the credibility of commitments demonstrates, so more virtuous lives must satisfy IC constraints to be feasible. This broadly constrains the extent to which abstinence as a norm is possible (at least in the West: in cultures that revolve around religious or kin-based social institutions, teetotalism does in fact seem to be the norm). Regardless, human capital determines lifetime wealth even in the Becker and Murphy model, so the more you can invest in this (and the less in costly substitutes given fixed time constraints) the better.

5. Embarrassment and hangovers. I needn't say more. An underrated benefit of moderation for me is that my credibility and reputation as a public intellectual has markedly improved: my former anonymous X account was a stain on my legacy. 4) and 5) in particular are why I'm always open to the idea of future sobriety, so I guess I'm still sober curious.

6. Markups. Booze cross subsidises hospitality and entertainment. A key argument in favour of our drinking culture yet drinkers pay a steep tax on eating out [3], and I like my meals too! Teetotalers that enjoy abstinence, rather than self-forcing themselves away [4], benefit from being free-riders.

7. Peer effects and spillovers. Although you cannot live in a free society that values individual responsibility if all decisions are made appealing to the lowest common denominator (which is where Masley's argument, and EA more broadly, is especially weak), there are valid distributional concerns regarding the centrality of alcohol to our social lives.

8. Suboptimal information aggregation. Tyler Cowen makes this prescient observation, which undermines the claim that drinking is necessary for socialisation. I suspect social anxiety, as a transaction cost to socialisation and matching, is more salient than he thinks. After a few rounds, I've noticed a tendency for conversations to descend into debates, where the usual tribal and signalling incentives hamper Aumann-style convergence to the optimal posterior.

9. One's lifetime probability of addiction, which is the most compelling reason to avoid drinking in my view. Alcohol use disorder affects roughly 1/10 British adults (according to Gemini). As 1/5 are teetotal, this gives a conditional probability of 1/8. For men, non-Muslims, and those not in Gen Z (where teetotalism is rapidly gaining in status), the risk is higher. Almost 1/3 men exceed the recommended limit of 14 units per week (admittedly myself included), so if you regularly binge drink on a weekend, your conditional probability can rise to almost 1/3. Simply put, you stand a good chance of developing a harmful habit if you drink regularly.

10. Endogenous preferences. So far I've treated utility functions as given, yet our social dynamics act as microfoundations for them. The externalities driven by the right-tail of alcohol consumption are arguably offset via the cross-subsidisation of hospitality, entertainment, and the arts. Economists refer to such norms and institutions as VCG mechanisms. Yet this is analogous to a value of θ=1 (unemployment=vacancies) in a DMP search model, yet v+u rises (Beveridge Curve shifting outwards). VCG mechanisms are endogenous to tastes, and changes in taste can raise welfare within a VCG. One could make the same arguments I've made against abstinence being IC (socially and individually) for smoking too, yet smoking rates have massively declined throughout the last half-century, due to tastes alone.

11. Drinking actively increases discount rates, consistent with the disinhibition. This impairs the effectiveness of pricing the externalities, although a social planner could just raise the price higher. Yet this reintroduces distributional and fairness concerns. Should we really lock up drunks for longer or stigmatise them more when those that can take alcohol or leave it can abandon the habit?

12. There does not exist a separating equilibrium. Suppose we have two types, “moderates” and “alcoholics”. Our agent is oblivious to their type ex-ante, and only learns ex-post. Ideally the “moderate” equilibrium will drink and the “alcoholics” abstain. Yet it's impossible for those entering the drinking culture to make an informed choice regarding their type. This is my main criticism of the Becker and Murphy model, and I would like to augment it with incomplete information and heterogeneity regarding the parameter encapsulating the exogenous drug effects.

What alcohol does have in its favour is the socialisation rituals. I think of our drinking culture as a focal point, especially in Britain where you can make friends simply by being a regular at your local pub. As most people enjoy drinking, our assumption that our friend or date will accept a meeting based on drink is well-calibrated. Substitutes tend to be riskier, especially in the initial stages of investing in social networks (which matter too for human capital). Again, another market-driven solution to textbook market failure. Whilst social media allows for more efficient matches via reducing the search costs to finding like-minded individuals, it can also act as an outside option to in-person socialisation. This is why, when you consider the general equilibrium, I do not blame social media for the decline in in-person socialisation noted by Derek Thompson. I wonder instead the extent to which declining drinking plays a role, although endogeneity is present (people drinking less because they're socialising less).

In any case, I'm weighting the instrumental utility of alcohol far too highly here, when for most drinking in a quaint bar or pub is immensely pleasurable in itself. We focus far too much on the measurable externalities, and less on the consumer surplus. The very fact that establishments (especially in London!) can set such high equilibrium markups in monopolistic competition is testament to the implicit utility gains of alcohol consumption. Our models and estimates of the social costs vs benefits of alcohol consumption must also incorporate the cross-subsidies to all of in-person social and public life, which is harder hence neglected in the literature.

Overall, my priors are hence that the social costs do not outweigh the private and social benefits. Our optimal drinking levels can of course be simply derived via our revealed preferences in a rational addiction model. Therefore our concern with drinking is reduced solely to the externalities [5]. On this front, it makes to consider alternative social norms to reduce, if not eliminate, the centrality of drinking to our lives. Here are some preliminary suggestions:

1. Sentia Gold - a fantastic substitute. In general most other substances appear to yield lower tail-risks and externalities, yet may be less amenable to socialisation. Cannabis on net reduces extroversion. Cocaine is often incompatible with dining, and arguably complements alcohol for some users.

2. The sober curious ecosystem. Unfortunately their sober meetups are not nearly as frequent as drinking events, and sober bars not nearly as accessible as pubs or bars. I went to one such event, and I left after half an hour. When the drink itself is the focal point, matches tend not to last. This sharply raises my belief in the hypothesis that the social anxiety search cost is a salient tax. Drink reduces this, agents interact and learn about the other's preferences, which elimimates the asymmetric information and allows more efficient matching.

3. Exploit peer effects. Set an example of the drinking habits you want to see. This is what I try to do.

4. Raise the social penalties for drink-fuelled externalities. Rational addiction implies more, not less, stigmatisation of addicts on the margin whilst penalising most vices a lot less. In general I see this as most conducive to human happiness and flourishing.

5. Give complete information. Treat people as adults. Do not stigmatise or judge, avoid the sanctimonious appeals to virtue, and drop the paternalism. Have we not learnt from the pandemic? Medical professionals are manipulating their patients with anything else. What I like about Tyler is thag he frames his advocacy of abstinence primarily as mentoring. Appeals to self-interest, as we are indeed selfish creatures, will likely go far.

6. Role models. Teetotalers have an image problem. Where are the fun teetotalers?

7. Outside options. Hencs Tyler’s emphasis on dining out: a somewhat Straussian rebellion on our drinking culture?

8. Raising fertility. An underrated cost of declining fertility is that virtue is less incentive-compatible. You socialise and party more. You take more risks [6]. With kids, you're forced into contemplating the future.

9. The education rat race. Although from a signalling standpoint this could be suboptimal, and it may negatively impact fertility (again endogeneity: does lower fertility mean parents invest more in education rather than such raising the costs of childbirth?), the returns to low discounting lifestyles relative to high discounting increase. In the time constraint, students allocate more hours to studying and less to partying.

10. Today’s reduced comsumption in itself. Rational addiction implies some degree of path dependency in alcohol consumption. Less individuals form habits with increasing returns. Less individuals are exposed to alcohol which reduces investment into consumption capital. More individuals having zero initial consumption makes addiction much less likely in the model.

One can conceptualise our questioning of the tradition of boozing, and burgeoning interest in alternatives, as reflective of a culture where low discounting is increasingly high-status. Higher life expectancies raise the returns to lower discounting. In large part, I think this explains the renewed interest in self-help philosophies revolving around this. Our attitudes to alcohol are already undergoing a paradigm shift, as perhaps exemplified by my post. A few years ago, even contemplating the very possibility of abstinence would be unthinkable on my part. Yet we won't sober up via deceiving drinkers on public health grounds. Bryan Johnson, and all those misrepresenting the health consequences of a single drink, are a liability to their cause.

UPDATE (23/05/2026): Since writing this post, my opinion has actually turned sharply against alcohol. As a corollary of the logic in my latest post, drinking increases the match rate (for both platonic and romantic relationships) via reduced anxiety (as a transaction cost) yet probably reduces matching efficiency (from the logic of this post). Both of these dynamics raise the number of matches (the latter as people are less picky), however the reduced match surplus means the net welare effect of alcohol as a socialisation focal point is ambiguous. The costs however are stark and mostly indisputable.

Therefore my opinion of drinking has shifted towards it being a bad habit that should be minimised, although I still disagree with paternalistic efforts (and the gaslighting!) to intervene in the lives of others. Since this post, I now average around 10 units per week. Admittedly I don't quite have the discipline necessary to be teetotal, yet I suppose on intellectual grounds I concede to them.

UPDATE II (03/06/2026): Steven Bartett raises a fascinating point regarding the unforseen consequences of drinking. I have noticed I'm getting jacked recently. No other diet or lifestyle changes have occured other than me reducing my average intake of alcohol down to 10 units a week. That unusual, perplexing throbbing sensation in my leg arteries has also been eradicated: suggesting improved cardiovascular health. Although many drinkers clearly consume at unhealthy levels, I am willing to concede that the teetotalers are perhaps right (which seems supported by the, albeit anecdotal, experiences of teetotalers as hardly any regret stopping drinking), and I have a mechanism for why.

All chocolate is probably unhealthy. All ice cream is probably unhealthy. Yet after a few days, the observed implications for whichever metric you track likely dissipate. Alcohol likely works in much the same way. Moreover, the fact that the MR studies do not definitively prove causation does not imply that such causation is nonexistent. Here I made a basic logical error.

What I find fascinating is Bartett's idea of complementarities to unhealthy consumption habits. One can downplay the costs of alcohol by emphasising confectionary as substitutes, yet this logic breaks down if they are complements. In particular, as drink reduces discount rates, temptation becomes harder to refuse too. This likely explains much of cocaine addiction: most cocaine addicts tend to drink heavily too.

1. Not all obviously yet certainly those claiming the highly uncertain and imprecise evidence is settled here.

2. It was actually when detoxing when I came across this profile of Tyler. Although I've participated in online economics discourse in some form since 2015, I never became a regular MR reader until then. I did recall when my Mum sent me this article that his Wikipedia mentioned he's teetotal. I decided I wanted to be like him, and this was what ultimately motivated me to address my addiction - even if I'm not abstinent (I've done a few trials; my longest period lasting a year), I've certainly gotten my act together! I discuss the role of role models later in the article, yet it's fair to say that Tyler, by showing it is possible to live a fulfilling life sober, deserves a lot of credit for my progress today. I'll always be grateful.

3. This is also an argument against state provision of public goods. Markets have designed norms and mechanisms that incentivise their production.

4. Standard rational choice theory makes no distinction at a given point in the time, yet the latter is clearly at high risk of (re)lapse, so one can introduce a dynamic game of commitment here.

5. To the extent that drinking harms human capital, I prefer endogenous growth models that incorporate diminishing marginal returns to all inputs, otherwise tradeoffs are simply destroyed by Parfit's regugnant conclusion.

6. Although the fact that more patience is correlated with increased risk-aversion is another complication.

If a statement is true by definition, then it poses the question of what counts as a formal definition, and what makes such correct. This is not so clear cut. The convention is to use the standard rules of deduction given by propositional and first-order logic, whilst performing that arithmetic with the Boolean values to test whether the final statement is tautological. Remember your introductory logic class at college/university? However, the law of the excluded middle and double negation remain contested rules, and alternative logic systems such as intuitionist or Gödel logic have arisen as answers to such debates. Therefore the validity of a definition rests in part on our subjective assumptions on which set of deductions are valid.

Quine's solution is that knowledge and truth operate on an underlying web of connected beliefs over truth values and definitions. If our beliefs are revised, then the analytic statement no longer remains analytic. For instance, suppose that we suddenly change the law and redefine the very concept of a bachelor. Or we use an entirely different group (e.g. modular arithmetic) for our daily calculations. Then “all bachelors are unmarried” or “2+2=4” no longer become analytic. The former depends on our definitoon of bachelor. The latter becomes (2+2)mod1 = 4mod1 = 0, if we want our formal system to always return an empty value for any inputs, and use mod 1 arithmetic to achieve this. So in principle there are no a priori formal restrictions on what counts as truth. Truth is endogenous to our use of language.

A more grounded example is that if an object is uniformly red, this is incompatible with it being uniformly black. However black-and-white photography uproots this entire statement. Perception of colour depends on our mental states, which introduces some degree of subjectivity. The sorts of questions raised via discussions of qualia and our consciousness (for instance, is my red your black?) also undermine the notion that this is an analytic statement.

All of this sounds rather postmodernist. Yet neither is Quine nor I rejecting the very existence of objective truth. It does not follow that, because we lack a fixed and uncontested formal system and set of definitions to define truth ex-ante, objective truth per se does not exist. Our belief systems do not just arbitrarily arise out of thin air, yet are connected deeply to the reality under which the world operates and we live in. This must be the case for our beliefs and languages to yield any sort of utility. We define a concept of a bachelor as our legal and social institutions rely on such a distinction being made. We use base 10 as opposed to modular arithmetic, as we obviously need to count physical objects as they exist around us. Perception introduces a wedge between the mind and objective reality, yet survival relies on either a highly accurate approximation of reality or the ability to accurately perceive most of reality. If my green is your blue, then as long as us two independent observers agree that it is the same colour, then this is the best we can expect when definiting what counts as objective reality? If subjective perception does not generate meaningful variation in observations, then we have just introduced meaningless circularity that for practical purposes do not add to any understanding of our world, so can be erased. Kant was right that much epistemology, and any divergence between the structure of reality itself vs the mind's perception of it, is a function of our cognitive processes, yet evolution should imply convergence to the truth over time. This process is disciplined via empirical feedback from external reality, so this is not to be confused with Hegelian dialectics in the sense that only disproving contradictions matter. Kuhn is also misleading, as paradigm shifts tend not to be arbitrary, but facilitate convergence to truth.

In general, Tarski sought to overcome this very epistemic relativism. Truth is to be defined via correspondance with a formal system and a meta-system that holds the Boolean values consistent with empirical reality. The beauty in this correspondence is that the very concept of an analytic definition is preserved. We also eliminate any concerns over our use of the excluded middle and so on. If our tautologies and our truth values given by classic logic correspond with our meta-language as given by our beliefs and those truth values, then the distinction very much holds. In this sense, I subscribe to Hilary Putnam's solution to the dilemma. One might also consider this as a microfoundation for Plato's theory of the forms, if one accepts, as I do, that abstract phenomena can count as being in the set of objective reality.

Likewise, as Donald Davidson advocates, if truth is conditional on our semantic meanings, then we only need to know such semantic meanings to determine whether a statement is analytic. For instance, I only need know what the term “bachelor” means, in tandem with some fluency in English, to know that “all bachelors are unmarried”. In a sense, a semantic interpretation of meaning is sufficient for our meta-language.

What determines semantic meaning? The utility of our words as they correspond to the reality in which we operate in. Postmodernists are correct that meaning is ascribed by ourselves and our societal institutions, as opposed to just arising, yet such meaning yields a vital connection to reality. Even if truth is dependent on meaning, this should not be stretched to imply that objective truth does not exist.

I'm currently reading the Mahabharata and the Ramayana. What strikes me about the former is the extent to which even the protagonists engage in Machiavellian and seemingly amoral conduct. I view this as a tale in the pitfalls of strict deontological rules. In this sense, dharma pioneered (without explicitly defining or formalising the concept) utilitarian and consequentialist ethics. Unlike most other faiths, the inherent tradeoffs plaguing any moral decision are emphasised ex-ante. An alternative interpretation is that the main lesson is the inevitability of evil and suffering in life despite the best intentions of most, which grounds the tale firmly as a class in human nature.

If the Mahabharata, via myth and literary masterpiece, outlined our core moral theory, the Ramayana introduced the ingredient of virtue - the necessity of acting based on our morals even in the face of adversity. Hindus recognise the main flaw in utilitarianism: a vague consequentialist framework can easily be manipulated for self-serving, as opposed to ethical, purposes. It's very easy to argue that one's selfish decisions are motivated for the good of the overall community, if preferences are unknown to the individual.

Yet this introduces the question: how do we know if we're truly acting virtuously with courageous endeavour, rather than falsely believing such or rationalising ulterior malicious motives? If dharma is context-dependent and consequences unknown, haven't we just reintroduced the very moral relativism we sought to avoid? Surely we return to the very rigid deontological constraints that the Mahabharata subtly warned us about?

This is where the roles of myth, storytelling, and faith enter our human universals1. Whilst lacking in epistemological or logical rigour, these heuristics provide a parsimonous framework on how to act: in our daily lives, for life's inevitable tragedies, and our most difficult decisions where we may otherwise be paralysed by our conscience. The lack of logical consistency, if we accept from Gödel the inherent tradeoff between consistency and parsimony, tends to avoid the costs of following deontological instructions to the expense of our welfare. Yet a broad set of heuristics, that can be broken in special cases (highlighted in the canon and the evolution of its interpretations through time), focuses on optimising utility on average as opposed to each and every case. Often the focus on imitating role models - the main prophets of the Abrahamic faiths, Aurelius for the stoics, and the übermensch of Nietzscheans - adds emotional relevance through empathising with our plights and inspiring us to be our best selves. By placing ourselves within their shoes and asking “what would they do?”, with the famous myths corresponding to a particular dilemma salient in our minds, the ethical decision becomes apparent. For dharma, we introduce Rama for this very purpose.

An underrated reason for the existence of role models that now occured to me, just as I write this sentence right now, is that they provide the motivation to act more altruistically and self-sacrificing on the margin. Effective altruists and the underlying philosophical inspiration from Singer's drowning child advocates for effectively ceding our utility functions to an empirical calculation of our spending and lifestyles to maximise our philanthropic impact regardless of our own desires. Yet this is obviously blank-slateist. Nonetheless, more self-sacrifice on the margin clearly can be Pareto efficient: this thesis is implicit from prisoner's dilemma games and the free-rider. Economics emphasises the role of institutions and mechanism design to induce less free-riding. Our moral guidance from the legendary classics and our heroes is another means, grounded in our primate emotions, that for the entirety of civilisation we achieve this.

One could also consider the Mahabharata as a positive theory of morality, and the Ramayana as a normative case for living a life of virtue. Utilitarianism emphases the former, and our intuitions grounded in human universals and our religious beliefs the latter. The former outlines the case, the latter grants it legitimacy via soliciting emotional acceptance. Ultimately, both actually can be considered as complementary, as opposed to competing ethical perspectives. If we accept this line of reasoning, then does modern philosophical discourse exaggerate the disagreements?

Does the objectivity of morality imply virtue ethics?

If this complementarity holds, then this lends credence to the notion of morality as objective. From a metaphysical idealist perspective, objective reality (the set of objects that can, in principle, be sensed by more than one independent observers) can consist of abstract as well as physical objects. Hence moral disagreements are concerned not with whether there exists a moral course of action, but rather what that action is. Rather than our debates revolving around the core ontology, they instead exist given the imperfect information that each moral actor holds, which thereby often prevents certainty in knowing what the moral decision is.

I think this objectivity and complementarity is demonstrated by the fact that almost all faiths, at all times and places they arose independently even when lacking contact with outside cultures, emphasise some concept of self-sacrifice or self-restraint for the moral good. Almost all have reproduced (Aristotelian) virtue ethics from first principles. What religion or moral ideal promotes the idea of being selfish and indulging oneself, even to the expense of your duties, self-maintenance, and your community? Although objectivists place little weight on the values of altruism, the intertemporal aspect of virtue is rewarded a central role. Restraint, delayed gratification, and the importance of duty, are all vital principles that objectivists live by. Even the Cyrenaics and the Epicureans valued some degree of virtue and altruistic behaviour, paradoxically as a means towards achieving individual hedonistic pleasure. Note that all of this is consistent with adopting a utilitarian and consequentialist approach to evaluating ethical tradeoffs, and that such does not imply relativism. In fact, many of the most famous utilitarian philosophers, such as Peter Singer, argue that utilitarianism implies the opposite, that morality is objective.

A reasonable conclusion to draw is that virtue ethics, the set of lifestyle philosophies emphasising low discounting and altruism, can be considered a human universal. I have at times throughout the past advocated vociferously against attempts to reduce (via social stigma) the relative status of hedonism, and still consider this paternalistic sentiment to render most individuals worse off overall. The particular avocations of vice we stigmatise, and the magnitude of this social penalty, is not exactly correlated with the social harms such indulgence yields anyhow. At the margin, we could clearly do with less judging, and more of “live and let live” approach to the habits of others. However, concerning the total effect, the optimal level of belief in virtue ethics is probably nonzero. If such a philosophy has been replicated independently across distinct times and places throughout human history, and remained prevalent since, then a Burkean view on institutions considers such as facilitating our social and cultural wellbeing, and therefore is culturally and biologically adaptive.

However, an alternative interpretation is that these philosophies instead aim to spin the ubiquity of suffering in a positive light. Tragedy is instead framed as heroic glory, and a courageous pursuit of a higher virtue. By raising the relative status of suffering, not only are individuals more likely to sacrifice themselves for the good of their polity (necessary when war was so commonplace to in effect be our default state), but the cardinal costs such suffering imposes are alleviated. To an extent, the Stoics and Buddhists are right that some proportion of our pain is influenced by our emotional perception of it. Whilst such appeals to virtue may have dulled the pain of the Malthusian poverty characterising most of human existence, in an era of unprecedented abundance, should we not rethink the valorisation of tragedy?

Even I find the concept of virtue ethics intuitively appealing on aesthetics, yet this line of reasoning amounts to little more than “the vibes align with my emotional priors”. If virtue is indeed always in our self-interest, then why do lofty appeals to such prevail over binding contracts? Why do even self-enforced commitments to lifestyle changes tend to fail? When the incentives of committing to an ethical heuristic tend to fail, then should we update our normative views on its merits? The answer is that, as is the central lesson of game theory, that what is biologically and culturally adaptive hence socially optimal is not always consistent with individual optimisation. Therefore we design a set of social norms to coordinate this shift to the greater good. This is how human universals accelerate cultural and natural selection.

UPDATE (11/05/2026): Recently I was thinking about optimal taxation, and on reflection, there are lessons for the utility of utilitarianism. Generally, in optimal taxation and tariffs papers, a benevolent social planner maximises a social welfare function subject to constraints such as labour supply responses. This set up makes the flaws in utilitarian calculation apparent.

As Hayek and Burke demonstrated, there does not exist a social planner that holds complete information regarding everyone's cardinal and ordinal preferences (assuming rationality, the two are equivalent). In Walrasian equilibrium, this is not an issue provided that the axioms of Arrow's first theorem are satisfied, yet that clearly is not the case. There also exists suboptimal NE with limited participants. In selecting policy, Arrow's impossibility theorem and Gibbard-Satterthwaite demonstrates that all aggregation rules are inefficient or dictatorial. Moreover, public choice theory disproves the notion of a social planner as benevolent (maximising a SWF as opposed to the interests of those participating in political institutions).

Therefore, we lack a reliable means to state with certainty what the utilitarian outcome is. Effective altruists not only ignore the calculation problem, but go further than most economists in assigning utils (cardinality), which is why I am not entirely convinced by the movement. However, if we were to adopt a Burkean stance on ethics to its most extreme, we get something along the lines of the Confucianism of the Chinese Empires, which is not only excessively rigid but a major factor in China's relative isolation from the world until the last century or so.

As such, perhaps the role of stoicism is to tradeoff these costs and benefits of crude pleasure maximising (which as we have seen may not generate efficient outcomes) vs being responsive enough to incentives and facilitating creative destruction. Stoicism departs from Eastern concepts of virtue by emphasising the role of individual agency over social hierarchy and authority. Of course, in our models, the relevant margin of adjustment is the discount rate. Whilst discounting is justified, on the margin our world can clearly benefit from a reduction in this parameter. Besides, low discounting tends to produce happier lives in the long-run anyhow. How else would you explain why unemployment or inactivity causes depression, if labour is considered a disutility in our models?

UPDATE II (19/06/2026): Despite my writings I'm actually a practitioner of the Stoïcs myself. One can consider this Substack partly as my own version of Meditations (albeit ChatGPT fulfills most of that). Stoïcism, and Aristotelian virtue in general, doesn’t just work because it's a philosophical thesis with a strong claim. It works as you apply the lessons and experiment in daily life. However this “trial and error” involves more tacit knowledge than any model can quantify or aggregate. Economics has settled for a measurable (in ordinal terms) utilitarian framework, yet only superficially does this imply crude hedonism.

Although I still align with the Epicureans, in the sense that virtue is the means to obtain a pleasurable life, rather than a good in itself. This is probably my strongest point of disagreement with many Stoïcs.

“Old, digitized astronomical images taken before the human spacefaring age offer a rare glimpse of the sky before the era of artificial satellites. In this paper, we present the first optical searches for artificial objects with high specular reflections near the Earth. We follow the method proposed in Villarroel et al. (2022) and use a transient sample drawn from Solano et al. (2022). We use images from the First Palomar Sky Survey to search for multiple (within a plate exposure) transients that, in addition to being point-like, are aligned along a narrow band. We provide a shortlist of the most promising candidate alignments, including one with 3.9 sigma statistical significance. These aligned transients remain difficult to explain with known phenomena, even if rare optical ghosting producing point-like sources cannot be fully excluded at present. We explore remaining possibilities, including fast reflections from highly reflective objects in geosynchronous orbit, or emissions from artificial sources high above Earth's atmosphere. We also find a highly significant (∼22σ) deficit of POSS-I transients within Earth's shadow when compared with the theoretical hemispheric shadow coverage at 42,164 km altitude. This deficit is still present though at reduced significance (~7.6 sigma) when a more realistic plate-based coverage is considered. This study should be viewed as an initial exploration into the potential of archival photographic surveys to reveal transient phenomena, and we hope it motivates more systematic searches across historical datasets.”

In particular, the novelty in these findings is in the methodology used to test for possible UFOs. Rather than simply count all pre-Sputnik transients (vulnerable to measurement errors and interference, or random noise in the atmosphere1), they exploited the geosynchronous orbit and isolated the effects of Earth's shadow. For any given orbit, we can calculate whether the Earth blocks sunlight at a particular point. If transients arise from glints, we should see far more transients in regions exposed to sunlight relative to within the shadow.

This is indeed the case, using the rate at which we'd expect transients in the same spot in the absence of shadow. As Robin Hanson noted, the discrepancy between this rate and actual observations (around 2/3) is roughly the same for two orbits at different radiuses, so our findings are consistent with the constraints from math. This discrepancy, from our calculations, also gives the expected number of transients that are glints.

Moreover, Hanson calculates (adjusting for advances in telescopes) that these are likely to be flat objects:

“At GEO it would take an ideal ~22-35 cm diameter circle (perfectly flat, aligned, and reflective) to create a magnitude 15-16 image in these photos from a 0.04 sec glint. (Or so says ChatGPT5.) For magnitude 17-19, that changes to an ideal 6-14 cm circle. Objects less than perfectly flat, aligned, or reflective, or with shorter duration glints, would have to be larger. Objects close to Earth could have smaller areas due to that, but would also have to be larger due to closer objects orbiting faster.

Thus while today Earth sees ~1800 glints per hour off of the roughly ~1-2cm equivalent diameter ideal circle sized human satellites and debris, in the decade before Sputnik, the space around Earth generated ~340 glints per hour mostly from brief (<0.04s) glints of sunlight off of at least roughly 6-14cm equivalent ideal circles. While this makes the older glint distribution seem different from today’s, note that the older telescope couldn’t see the fainter glints that today’s telescopes can see. So the two distributions might actually be the same.”

Therefore we can plausibly rule out random atmospheric or measurement disturbances, at least at the distances involved. Whilst the possibility of these transients being cosmic ray flashes is very much alive, the correlation with nuclear tests is consistent with alien civilisations subtly2 warning us about the existential risks involved for us and possibly extraterrestrial civilisation. On a more general note, it's useful to discipline our thinking by considering the incentives alien civilisations have to communicate with us in only a discreet manner - interpreted by (possibly) only top militaries. It seems odd that aliens wouldn't want to make themselves known right? Hanson3 arrives at similar conclusions to mine, with an added emphasis on our likelihood of casting them in the out-group should they attempt peaceful communication4.

Ergo if you accept these findings (it's worth noting that some have failed to replicate Villarroel's results, so admittedly the findings are noisy5), the very high likelihood of these transients being glints pre-Sputnik should update your priors. What are the odds that these glint observations are timed coincidentially with reports or nuclear tests6?

In any case, a reasonable base rate for our priors regarding the number of alien civilisations communication is possible with, is of course constrained by the Drake equation. Given the billions of stars and planets in the Milky Way, even tiny increases in the probability of extraterrestrial life (which nascent discoveries of biosignatures have confirmed7) massively increases this number. Assuming heterogeneity in technological development, surely at least some yield the ability to communicate across multiple light-years? So why should our priors be that the number of suspected reportings are in fact alien UFOs are zero? In my view, the burden of proof rests on those that discount this possibility.

Within occultism, there are a plethora of broad traditions. Hermeticism emphasises our divine transformation via gnosis (esoteric knowledge) as a path to revealing the hidden mechanisms of the cosmos. Kabbalah attempts to formalise esoteric theory and the interpretations of symbolic representations. The relationship with the neopagan faiths runs deep. Of course, occultism is defined just as much via practice and implementation as theory. Ceremonial magic explicitly seeks contact and alignment with supernatural forces via an elaborate set of rituals designed to achieve spiritual transformation. Thelema argues that we each have a hidden “true” purpose for our existence, and this defines the meaning of life - Thelemites tailor magic towards uncovering this individual meaning. Chaos magic is less dogmatic, and will borrow more heavily from established science.

Like rationalists (of the Déscartes rather than the LessWrong tradition!), the idea that reality is only what we can sense is rejected. Where they differ is obviously via the manner in which they seek this knowledge. The former emphasises formal logical deduction and proof from axioms; the latter vague symbols and superstition, unfalsifiable rituals, and the “supernatural” as a residual for our ignorance. There are no prizes for guessing which epistemological framework I'm more sympathetic to! Despite this, an advantage of esotericism over rationalism is that it seems more aligned with those seeking a life meaning: guidance, often in the form of personalised ceremonial activities, on how to live. Moreover, modern science essentially relied upon occultist practices in its heyday. Chemistry would not exist without alchemy. Kepler was infamous for practicing astrology. Arguably, a focus on decoding hidden processes broke us out of the Dark Ages into the Enlightenment, by shackling us away from a dependence on God as the default explanation for everything. Modern science, via wildly different methods, also seeks to decipher hidden phenomena, if we consider “hidden” as synonymous with “unknown” in this context.

We can test the robustness of astrology as a theory of the universe via tracking the accuracy of its predictions. Once we account for sampling error, astrological predictions seem to perform no better than random chance (although an astrologer would argue that randomness is determined by some esoteric mechanism). Hence this entire epistemological framework is not only unfalsifiable, but cannot be justified via a Bayesian standpoint either. So what explains its continued popularity?

I'd argue that the widespread demand for this pseudoscience is grounded in the desire to seek a secular replacement for the meanings and practices that religious institutions provide. Science has not displaced religion, as science cannot answer how we should live life. Religion provides a ready-made template for practical ethics that can be applied throughout our daily lives, with celestial enforcement and a philosophy of meaning to incentivise the prosocial. A core cognitive bias is that we seek and accept nonrandom explanations for random phenomena or coincidences, which is why we're often susceptible to astrological predictions proving correct or supposed miracles and answered prayers. To a large extent, wishful thinking - the desire to place a positive spin on life's ubiquitous tragedies, suffering, and evils - also prevails1. Then of course art and beauty is commonly motivated via the pursuit of a higher power - our cultural lives would be drastically inferior with the absence of religion.

Religion is increasingly being falsified, with claim after claim regarding metaphysics in the Bible being falsified, and hasty Straussian reinterpretations ex-post to maintain legitimacy via consistency with science. Nonetheless, it clearly serves a purpose in the advancement of our prosperity and a fulfilling life, so these superstitious institutions will never vanish. Faith, myth, and ritual, with the derived practices, are classic human universals.

Therefore, occultism is rising as a viable substitute. Indeed, practitioners of the occult seem to have learnt from the mistakes of established religions. Don't offer certain testable hypotheses that could be refuted in the future, but instead offer vague statements that can easily be interpreted to fit the facts. This is essentially the entire premise of astrology, which is why even devout atheists have an easier time accepting this philosophy. Moreover, occult practices also place an important emphasis on individual agency, so is compatible with the increasing demand for virtue ethics as exemplified via the nascent resurgence in stoicism. Plus unless you're in a literal cult, the demands for commitment and adherence are also a lot less rigid than established religions. You can essentially drop in and out, and take whatever belief you want, as you please. You cannot do this with religion.