Columns, posts, and notes.

Research, and consequent publishing, has been a team game for the last forty years or so, and teams grew in size steadily in all areas of research, even in areas where traditionally co-authorship was not taken very positively.

But in economics and management, this trend has been reversed from 2025 onwards. Teams did not necessarily shrink though. What happened is, the number of single-authored papers took an upward turn, especially in mid-tier journals, and especially among first-time authors.

Established authors, who earlier regularly collaborated with junior scholars, seem increasingly to be flying solo as well in some cases. Teams of established authors didn’t change at all.

This trend conspicuously matches the advent of generative AI. Suddenly, the need for collaboration seems to have been replaced by the machine, and the time and cost savings have no doubt been significant.

Although, I cannot say whether this reversal is about substitutability of the machine or secrecy about machine usage itself, which is still frowned upon. The second one is a moral question which secondary data cannot answer, and I’m preparing to collect data for that.

This is what my working paper is about: Machine Execution and the Entry Coauthor: Generative AI and the Structure of Scholarly Authorship. Comments and criticisms welcome.

First posted on LinkedIn, 20 August 2026 · View the original post

A post arguing that Shapley values are misleading at best and actively harmful at worst has been circulating with approval this week, and since the argument runs through cooperative game theory, I hope a game theorist may be forgiven for answering it.

Half of what it says is correct, and has been for a while. Conditioning on a mediator absorbs the indirect path, and the local attributions of a fitted predictor carry no warrant for an intervention. Kumar and coauthors made that case at ICML in 2020. Had the post stopped there, it would have stood as a fair reminder that these plots get misread in practice.

The trouble begins with what the Shapley value is taken to be. It is an allocation rule on a coalitional function, and whatever comes out of it inherits the meaning of the function fed in. The author builds that function from an observational reading of a predictive model, gets an observational answer, and faults the rule for not being causal. None of the axioms that define the value, efficiency and symmetry and the null-player and additivity properties, appear anywhere in the demonstration.

Then came two plots, and in both cases what the author read became more important than what the method supplied. The beeswarm sits centered near zero, but that centering is just arithmetic, since efficiency forces the attributions to sum to the gap between a prediction and the mean prediction, and taking that spread for a null effect is like taking a demeaned regressor for a zero coefficient. The individual-day plot turns on a baseline that already carries the average spend of $100,000 a day, so reading it as the payoff to spending nothing swaps an average-treatment baseline for a no-treatment one. Cutting spend to zero lands about twenty standard deviations outside the observed data, which no descriptive method answers honestly.

It actually comes closest to answering itself in its final paragraph. In the example as constructed the features are independent and the fit is additive, and there the Shapley dependence of spend is the partial dependence curve the author recommends, shifted by a constant. The challenge they set answers from its slope, around 1.15 dollars of revenue per additional dollar spent, the same marginal effect they compute by hand and the same reason to raise the budget.

And this is why Causal Shapley Values exist since Heskes and coauthors at NeurIPS 2020, and they ship inside shapr, the package the post runs its code in. Serious objections to Shapley-based explanation are out there, on the observational against interventional choice of value function, on evaluation off the support of the data, on whether additive attribution answers the question a decision-maker is asking. Any one of them would have made a sharper piece. What this one demonstrates is that an observational quantity is not an interventional one, and that was not a fault of good ol’ Lloyd Shapley.

First posted on LinkedIn, 15 August 2026 · View the original post

My years of training in game theory faced some serious questions yesterday, when I tried to teach the ideas of the winner’s curse and Rubinstein bargaining to a group of super senior Group A officers of the Indian government, the likes of Directors of various departments, Joint Secretaries of States, Financial Advisors of defence establishments, and so on.

They played a common value auction in class, and there was no winner’s curse despite my desperate attempt to rig the game!

They played Rubinstein alternating-offers bargaining and reversed the patience–surplus relationship!

I’m in serious doubt now. Either something serious is wrong with the universality of the theory, or…

First posted on LinkedIn, 14 August 2026 · View the original post

I have never been taught using cases, and I have rarely used cases in my classroom, simply because that was not part of my training. But as I have been teaching in business schools for the last twelve years, I must catch up in some way. So here is my first published case study: JSW MG Motor India: The Windsor Pricing Decision, published by Gyanodaya, the case centre at IIM Kozhikode.

It is a strategic analysis of how MG Windsor’s pricing decision, and the introduction of battery-as-a-service (BaaS) in the Indian automobile market, was a brave and absolutely brilliant move despite the market leader’s initially discouraging stand on BaaS.

It is an instructive case for whoever is interested in pricing, game theory, policy design, and strategy.

First posted on LinkedIn, 12 August 2026 · View the original post

Research note, 28 July 2026. Shared on LinkedIn with the remark that, to a political being, the thirty-six days at Jantar Mantar were fascinating and heartening, while to a game theorist they were a little depressing, since events followed the playbook rather diligently.

1The puzzle

On 25 July 2026 the Union Education Minister resigned, the government undertook to withdraw the police cases filed against protesters, compensation was promised to the families of students who had died by suicide in the aftermath of the examination leaks, and a satirical outfit that had taken its name from a Chief Justice’s insult declared victory and sent everybody home. The Government of India, which possesses the police, the treasury, a parliamentary majority, and roughly three years before it must face a general election, had conceded to a movement founded ten weeks earlier by a single post on X.

Bargaining theory says the patient party wins. Something has therefore gone wrong, either with the theory or with the way we habitually apply it. This note argues for the second, which is the duller of the two possibilities and, as usual, the more likely one.

Three days later the picture has become untidier, and the untidiness became so very much loaded with information. The Cockroach Janta Party now reports detentions and arrests in Assam, West Bengal, and Bihar, describes the promised written guarantee as still outstanding, and has warned that it will resume the agitation if nothing arrives by 28 July. That happens to be today. What looked on Saturday like the terminal node of a bargaining game looks by Tuesday like an intermission, and the reason for the extension is the most interesting thing in the episode.

2What Rubinstein actually says

Two players, one pie, alternating offers. I propose a division, you accept or reject. If you reject, a period elapses and then you propose. This continues, in principle, forever.

Rubinstein (1982) showed that under complete information this game has a unique subgame perfect equilibrium, that agreement occurs immediately, and that the division depends on nothing whatsoever except how badly each party dislikes waiting. Write r_i for the rate at which player i discounts the future. In the continuous-time limit the first proposer’s share converges to

r₂ / (r₁ + r₂)

It is worth staring at that expression, because it is very nearly the whole of what this literature means by bargaining power. My share is governed by your impatience. I am paid nothing for being tough, nothing for having a larger army, and nothing at all for being in the right. I am paid entirely for your inability to sit still. Binmore makes the point somewhere in his textbooks in words to the effect that bargaining rewards you for what your opponent cannot bear to endure rather than for anything you are able to do to him, which is a sentence that should be printed, framed, and mounted on the wall of every negotiation classroom and probably a few ministries in India.

The engine driving all of this is the cost of delay. If nobody minded waiting, the model would have nothing to say, since every conceivable division would be sustainable and the theory would reduce to a shrug. Delay has to hurt. It hurts either because the pie physically shrinks, as with an ice cream in Kozhikode in April, or because the parties discount, which is the same proposition in a better suit.

3The first bad idea

The obvious move is to reach for the shrinking-pie version. Protests do fizzle. Attention wanders, the monsoon arrives, examination dates approach, tents get damp, and the crowd thins.

The difficulty is that this is not what occurred, at least not on the way up. The movement began on 16 May as a satirical platform for the unemployed. The protests opened at Jantar Mantar on 6 June with the support of several left student organisations. A second Delhi mobilisation on 20 June drew farmer unions, and the occupation from that point was declared indefinite. Sonam Wangchuk began a hunger strike on 28 June, importing his own long-running constitutional demands regarding Ladakh into a dispute about an entrance examination. Somewhere along the way the ethanol blending policy was raised against the Road Transport Minister, which is either evidence of remarkable thematic range or of a protest site behaving the way protest sites behave. Roughly twenty thousand people marched on Parliament on 20 July.

So we require the growing-pie version instead. This turns out to be a considerably worse idea than the first one.

4The second bad idea, with arithmetic

Let the stake grow at rate g and leave everything else alone. Each party now faces effective impatience r_i − g, since waiting costs interest but earns a larger pie. Agreement requires g < min{r_G, r_C}, and otherwise both sides prefer to wait indefinitely and the model predicts a permanent sit-in, which is closer to the truth than most models manage and still wrong about 25 July. Within that range the government’s share converges to

(r_C − g) / [ (r_G − g) + (r_C − g) ]

Now put numbers in it. Suppose the government is five times the more patient party, with r_G = 0.1 and r_C = 0.5. With no growth the government takes 83 percent of the surplus. Set g = 0.05 and it takes 90 percent. Set g = 0.09, just short of the government’s own impatience rate, and it takes 98 percent.

Growth makes the patient party stronger, and it does so with some hulk-like enthusiasm. The intuition is shallow, though, once one sees it. Growth is a subsidy to waiting, and a subsidy to waiting is worth most to whoever already found waiting comfortable. A commonly growing pie is the nicest thing that can happen to the side with time on its hands.

We have accordingly constructed a model of the Cockroach protest which predicts that the government would have won by a wider margin than the standard model predicts. This counts as progress in roughly the way that eliminating a suspect counts as progress. The mechanism lies elsewhere.

5Who was actually impatient

Consider the claim that the government is the patient party. The usual argument is that it has a great many issues to attend to and this was only one of them.

That argument has the sign the wrong way round. In Rubinstein the discount rate measures the flow cost of one further period of disagreement. A government running a large agenda pays a higher flow cost for an unresolved dispute, because the dispute consumes the parliamentary calendar, the ministers’ diaries, the news cycle, and the general goodwill it needs for everything else it hopes to accomplish. Having a great deal on one’s plate is an excellent reason to want this particular item off it. The multi-issue portfolio raises r_G rather than lowering it. It is no accident that the concession arrived during the Monsoon Session.

What the government does possess is an unusually comfortable disagreement payoff. It carries on governing while the argument proceeds. That belongs in the outside option and not in the discount rate, and unifying these two is the error sitting at the bottom of the original puzzle. A party may be simultaneously well insured against breakdown and desperate to have the matter closed. Anybody who has negotiated with a large organisation has met this creature and knows that it does not bleed, although it does fidget.

The aspirant sitting at Jantar Mantar, meanwhile, has an admission clock running and is about as impatient as an economic agent is capable of being. The organisation, however, is a different agent from its members, and its payoff was never anybody’s medical seat. Whose discount factor belongs in the model is a genuine modelling question, and it is worth recording that the founder is by profession a political communications strategist. The party whose comparative advantage is manufacturing attention is unlikely to have been the party in a hurry.

6Salience, and where the growth belongs

Here is the reformulation. Nothing was growing about the pie. What was growing was the cost to the government of failing to settle. Let σ_t denote the salience of the issue. Let the movement expend costly effort raising σ and let the government expend effort lowering it. Salience does not enter the surplus. It enters the two disagreement payoffs, and it enters them with opposite signs. The government’s payoff from one further period of stalemate falls in σ, through the reputational cost of visible coercion and through the widening of the set of voters for whom the issue is decision-relevant. The movement’s continuation value rises in σ, because its capital was being accumulated rather than consumed.

The reversal now appears without any strain at all. The party that was patient on 6 June found itself bargaining on 25 July with an endogenously degraded discount factor, so that the ordering of the δ’s at the settlement date is the reverse of the ordering at the start. Patience is power, exactly as advertised, but patience is a state variable and the other side has a hand on it. Stationarity is lost and the closed form goes with it, which is a nuisance for the algebra and no loss whatever for the economics.

One virtue of putting salience in the state rather than in the surplus is that the state is visible. On 20 July the authorities cut internet access around Jantar Mantar and closed five metro stations. After the site emptied on the twenty-fifth, municipal workers arrived with paint rollers and scrubbed the slogans off the surrounding walls. These are not metaphors for lowering salience. They are the literal operation, performed first on the channel and then on the wall, and anybody who has modelled information as a perishable deposit that one side lays down and the other side cleans away will recognise the second image immediately.

The framework also brings up Schattschneider, who described the entire phenomenon in 1960 without recourse to a single equation. The stronger party works to contain the audience, the weaker party works to enlarge it, and it is the losing side that calls for outside help. Sixty-six years on, the argument still has no non-cooperative treatment, which is either an oversight or a warning to anybody who goes looking for one.

7Expansion, and then contraction

One reading of the episode has the demand set growing steadily until the government buckled under the accumulated weight. The record does not support that reading, and what actually happened is the more instructive pattern.

The demand set did broaden through late June and mid-July, from the conduct of one entrance examination to the on-screen evaluation process, then to the treatment of protesters, then to constitutional questions in Ladakh, then to the ethanol policy. At the settlement, however, most of those items had quietly evaporated. Reporting after the twenty-fifth describes several demands as having surfaced and faded, leaving two central ones standing, with a third on the police cases and a fourth in the form of a five-point policy charter on examination reform.

Expansion followed by contraction is a more sensible strategy than monotone escalation and fits the model better. Breadth is how one raises salience, since each additional constituency brings its own aggrieved population into the audience. Breadth is a terrible basis on which to close, because every additional dimension is another thing the government must concede and another thing on which the coalition can split. So one expands to build the state variable and contracts to spend it. The demand that survived to the end was the one that could be paid immediately.

8Two kinds of promise

This brings us to what the government actually paid, which is where the last three days have been most instructive.

Sort the concessions not by how much they cost but by whether they execute themselves. A resignation is self-enforcing. Once the letter is on the Prime Minister’s desk and the news cycle has moved, no further compliance is required from anybody, and the minister does not quietly return the following Thursday. Withdrawal of police cases is a different animal altogether. It is a flow, it requires continued forbearance by police forces and state governments the centre does not entirely control, and its non-delivery is visible only to the individuals being prosecuted, who by then are dispersed across several states and no longer camped in front of Parliament. Compensation was conceded in terms whose escape clause was audible at the time, the government undertaking to pay whatever maximum is possible under the existing rules. So the government settled in the currency that is cheap, instantly verifiable, and irreversible, and issued paper in the currency that is expensive, unverifiable, and revocable. In the arithmetic of this game a minister turns out to be a remarkably liquid asset, and a promise about future prosecutions turns out to be a remarkably illiquid one.

The movement understood this perfectly well and said so. Its spokespersons demanded a written sovereign guarantee rather than an assurance, submitted a draft agreement to the two ministers handling the talks, took advice from senior counsel before and after calling off the occupation, and stated the mechanism plainly, which was that once the momentum went down the government might begin targeting people individually, as had happened in earlier agitations. That is a lucid statement of the model. Enforcement power was salience, salience decays, and settlement means surrendering the instrument that made the promise credible in the first place.

Which produces the position as of this morning. Arrests and detentions have been reported in Assam, West Bengal, Uttar Pradesh, and Bihar. The written guarantee is due today. The movement has said it will return to the streets without it.

If one wanted a single sentence for the whole episode, it might be that the government bought the termination of a flow cost with an irreversible payment and a revocable one, and the revocable one is now being tested at exactly the moment when the buyer has the least capacity to enforce it.

9The entomology problem

The government’s instrument for lowering salience was to deny the protesters standing. The resignation letter of 25 July contrived to invoke anti-national forces in the same document that granted the demands, which is precisely what a two-currency model predicts. Pay the symbolic price and decline to validate the frame.

The instrument has a defect. Delegitimation must be broadcast in order to work, and broadcasting it raises common knowledge of the grievance and of the number of people who hold it. In a participation game with strategic complementarities the informational effect can dominate the intended cost effect, so that turnout responds to attacks by rising.

The movement, one recalls, is named after the insult. A Chief Justice observed in mid-May that there are youngsters like cockroaches who cannot find employment and who attack everybody, and within twenty-four hours somebody had registered a website, adopted the slogan Voice of the Lazy and Unemployed, and specified the eligibility criteria as being unemployed, lazy, chronically online, and able to rant professionally. Three hundred and fifty thousand sign-ups followed within days. As far as demonstrations that the delegitimation technology can exhibit a negative marginal product go, it is a difficult one to improve upon. There ought to be a term for this somewhere in the model, and until somebody writes it down, the received wisdom that repression backfires will remain a large and confident qualitative literature containing no equations at all.

The timing around the twentieth is the fulcrum of this backfiring-of-repression mechanism. The hunger striker was removed by force on 18 July. The march met batons and tear gas on 20 July, with the police reporting around one hundred and eighty injuries including one hundred and eighteen of their own personnel, and the movement claiming a hundred and fifty of its people in hospital. The first formal talks were announced the same afternoon. Coercion and negotiation arrived inside the same news cycle, which is what one would expect if the coercive act had revealed to the government how large the cascade already was.

10The third player

On 27 July the Supreme Court took up petitions alleging police excess on the twentieth. The bench held that the right to peaceful protest is constitutionally guaranteed, that mere agitation cannot justify a baton charge, that allegations of excess must be independently examined, and that the matter may warrant a uniform national protocol governing how demonstrations are to be conducted and policed. It also agreed to hear a separate petition from the families of police personnel injured during the same events, which is a useful reminder that the audience whose expansion matters is not always sympathetic to one side.

Two observations follow. The first is that the presiding judge is the author of the remark that named the movement, an arrangement of facts that no referee or reviewer 2 would accept in a stylised example. The second matters more. The institutional reform that the bargaining table deferred may now arrive from a party that was never at the table. A uniform protest protocol is a far more durable good than a minister’s resignation, and the movement will not have paid for it in the bargain.

That makes the two-player formulation misspecified. The court should enter through the disagreement payoffs, with the probability of judicial intervention increasing in salience, which gives the movement’s investment in salience a second payoff channel it never bargained over. It also gives the government a reason to settle quickly that has nothing to do with the protesters, namely that a settled dispute is less likely to generate a binding precedent than a continuing one.

11An awkward date

There is one slightly confusing sequencing fact that refuses to fit tidily within the game-theoretic model.

Wangchuk ended his fast on 23 July, after roughly twenty-six days, in exchange for an assurance that no legal action would be taken against those protesting at the site. The resignation came two days later. So the government retired the movement’s most credible commitment device in advance, paid for it in the same currency it now appears to be defaulting on, and then conceded anyway. A model in which the government concedes because salience is high should have predicted that it would hold on once the hunger strike was over.

Two options come to my mind. Either the concession was already locked in through the talks by the twenty-third, in which case the fast was bought out as a cheap cleaning-up operation rather than as a cause of anything, or salience had by then detached from any individual instrument and become a stock the movement held independently of its most photogenic member. The second is the more interesting possibility and the harder one to test.

12What is already known

This note has almost zero novel claims, and it is a good practice to be explicit about how little of it is new. That a growing surplus can overturn the patience result is established. Flamini’s dynamic bargaining models, in which Rubinstein players choose investment as well as division, show that a more patient proposer may consume less than his opponent and that patience can leave a player worse off, with Muthoo as the antecedent. Coalition growth as a source of challenger bargaining power during a confrontation has been modelled in the civil conflict literature. Investment in organisational capacity as a route to a credible protest threat has been modelled by Inata. Protest as a device for aggregating dispersed information has been worked over thoroughly since Battaglini. And the regularity that concessions provoke further mobilisation rather than demobilisation is documented across a large sample of autocracies by Leuschner and Hellmeier, whose stated reason is exactly the one on display this week, namely that protesters have little cause to trust that agreements reached during a contentious episode will be implemented afterwards.

What appears to remain unoccupied is the two-sided contest over the salience state, since in the accumulation models the investment is jointly funded and both parties want the stock larger, whereas here one side builds and the other demolishes. Also apparently unoccupied are escalating demands as equilibrium behaviour rather than as a hardball tactic in a negotiation manual, and a formal account of why the containment instrument occasionally carries the wrong sign. Whether any of that is worth an academic paper is a separate question that I do not want to answer here.

13The control case

Any account of July 2026 has to survive contact with the summer of 2024, when students sat at the same Jantar Mantar, over irregularities in the same examination, demanding the resignation of the same minister, and were detained rather than negotiated with. Same demands, same site, same minister, opposite outcome. Whatever parameter flips the result had better be something an outsider can observe, or the exercise collapses into fitting a story to a headline.

The candidate on offer here is a threshold in salience above which conceding the symbolic demand becomes cheaper than continuing to deny the movement standing. Below the threshold, delegitimation is the efficient instrument, which is roughly what 2024 looked like. Above it, the resignation is, which is roughly what the twenty-fifth looked like. The threshold has the merit of being measurable, at least in principle, through mobilisation counts, coalition breadth, and the observable coercion the state was willing to apply in public.

Whether the threshold was crossed permanently or only for a fortnight is a question about the future rather than the past, which makes it the natural place to stop describing and start predicting.

14Forecasts, and how to be wrong about them

A model that only accounts for the past is a description with equations attached. What follows is an attempt to make this one falsifiable within weeks rather than years. The probabilities are subjective, rounded, and derived from the model rather than from any independent reading of Indian politics, which means they inherit whatever is wrong with the model.

Begin with the state variable. Salience is falling and the movement knows it. The site is cleared, the walls are painted, the hunger striker has gone home, and the focal moment has already been spent on declaring victory. Two things work against the decay. Opposition members have carried the police action in Bihar into the Monsoon Session and are pressing the Home Ministry on it, which means salience is now being maintained by an actor whose costs the movement does not bear. The pending petitions keep the twentieth of July in the news on a schedule the government does not control.

Short term, two to four weeks. The written instrument probably arrives in narrowed form. Roughly seventy percent that something in writing is delivered within a week, and conditional on that, roughly eighty-five percent that it binds the Delhi Police and the central agencies while treating the states as a matter for advice rather than instruction. A blanket forward-looking guarantee against future cases across all jurisdictions is close to legally unavailable, and the demand was probably always more useful as a test of good faith than as an achievable instrument.

The movement then faces the harder problem. Remobilising after a declared victory costs considerably more than sustaining an occupation, because the coordinating belief has already been consumed. A full resumed sit-in during August sits at roughly twenty-five to thirty percent, with a single-day action or a legal offensive the more likely response to partial delivery. The decision to take senior counsel before and after calling off the occupation suggests the organisation has already worked out that the courts are now the cheaper instrument. Cases in states the centre does not control will probably remain unresolved, since it cannot deliver there and has an obvious interest in saying so.

Medium term, three to nine months. The legislative response is close to certain in form and doubtful in substance. Anti-paper-leak legislation formed part of what was agreed, and since a statute on unfair means in public examinations already exists, the likely output is an amendment, a restructuring of the testing agency, and a committee. Something announced by the close of the next Budget session sits at around seventy percent. Whether any of it changes examination outcomes is a separate question on which the record invites pessimism.

Compensation will be slow and smaller than demanded. The concession was made in terms whose escape clause was audible at the time, and ex gratia schedules rarely reach the figures that movements ask for. Around eighty percent that disbursement is materially below the demand and takes more than six months. The police cases will mostly decay through non-prosecution rather than through formal withdrawal, since quiet attrition costs nothing and requires no visible admission of anything.

The sleeper is the court. A uniform national protocol for demonstrations is the one durable institutional output currently in play, and it may not turn out to be a victory. Codification constrains police discretion and protester discretion at the same time. A movement that mobilised in part by overstaying its authorisation may find the resulting framework harder to work with than the ambiguity it replaced.

Long term, one to five years. The organisation is more likely to fragment or professionalise than to become an electoral vehicle. Base rates for the conversion of protest movements into durable parties are poor, and the founder’s own history with a party built on exactly that conversion is evidence in both directions. Formal registration and contesting sits under thirty percent, with a standing advocacy organisation the more likely destination.

The repertoire will outlive the organisation. Adopting the opponent’s insult as a brand, running the coordination layer on commercial platforms, and insisting on written terms rather than assurances are all cheap to copy and were publicly validated in July. Expect imitation, and expect the containment side to adapt toward earlier and quieter intervention, since visible coercion is precisely what the model says carries the wrong sign once salience is high.

For the government the interesting parameter is precedent. If concessions raise the expected return to future mobilisation, as the cross-national evidence suggests, then the shadow cost of the twenty-fifth is considerably larger than one resignation. That argues for harder early containment in the next episode, which happens to be the condition under which the backfire mechanism operates. The state elections due in early 2027 are the first occasion on which youth unemployment becomes a competitive issue rather than a protest issue.

How to be wrong. If a full occupation resumes and reaches July’s scale within a month, the salience-decay account fails and salience behaves more like a stock than a flow. If the government delivers the complete sovereign guarantee including the forward-looking clause, the argument of Section 8 fails and irreversibility was not what made the resignation cheap. If the court declines to move, the third player is doing no work and the two-party formulation was adequate after all. Each of those is observable within weeks, which is more than most of this literature can say for itself.

The movement has said it is resting and regrouping, that it means to take up one issue at a time, and that this is not the end. The government, on present evidence, is testing whether the stock of attention has depreciated far enough to make the second half of the bill optional. Both of them are behaving exactly as the model says they should, which is either reassuring or slightly depressing depending on where one is sitting.

Chronology

References

First shared on LinkedIn, July 2026 · View the original post

Check the PDF in the post. This is a masterclass in how NOT to present data, ever, to anyone, UNLESS you are too desperate to hide facts.

There’s not a single meaningful piece of information across the entire PDF. Female workforce participation increasing from 1000 in 130 crore to 2000 will mean a 100% increase. I should never make a poster out of that! Out of all Jan Dhan account holders, some 55% are women. That means nothing unless you tell us what percentage of women have a Jan Dhan account!

Maybe the Gates Foundation has done very good work. But this type of data obfuscation definitely smells very, very fishy.

So, never ever present data like that.

The poster referred to is the Gates Foundation India document attached to the original post.

First posted on LinkedIn, 15 July 2026 · View the original post

Every few days my other social media feeds serve me another post crowning Messi the undisputed GOAT of the World Cup, and the proof is always the same stack of raw totals. Most goals ever, most assists ever, most appearances ever, nobody has done this, nobody will again, cue the reverent emoji.

As an economist I have a small, pedantic, and deeply annoying objection. Every one of those is a counting statistic, and the thing being counted has been quietly getting bigger for almost a century. The tournament had 13 teams in 1930 and will have 48 in 2026. More teams means more matches, more knockout rounds, and more opportunities to pad your tally before anyone sends you home. A player who appeared in six modern World Cups is not automatically better than one who appeared in two older ones. He was mostly handed a much larger denominator.

So I did the boring thing and divided by matches played.

On goals per game, Gerd Müller and Mbappé sit comfortably ahead of Messi, and Just Fontaine, with 13 goals in 6 games back in 1958, is operating in a separate universe that the rest of us are not invited to. On assists per game, Maradona and Pelé come out ahead of him as well. Messi owns the raw records because he showed up 31 times, which is its own rare and unrepeatable achievement, only not the achievement the posts are actually claiming.

None of this demotes Messi from greatness. It only reminds us that “most ever” and “best ever” are different quantities, and that we keep reporting the numerator while silently hoping nobody asks about the denominator.

Footnote for the colleagues already drafting an angry comment. Official assist records begin only in 1966, so Fontaine and half of Pelé’s creative output never made it into the ledger, which means even my correction is politely incomplete. Peer cursing awaits.

First posted on LinkedIn, 14 July 2026 · View the original post

Applied game theory note, May 2026.

01A cartel under structural stress

The United Arab Emirates formally withdrew from the Organization of Petroleum Exporting Countries and the wider OPEC+ alliance on 1 May 2026, ending nearly six decades of membership. The UAE’s production capacity had grown to approximately 4.8 million barrels per day against an OPEC-mandated ceiling of roughly 3.2 million, a gap that had generated persistent frustration within its energy ministry. The UAE’s stated rationale for departure was to recover the freedom to respond to market conditions “at the right time and at the right pace,” a phrasing that carried significant strategic content.

The proximate context was the disruption caused by the US–Israel war on Iran, which throttled Persian Gulf exports through the Strait of Hormuz and temporarily forced all Gulf producers, including the UAE, into de facto production cuts regardless of formal OPEC commitments. The UAE’s energy minister acknowledged that the disruption created “an opportune time for the move,” suggesting that the exit timing was strategic rather than reactive.

“They are clearly preparing for the period after the war. Now that we have reached peak oil demand and are entering a new environment, they want to be free from the constraints of OPEC.”Kingsmill Bond, energy strategist, Ember Future, April 2026

The departure dealt a structural blow to OPEC that goes beyond the loss of one member’s quota compliance. The UAE was second only to Saudi Arabia in spare production capacity, the instrument through which the cartel enforces compliance by threatening to flood the market. With that second anchor gone, the cartel’s enforcement mechanism rests entirely on Riyadh. This note sets up the situation as a game-theoretic problem across three nested games: cartel sustainability, voting power, and geopolitical alliance formation.

02OPEC as a repeated prisoner’s dilemma

OPEC has always been a cartel solution to a prisoner’s dilemma. Each member, considered individually, has a dominant strategy to produce above its assigned quota: more barrels at the prevailing price yields more revenue, regardless of what other members do. The problem is that if all members follow this dominant strategy simultaneously, prices collapse and everyone is worse off than if all had held to the quota. The cartel’s function is to coordinate on the Pareto-superior outcome when the non-cooperative Nash equilibrium is individually dominant.

OPEC / Saudi Arabia

Honour quotas (cut)

Flood market (punish)

Stay in OPEC
(honour quota)

UAE +8 / OPEC +10High price, stable market

UAE −5 / OPEC +4UAE penalised, low price

Exit OPEC
(flood market)

UAE +13 / OPEC +6UAE free-rides on the cut · current position

UAE +2 / OPEC +2Price crash · Nash equilibrium

The UAE’s exit can be read as the moment at which it judged the punishment mechanism to be below the threshold required to sustain cooperation. Three factors converged to bring it to that threshold. First, Iraq and Russia had routinely exceeded their quotas without triggering a price war, which degraded the credibility of the enforcement threat. Second, the UAE’s outside option improved substantially via the Fujairah terminal and the Abraham Accords. Third, and most important theoretically, the UAE’s assessment of the game’s horizon shortened decisively.

03Peak oil demand and the endgame logic

The most theoretically significant dimension of the UAE’s stated rationale is its explicit appeal to the finitude of the game. The UAE’s strategy is premised on the belief that global oil demand has reached or is approaching its historical peak, and that the window for extracting maximum revenue from its reserves is closing. This is precisely the condition under which the folk theorem’s support for cooperative equilibria unravels.

Saudi Arabia’s position is structurally different. Its fiscal architecture is built around high and stable prices over a long horizon, and its Vision 2030 programme presupposes decades of continued oil revenue. Its breakeven oil price is higher per unit of government expenditure than the UAE’s. Saudi Arabia thus has a longer subjective shadow of the future, and its optimal strategy remains cooperation, even as the UAE’s shifts to defection. This divergence in subjective T is the root cause of the disagreement that finally produced the exit.

04Shapley–Shubik power shift within remnant OPEC

Political and economic power in a cartel need not be proportional to seat or quota share. The Shapley–Shubik power index measures a player’s influence as the fraction of all orderings in which that player is pivotal, the player whose inclusion first brings the coalition to or above the winning threshold. An analogous logic governs OPEC’s internal power structure, where the pivotal resource is spare production capacity rather than legislative seats.

A member is pivotal if its capacity to expand or withhold production determines whether a price war is credible. Prior to the UAE’s exit, the two members with meaningful spare capacity were Saudi Arabia and the UAE, who together controlled the majority of the world’s approximately 4 million barrels per day of idle capacity. This placed them in a two-anchor configuration: either could credibly threaten a price war independently, and together their threat was unassailable.

Before the UAE exit · dual anchor

Saudi Arabia, 2.4 mbpd spare

UAE, 1.6 mbpd spare

Others, negligible

Power index (approx.): KSA 0.50 · UAE 0.33 · Rest 0.17

After the UAE exit · sole anchor

Saudi Arabia, 2.4 mbpd spare

UAE, outside the cartel

Others, negligible

Power index (approx.): KSA 0.83 · Rest 0.17

The paradox of the UAE’s departure is that it simultaneously raises Saudi Arabia’s formal dominance within OPEC and weakens the cartel’s collective enforcement capacity. Saudi Arabia controls a larger fraction of a smaller enforcement capability. The residual members’ individual indices increase, but the coalition’s total capacity to exercise authority falls. Removing the second anchor from a two-anchor system is not merely a linear reduction in power; it eliminates the redundancy that made the enforcement threat robust.

05Geopolitical coalition reconfiguration

The departure also restructures the geopolitical coalition game at a broader level. In cooperative game theory, a coalition forms when its characteristic function value exceeds what its members could obtain by acting individually or in alternative coalitions. The UAE’s implicit claim is that its characteristic function value outside OPEC exceeds its value within it.

The Abraham Accords relationship with the United States and Israel has raised the UAE’s outside option as an oil exporter in ways that are not available to most other OPEC members. It gains preferential access to Western capital markets and upstream technology partnerships, and it receives political protection from US military presence in the Gulf. The remaining OPEC members with strained US relations, Iran, Venezuela, and Libya, have no comparable outside options, which keeps their OPEC payoff relatively high and their defection temptation lower.

Emerging UAE axis

UAE · USA · Israel · Fujairah

Abraham Accords and market freedom. High outside option.

v(S) outside > v(S) in OPEC

Remnant OPEC

Saudi Arabia · Iraq · Kuwait · others

Quota regime and price floor. Lower outside option.

v(S) in OPEC > v(S) outside

06What the model predicts

The analysis across the three nested games converges on the following scenario assessments. The near-term picture remains constrained by the US–Iran war and Strait of Hormuz disruptions, which force a temporary cooperative outcome regardless of formal OPEC membership. The predictions below concern the post-war medium term.

Scenario	Mechanism	Price outlook	Assessment
Saudi accommodation	Saudi Arabia absorbs the production burden alone, accepts the UAE’s increased output as a fait accompli, and continues managing supply through its own adjustments. Rising volatility as the sole-anchor system is less nimble.	$80–90 per barrel, with rising volatility	Most likely
Informal bilateral deal	Saudi Arabia and the UAE negotiate a production coordination agreement outside the formal OPEC quota structure, effectively recreating a two-anchor system without the institutional overhead. The UAE’s stated willingness to “continue engaging with producers” is consistent with this.	$85–95 per barrel, lower volatility	Plausible
Saudi-led price war	If the UAE floods the market aggressively once the Strait reopens and other OPEC+ members follow suit, Saudi Arabia triggers a disciplinary price war analogous to March 2020. Marginal producers face fiscal crisis; some may exit or face political instability.	$55–65 per barrel for 12 to 24 months	Lower, non-trivial

Regardless of which scenario obtains, rising oil price volatility is a near-certain outcome. When spare capacity is distributed across fewer coordinated parties, supply shocks are absorbed less efficiently and the market’s ability to price in geopolitical risk degrades. Future demand spikes or disruptions will be met by a less nimble collective response, producing sharper price swings in both directions.

“Without a unified mechanism to balance supply and demand, markets may experience sharper fluctuations, particularly during periods of geopolitical stress or demand shocks.”The Global Economics, May 2026

The deepest prediction the model offers is structural rather than price-level. The cooperative equilibrium among Gulf oil producers is now conditional on Saudi Arabia’s willingness to bear the full enforcement cost alone, and that willingness has a finite limit. If UAE production growth erodes Saudi market share to the point where the price of enforcing cooperation exceeds the benefit, the Saudi switch from accommodation to price war becomes a rational move rather than an emotional reaction. That threshold, not bilateral diplomacy and not the Hormuz situation, is the variable that determines the long-run equilibrium of Gulf oil markets.

Applied game theory note, May 2026.

A particular pattern of phrase has become so common nowadays that it now functions as a telltale sign of AI-assisted writing. It shows up in LinkedIn articles, op-eds, consultancy decks, grant proposals, teaching notes, and even academic drafts. Many readers can spot it instantly, even if they cannot yet name it.

The pattern usually takes a form like: “it’s not just X, it’s Y.” Or: “this isn’t merely A; it represents B.” Or: “the issue goes beyond P and speaks to Q.”

Once you start noticing these constructions, they appear everywhere. They tend to surface at moments where the writer wants to sound reflective, signal depth, or mark a transition in the argument. They bear a promise of some kind of insight while absolutely avoiding any explanation. They reassure the reader that something important is happening. This is called the contrastive pivot.

This pattern actually did not originate with generative AI. Long before large language models, it flourished in management writing, opinion journalism, TED-style talks, and academic introductions, because specifically these genres reward rhetorical confidence and narrative flow. They often value readability and punchlines over analytical precision. Contrastive pivots thrive in such environments because they compress the persuasion that could have taken multiple paragraphs into a single sentence.

Generative AI systems learn this habit faithfully. They are trained on vast amounts of publicly available prose drawn heavily from these domains. During fine-tuning, human evaluators consistently reward writing that feels polished, clear, and confident. Sentences that perform rhetorical upgrading score well on those dimensions. They create the impression of insight while avoiding specific commitments that could be evaluated as incorrect. It’s a very attractive shortcut for many. And thus it became so prevalent in the training materials of generative AI, because the goal of many public-facing texts is precisely to deliver the punchline in a succinct manner without explaining. This is what we call a sampling bias.

From an algorithmic perspective, these sentences are also very convenient. They offer an easy way to move from one paragraph to the next. They close a topic without fully resolving it and open a new one without fully defining it. For a system optimising next-token probability and coherence, this is THE lowest-risk and maximum-scoring path through the text.

Moreover, this phenomenon feeds itself because of the mindless copy-paste job of AI-generated texts. People copy AI-generated drafts into blogs, reports, lecture slides, newsletters, and posts. Those texts circulate publicly. They are indexed, summarised, and scraped. Over time, they re-enter the training ecosystem. A stylistic mannerism becomes statistically dominant through repetition. The model increasingly encounters its own rhetorical habits reflected back at scale, and thus the usability of and confidence in using these phrases grows exponentially.

This explains why the pattern feels unusually visible now. It has crossed a threshold from familiarity into absolute saturation. Readers now encounter it often enough to register it consciously. What once passed as a stylistic flourish now reads as a lazy formula to look cool (see what I did here?).

At that point, responsibility shifts from the tool to the writer. Using AI to generate a first draft is not the issue, but leaving these sentences untouched is. Their presence often signals that the writer did not reread closely, did not translate the draft into their own voice, and did not clarify what they actually meant to say. For attentive readers, this has become a proxy for measuring or silently judging the effort of the writer.

Another problem, that is likely a deeper one, is intellectual. These sentences frequently substitute relabelling for reasoning. They rename a phenomenon instead of explaining its structure. They gesture toward depth instead of arguing it. And hence, when these unsubstantiated oversimplifications are overused, they flatten arguments and dull distinctions. Editing them out usually improves a piece immediately. Removing the sentence often reveals that nothing essential is lost. Rewriting the paragraph with explicit mechanisms, concrete claims, or clear transitions forces the writer to take ownership of the argument. But once we sit so comfortably on the lazy boy of ChatGPT, who wants to get up and do some real work? Our utility function for creating meaningful texts has altered for good, it seems.

Generative AI has made producing fluent prose almost costless. But it is high time to understand that it has not made thinking costless. Writers who recognise this difference and edit accordingly will remain legible and credible. Writers who do not will leave behind a trail of sentences that quietly shout how the text was produced. And don’t you dare take the readers for granted. Readers are already learning to listen for that signal.

First published as a LinkedIn article, 9 February 2026 · Read on LinkedIn

Generative AI has entered public life with a familiar promise. It will save time, boost productivity, and perhaps even democratise excellence. Alongside this promise sits an equally familiar fear that it will destroy jobs. My experience as a teacher suggests a different diagnosis. The real disruption is not that generative AI replaces human capability, but that it exposes how much of our education and early career culture has been drifting away from judgement, critical thinking, and responsibility.

In my classroom, ChatGPT and similar tools are everywhere. Students use them to draft answers, write case analyses, summarise readings, prepare presentations, and even generate research ideas. This is no longer a marginal behaviour. It is normal. Yet what is striking is that access is not what separates outcomes. Everyone has the same tool. What separates outcomes is the user. Only the very good students produce something meaningful and impactful with generative AI. They treat it as a starting point, not a proxy of their own self. They ask better questions, demand alternative framings, impose constraints, and then check whether the output is logically sound and factually credible. They edit, tighten, and take responsibility for what they submit. For them, AI is a productivity amplifier. Weaker students, by contrast, often generate trash. More worryingly, many do not even know it is trash. Many lack the domain understanding needed to detect what is wrong. Others are not incompetent so much as disengaged. They are willing to submit output they have not interrogated. And since generative AI writes in a confident, fluent tone, it creates an illusion of quality. The language looks polished, so the thinking is assumed to be good. This is precisely the trap.

That trap tells us something important about the labour market debate. Generative AI does not reliably convert an average worker into an excellent one. It speeds up the production of text, slides, and code, but it does not uniformly supply judgement. It produces plausible sentences, not reliable decisions. It can mimic expertise, but it cannot guarantee truth, context, or relevance. When the tool becomes universal, the differentiator is not access. It is the ability to formulate the problem properly, verify claims, and own the consequences. This is why the comforting belief that “AI will make everyone equally productive” is misleading. When everyone can generate output quickly, the scarce resource becomes credibility. Employers and managers will increasingly value people who can do what AI alone cannot: identify what matters, separate signal from noise, test assumptions, notice missing variables, anticipate risks, and explain trade-offs in human terms. In other words, the premium shifts toward judgement.

The classroom reveals another uncomfortable shift. We did not use to see so much low-quality work. Now it can feel like a flood. Not because students suddenly became less capable, but because generating large volumes of plausible material has become nearly costless. Earlier, producing pages of nonsense required time and effort. Today, a few prompts can create a long, polished answer in minutes. That changes incentives, especially in a system already tilted toward speed, credentials, and the next placement season.

In India, this matters a great deal. We have long devalued critical thinking in practice, even when we praise it in speeches. The race for jobs often rewards “getting it done” over doing it well. Many students are trained to optimise for marks, templates, and interview heuristics. In that environment, jugaad is not merely a stereotype. It becomes a survival strategy. Generative AI fits perfectly into that culture because it offers the ultimate workaround. It can produce something that looks complete even when understanding is thin. It makes shortcuts scalable.

This also explains why frustration is rising among teachers and evaluators. The tool is not merely helping students. It is changing the noise level in the ecosystem. When a large share of submissions is AI-generated and lightly edited, we spend more time detecting superficial work and less time nurturing deep work. The overall quality distribution shifts in an odd way. The best students may become even better because they can iterate faster. The bottom end, however, becomes louder, because producing poor work has become easier than producing silence. If this is the reality in classrooms, it will soon be the reality in workplaces. Organisations will see more polished outputs, but not necessarily better thinking. The cost of producing documents will fall. The cost of evaluating them will rise. Trust will become more expensive. Verification and accountability will become more valuable. A manager will not only ask for a report, but will ask who can defend it, who checked it, and what would change if key assumptions were wrong.

There is also a policy problem hiding inside what looks like a productivity boom. Entry-level work has traditionally been a training ground. Juniors learn by drafting, summarising, doing routine analysis, and receiving feedback. If AI performs the first draft, many organisations will reduce these roles or compress training. Over time, we risk weakening the pipeline through which judgement is built. That matters for India because we already struggle with the gap between degrees and job-ready skills. AI can widen that gap if it becomes a crutch rather than a tutor.

So what should be done? The answer is not to ban generative AI, nor to pretend it is harmless. The answer is to redesign incentives and evaluation, in both education and employment.

In education, we need assessment methods that reward reasoning, not mere output. Students should be asked to explain their choices, defend their assumptions, and revise under critique. Oral exams, supervised writing, open-book but time-bound analysis, and iterative submissions with feedback can reduce the returns to copy-paste work and increase the returns to genuine understanding. The goal is not to catch students using AI. The goal is to ensure that using AI does not substitute for thinking.

In hiring and early career training, organisations must change what they test. If AI can draft a case solution, then the interview should test whether the candidate can critique that solution, identify missing pieces, and make decisions under constraints. Firms should explicitly train juniors in verification, fact-checking, domain logic, and responsible use of tools. Otherwise, they will hire people who can generate outputs but cannot be trusted with outcomes.

Most importantly, we should stop treating generative AI as a miracle that upgrades everyone equally. It does not. It scales the user. For disciplined, thoughtful people, it is a force multiplier. For those trained to seek shortcuts, it multiplies superficiality. That is why the long-run impact is unlikely to be a simple story of jobs disappearing. The deeper story is that the market will reward judgement more, because everything else becomes cheaper to imitate.

Therefore, generative AI will not eliminate the labour market’s hierarchy of competence. It will expose it. And if India wants this technology to be a genuine opportunity rather than a factory of polished nonsense, we must restore the value of critical thinking, not merely celebrate the speed of getting things done.

First published as a LinkedIn article, 28 January 2026 · Read on LinkedIn

In one of the earlier sessions, I was teaching coordination games to my students. After that, I started following the news of Arattai, and found a nice case study for my class. How can Arattai be successful?

It is stuck in a classic assurance game, where there are two pure equilibria. One is where everyone takes up Arattai, and one where everyone stays with WhatsApp. It’s optimal to choose one specific social network, be it Arattai or WhatsApp, when there’s reason to reasonably believe that everyone else is choosing the same social network. The formula is “do what you THINK everyone is doing”. So, if you can somehow convince me that everyone is shifting to Arattai and businesses are also shifting their communication to Arattai, I’ll follow suit. Else, the status quo will prevail, however great their product may be.

That’s the case for the rise and fall of every social networking idea so far. Either using the new network has to be a viral trend, or regulations and fear (real or otherwise) have to push people away from the incumbent network. This is why we rarely see two social media giants in the same place. In each niche category (within each market) there’s only one giant, and the rest are at the fringe.

A developed version of this argument appeared as “Arattai and the coordination game” in Deccan Herald on 7 October 2025.

First posted on LinkedIn, 3 October 2025 · View the original post

After the sudden change in the H-1B visa rules, many experienced professionals, some rather filled with unbridled joy, started predicting that all the great talents will now come back to India.

I am not an experienced professional of that world. But my limited understanding fails to comprehend such logic.

There are several types of workers who work on an H-1B visa.

(A) Those who are already in the USA will most likely try to shift to similar projects in other friendly countries like the UK or Canada, and their companies would be willing as well. Who will be successful? The most useful ones (one can even use the proxy of “talented” here). The rest, for whom the company cares less, will probably have to come back.

(B) Those who were supposed to move for onsite projects will face another shortlist. Only the highest-RoI people will be sent to the onsite project team, and the rank and file will have to either stay back or seek projects in other countries.

Then there will be movements between companies, and suddenly those companies who have good establishments in other developed countries will start to look attractive. Because a large part of wannabe immigrants primarily move out because they want a better life, not out of love for a specific country.

So at the end, what remains back here is the lower-level, low-RoI, less impactful people, because for the rest, the companies should be willing to invest in the USA or other countries.

So, is that such a great outcome?

The argument appeared in developed form as “Game theory of Trump’s H-1B shock: why India might inherit the leftovers” in The Economic Times a few days later.

First posted on LinkedIn, 21 September 2025 · View the original post