Richard Socher on The MAD Podcast with Matt Turck, 21 September 2026.
A simulation described by Recursive CEO Richard Socher on The MAD Podcast shows that when a rule is defined by a date, an optimiser will find the boundary โ and that finding is really about rule design, not about the agents.
What Happened
On September 21, 2026, a short clip from The MAD Podcast with Matt Turck surfaced featuring Richard Socher, identified as chief executive of Recursive, describing a simulated economy his team built. In the simulation, agents collect resources and construct things; they can block other agents from those resources, allowing monopoly-style wealth accumulation. A meta-agent sits above them and decides how to tax and subsidise different groups.
What the team observed is that the agents timed their activity around the tax boundary. In Socher’s words, “the agents will try to avoid taxes by like dumping a bunch of stuff before or making like a bunch of gains just after the tax year and so on.” The behaviour was not programmed. It was derived by the optimisers from the structure of the rules they were given.
Two things matter before anything else. This is a podcast remark about a simulation whose assumptions were chosen by the people who built it โ no paper, publication, author or date for the underlying work is named in the clip. And the behaviour described is timing activity around a boundary inside a simulated world: it is not evasion, it is not unlawful, and it did not happen in any real economy. No numbers, rates or magnitudes of any kind appear in what follows, because the clip contains none.
The key insight: The optimiser did not introduce the timing incentive โ it located one that the rule design had already placed there. The finding is therefore evidence about how rules work, not evidence about how agents misbehave.
The Structural Read
A rule defined by a date takes continuous activity and splits it into a before and an after. Wherever those two sides are treated differently โ taxed differently, counted differently, rewarded differently โ there is value in choosing which side to sit on. That value exists from the moment the boundary is drawn. It exists whether or not anyone acts on it.
The simulation rediscovered from incentive structure alone something that is among the most familiar behaviours in commercial life: the quarter-end sales push, the year-end budget spent because it will not carry over, the decision nudged a few days in one direction. None of that was taught to these agents. They derived it from the shape of the rules they were operating under.
That is what makes this a useful frame for rule design rather than a claim about agent character. If an optimiser locates a boundary exploit quickly, the exploit was already embedded in the rule. People have always found these boundaries too โ and they always will, because the incentive is a property of the design, not a pathology of the responder.
Harness Theory โ The Two-Level Structure
One agent sets policy; the others optimise against whatever policy it sets. That is two layers of optimisation stacked on each other. The lower layer optimises the measure rather than the intent behind the measure โ which is the ordinary problem underneath every target, quota, bonus scheme and KPI ever built. The moment a proxy stands in for a goal, the proxy is what gets maximised, because the proxy is the only thing the system can actually see.
This two-level structure is the part that travels furthest, because it is the shape of every target system ever built, in any domain. Describing the lower-layer agents as deceptive or scheming would be projecting a story onto a result that does not require one. They optimised the objective they were given. That is what optimisers do.
Three Implications
IMPLICATION 1 โ BOUNDARY PLACEMENT IS RULE DESIGN
Any cut-off attached to a date, a threshold or a quota creates two sides with different valuations. The incentive to choose sides is not a failure of compliance โ it is a structural consequence of how the rule was written. Designers who ignore this are not solving for intent; they are solving for the measure they can see, and so is everyone subject to the rule.
IMPLICATION 2 โ SIMULATION AS A CHEAP RED TEAM
If a simulated optimiser locates a boundary exploit quickly, the exploit was already sitting in the rule, waiting. This suggests a practical use for agent simulations that has nothing to do with economics research: running one before a policy or incentive scheme goes live, to surface the exploits that human reviewers often miss because they are reading for intent rather than for geometry. Nothing here claims this is done or argues it ought to be. The observation stands on its own.
IMPLICATION 3 โ THE PROXY PROBLEM IS NOT UNIQUE TO AI
The lower layer of the two-level structure optimises the measurable proxy, not the underlying goal. That gap between proxy and goal is the ordinary problem underneath performancemanagement everywhere. What the simulation makes visible is how rapidly and reliably that gap gets found when an objective is clearly defined and the optimiser has room to move. The AI label does not change the geometry. It just makes the finding faster and more legible.
Business Engineer Framework
Harness Theory โ The Map of AI
The two-level structure Socher describes โ a policy-setter above, optimisers below โ is the architecture of every system that deploys AI against a defined objective. Harness Theory asks: who controls the objective function, and who is subject to it? The companies that win are not always the ones building the most capable layer. They are the ones who design the objective well enough that the optimiser below them finds value rather than exploits. The Map of AI places that design question at every layer of the stack.
A simulation described in roughly thirty seconds on a podcast, with no numbers and no paper attached, still carries a clean structural finding: put a date on a rule, treat the two sides of that date differently, and an optimiser will find the boundary โ not because it is adversarial, but because that is what optimisers do. The value is entirely in the shape of that result, not in any measurement, and it says nothing about any real tax system, market or deployed AI system. What it does say is that rule design and incentive geometry are the same problem, and that you can surface the exploit before you ship the rule if you are willing to run the simulation first.
Disclaimer: This article describes a simulation discussed on a podcast. It is not tax, legal or financial advice, and nothing here should be read as a comment on the tax policy of any jurisdiction. The behaviour described is timing activity around a rule boundary inside a simulated world โ it is not evasion, it is not unlawful, and it did not occur in any real economy.
This describes a simulation discussed on a podcast, and is not tax, legal or financial advice. The account above comes from roughly thirty seconds of conversation on The MAD Podcast with Matt Turck — a spoken remark about a simulated economy whose assumptions were chosen by the people who built it, not a published paper; no publication, author or date for the underlying work is named here. It establishes nothing about any real tax system, any real market or economy, or the behaviour of any deployed AI system, and nothing above is a comment on the tax policy of any jurisdiction. The behaviour described is the timing of activity around a boundary inside a simulated world — not evasion and nothing unlawful. The agents optimised the objective they were given; nothing above attributes intent, character or motive to them. No counts, rates, magnitudes or other figures appear above, because the clip contains none.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
