A 25-Year Conversation: Why We Started RandomBits

A Friendship Built on Abstraction

We have been friends for nearly twenty-five years.

Our careers have taken us through different domains. MJ has worked across financial risk, commercial analytics, marketing effectiveness, pricing, and operations. Vineet has worked in startups and established companies, building large-scale machine-learning algorithms and platforms for advertising recommendations, online retail, and financial technology. His work has also spanned applied probability, online experimentation, and game theory. Across those different domains, we have kept returning to the same way of thinking about problems.

We often start with a problem from work, strip away its immediate details, and ask what remains. Does its essential structure apply elsewhere? The goal is not abstraction for its own sake. We want to find formal approaches that solve the immediate problem and help us understand other problems with the same underlying structure.

For more than twenty years, our conversations—over whiteboards, phone calls, WhatsApp messages, and late-night debates—have centered on a recurring set of questions:

How should a leader act when the environment is changing?
How should we evaluate an action when its effects are delayed or difficult to attribute?
How do today’s decisions change tomorrow’s options?

Vineet often starts with a mathematical framing; MJ often starts with the operating decision. Our conversations move between the two: what can the model tell us, where are its limits, and how should we decide when uncertainty remains?

Probability is difficult to reason about intuitively. Even when we acknowledge uncertainty, we often struggle to account for it in our decisions. A well-informed choice can turn out badly; a weak one can succeed through luck. We want to get better at judging decisions without letting the outcome do all the judging.

Our shared conclusion is simple: management is the practice of making consequential decisions under irreducible uncertainty.

Data Informs a Decision; It Does Not Make One

The phrase data-driven suggests that data can identify the correct action independently of judgment. That is rarely true.

Historical data reflects earlier decisions and actions: which customers were approached, which prices were tested, and which risks were accepted. Those choices shaped both the outcomes we observed and the alternatives we learned nothing about.

graph LR
    A[Prior assumptions] --> B[Selected actions]
    B --> C[Observed outcomes]
    C --> D[Updated model]
    D --> A

When we train a model only on that history, we risk reinforcing the policy that generated it. It may improve exploitation within familiar territory while leaving unexplored alternatives invisible.

This is why data should inform a decision rather than claim to drive it. Leaders must ask what the evidence supports, what it leaves unanswered, and when to test an alternative rather than continue with the established approach.

State the Objective Before Judging the Outcome

Large organizations often evaluate an initiative after its outcome is known. By then, attribution is negotiable. Favorable results attract multiple owners, unfavorable results acquire multiple explanations, and executive sponsorship can matter more than causal impact.

A more disciplined approach begins before the action is taken.

This gives us something to return to when the outcome is known. An unfavorable outcome does not necessarily mean the decision was unsound, just as a favorable outcome does not prove it was sound. We can ask whether the decision made sense given what we knew at the time, what we missed, and what we should change.

graph TD
    O[State objective] --> T[Define trade-offs]
    T --> E[Price error asymmetry]
    E --> X[Choose explore or exploit]
    X --> N[Observe next state]
    N --> O

We cannot observe the full long-term consequence of a decision made today. We can, however, estimate how an action changes the near-term state, observe what follows, and update the policy. That is the connection between practical management and formal approaches such as causal inference, stochastic control, and reinforcement learning.

Why RandomBits Exists

RandomBits is an idea we have discussed for years. We have not had the opportunity to work on the same team for a long time, although we still hope to do so one day. This publication gives us a way to collaborate now.

Here, we continue a twenty-five-year conversation by formalizing ideas that would otherwise remain in messages and calls, testing each other’s reasoning, and connecting mathematical approaches with the realities of management and engineering.

RandomBits is how two old friends keep thinking together.