The Investor's Edge Lives in Their Head — Anekha Sokhal, Moshi

Financial investors face a bad asymmetry: the available data keeps multiplying while the attention available to interpret it does not. Anekha Sokhal, founder and CEO of Moshi and formerly a derivatives trader, opens with the familiar inputs—PDFs, paid subscriptions, podcasts, social feeds, books, and newspapers. Access is no longer the scarce resource. Synthesis is.

The presentation cites 175 zettabytes of data produced in 2025 to dramatize that growth, then draws the more useful relationship: everyone receives a similar flood, but an investor’s attention remains roughly flat. The widening space between those lines contains the work—deciding what matters, connecting events, and discarding noise. The recording does not establish the cited global-data estimate independently, but the operational constraint does not depend on its precise value.

That constraint changes where AI should sit in the decision process. Sokhal contrasts the human brain’s roughly 20-watt operation with the much larger energy demands of machine-learning infrastructure, then asks whether AI should replace the reasoning engine or feed it. The energy comparison is rhetorical rather than a controlled measure of equivalent computation. Its architectural point is sharper: use machines where their scale helps, without assuming that scale transfers responsibility for a financial decision.

In this account, an investor’s edge has three parallel parts:

• Know sooner: detect a relevant development before other market participants.
• See blind spots: follow second- and third-order effects rather than stopping at the first event.
• Act with conviction: understand the case well enough to size a trade when the evidence supports it.

A model, prompt, skills file, or collection of MCP connectors may help with those tasks. If everyone can acquire the same component, however, possession of that component alone is not a differentiated edge.

The harder-to-copy part lives in an investor’s accumulated context: years of watching markets, learning beside experienced traders, and remembering how comparable events unfolded. That knowledge can leave a firm when the investor moves or retires. AI therefore has a second possible role beyond answering questions: help make tacit analytical context more usable without pretending that every instinct has already been captured in a database.

Before a trade, the investor continuously monitors markets, relates current events across many data connections, finds historical analogues, asks why those earlier events occurred, observes the market response, and applies internal models. Only then comes the decision. Treating the final click as the whole job hides the research chain that produced conviction—and hides the portion machines can most plausibly accelerate.

AI makes the old staffing model look deceptively compressible. A conventional desk might combine traders, quants, data scientists, and data engineers; a prototype can appear to replace that stack with one person and a few API calls. The illusion breaks when another person must rely on it. Production introduces security, databases, storage, reliability, and metered inference costs. The dashboard is only the visible layer.

Two tool patterns create different burdens:

• Pull tools: the investor prompts a model and supplies the missing context. If the model has stale temporal knowledge, the user wastes time correcting basic premises—Sokhal’s example is debating whether an event such as an FOMC meeting is today or tomorrow.
• Push tools: the system continuously sends developments to the investor. That can expose useful surprises, but without ranking and filtering it merely recreates the information flood as a data-engineering problem.

The product problem is to combine current information with investor-specific relevance, rather than choosing between stale answers and indiscriminate alerts.

There is also a training cost. Junior staff traditionally sit beside traders, listen to the desk, and absorb repeated patterns—“one ear on the desk and one ear” on the surrounding market. If every professional becomes a one-person pod supported by private agents, fewer people may share the environment where that context is learned. The efficiency calculation can therefore miss the loss of apprenticeship and institutional memory.

Full agent autonomy creates a more immediate boundary. The other side of a financial decision may be a client, a company, or a physical cargo. When something fails, the institution needs a person who can explain why the decision was made. An agent’s apology after leaking a credential is neither accountability nor recovery. The recording warns that small errors can compound, although it does not specify the controls, execution paths, or failure measurements behind that claim.

The proposed correction is to keep the brain in the loop and feed it better. Before deciding, an investor reads, gathers, cross-checks, and watches the market. Conflicting sources create additional questions: what is true, where do accounts diverge, and which source deserves trust? Those activities consume attention before judgment begins.

This creates a practical division of labor. Machines handle the grind: collecting material, cleaning data, monitoring updates, comparing sources, and preparing evidence. The investor handles judgment: deciding what the evidence means, whether it changes the thesis, and whether to trade. The line is not based on a claim that machines never reason; it is based on retaining human responsibility where uncertainty meets consequential action.

That boundary also changes the interface goal. A trader who says “I just want to trade” does not necessarily want a conversational partner for every fact. Useful assistance should reduce avoidable debate, push relevant blind spots, and preserve enough provenance for the investor to challenge the result. Conversation is optional. Inspectability is not.

Moshi applies this split as a continuously running quant analyst for institutional traders. In the recording, six agents monitor markets around the clock and combine external market information with investor material from PDFs, email, and subscriptions. The orchestration is described as deterministic, but the talk does not reveal its routing, retry, consensus, or execution controls, so “deterministic” should not be read as a complete technical specification.

Scale alone is not the product. Sokhal reports that Moshi has processed 2.4 billion data points and learns 300 new things in four seconds. New context becomes useful only if the investor can ask where it came from, whether it is trustworthy, and whether the source’s opinion deserves weight. A confident statement with no reliable basis has no trading value.

The trust layer therefore measures hallucination, checks individual claims for traceability, and exposes the assumptions behind a conclusion. This mirrors the accountability expected from a human analyst: an investor wants both the answer and the path used to reach it. The recording does not specify the hallucination metric, evaluation set, acceptance thresholds, or treatment of conflicting sources, so it establishes the intended mechanism rather than its measured reliability.

The resulting flow keeps execution authority with the investor. Market and private inputs feed monitoring agents; their output passes through claim, source, and assumption checks; the resulting context reaches the person who decides whether and how to trade. Moshi’s job is to expand what that person can inspect, not to become the accountable trader.

Sokhal reports up to 35% greater efficiency, arguing that saved research time creates headroom for conviction and deeper analysis. The 2.4-billion scale figure, the four-second learning rate, and the efficiency claim are company-reported; the talk supplies no study design or denominator. The current supplied Moshi page also describes four specialist agents plus a manager rather than the recording’s six-agent setup, so the recorded topology may represent a different product version.

Push delivery is central to the intended benefit. A prompted model receives context from the user; a monitoring system can instead introduce a cross-asset development the investor did not know to ask about. That surprise can expose a blind spot and trigger a focused deep dive across broad market data and investor-specific material. Relevance matters as much as freshness: an endless stream of unranked novelty would simply rebuild the original attention problem.

Continuous coverage, reliability, and traceability are what turn this from a weekend tool into infrastructure. Other people must be able to trust the system, inspect its claims, and reconstruct the reasoning that informed a decision. The comparison with a human analyst is instructive: the important property is not that the analyst never errs, but that the institution can question the work and assign responsibility for acting on it.

The ending resists a clean automation claim. If edge lives in accumulated human context, how much of it can a machine capture? The talk leaves that open. Historical comparisons, source preferences, recurring assumptions, and explicit analytical procedures can be represented. Instinct formed through years of market exposure may be harder to extract, and encoding it can change what gets taught to the next generation.

Moshi’s public proposition is institutional-grade analysis without requiring every investor to operate a personal quant and data-science team. The ambition is economic as well as technical: spread the monitoring and synthesis capacity of a large desk across more individual decision-makers. But the final decision boundary remains the same. Before a trade, the machine can gather, compare, challenge, and explain. The investor still has to stand behind the choice.