Anekha Sokhal, founder and CEO of Moshi, argues on AI Engineer that financial AI should support an investor’s judgment. Her proposed split puts reading, collection and data cleaning with machines while a person owns the trade.

More information competes for the same attention

  • Sokhal describes investors reading across market feeds, PDFs, subscriptions, podcasts and social platforms. Collecting more material leaves less time to connect it.
  • She cites roughly 175 zettabytes of data in 2025 as background. The talk gives no source for that estimate.
  • Her brain analogy uses about 20 watts and 86 billion neurons to argue for supporting human judgment. It is a framing device, not a measured comparison of investment performance.

The investor brings the context

  • Sokhal defines an investor’s edge through three goals: learn early, find blind spots and gain enough conviction to size a trade.
  • Widely available models, prompts and tool connections do little to distinguish one investor from another in her account.
  • Experience supplies historical comparisons, relationships between events and a sense of which sources deserve trust. That knowledge can leave when a trader retires or changes firms.

A dashboard is the start of a production system

  • A weekend prototype still needs security, databases, storage, reliability and a way to pay for repeated model calls.
  • Prompt-driven tools require investors to keep supplying context. Push-based tools need to filter a continuous stream into useful information.
  • Sokhal also warns about one-person pods: replacing teams can remove the desk-side learning through which junior traders acquire judgment.
  • She keeps human accountability central because trades affect clients, physical cargo and companies. An agent’s apology cannot explain or take responsibility for those consequences.

What Moshi says it provides

  • Sokhal describes six agents, models running multi-step tasks with tools, operating in what she calls a deterministic manner. They monitor markets 24/7 alongside investors’ PDFs, emails and subscriptions.
  • She says Moshi has read 2.4 billion data points and adds 300 new things in four seconds. These are her scale claims; the talk leaves the units and measurement method undefined.
  • She says the system checks hallucinations, traces claims to sources and exposes assumptions. The talk gives no error rates or evaluation results.
  • Sokhal claims up to 35% greater efficiency. She supplies no baseline, sample size or method, and makes no quantified claim about investment returns.

Keep the decision with the person

  • Moshi is presented as shared quantitative research infrastructure that gives an investor more room to investigate and decide.
  • Fresh, relevant information can surface an overlooked connection. Traceability lets the investor question how the system reached its view.
  • Sokhal ends with an unresolved question: how much of the context in an experienced investor’s head can a machine capture?
  • Caveat: this is a founder’s product thesis and presentation, not independent evidence that Moshi improves trading outcomes.