AI for investors should sit on the grind before the decision, not on the decision itself. Anekha Sokhal, founder and CEO of Moshi, argues that machines are strong at reading, gathering, cleaning and cross-checking, while humans are strong at judgment. Splitting the work that way keeps a named person accountable for the trade.
AI for investors: where the machine belongs in the loop
Every investor is looking for the same thing: edge. Sokhal defines edge as three things together, knowing what is happening before anyone else, being told about your blind spots and the second- and third-order effects, and having the conviction to size a position. Nothing in that list requires the reasoning to happen inside a model.
The premise is arithmetic as much as philosophy. Sokhal says a human brain runs on roughly 20 watts, the power of a dim light bulb, across about 86 billion neurons, while the systems built to imitate a sliver of that capacity need far more power. She uses that gap to argue the technology should be pointed at a specific stage of the decision cycle rather than at the whole cycle.
The context problem has scaled up. In 2025 the world generated about 175 zettabytes of data, an exponential jump on two decades ago when an internet connection and a newspaper were the main inputs. Investors now also work through PDFs, subscriptions, podcasts and a never-ending stream on X and LinkedIn.
More data does not produce more information. It gives everyone the same flood, while attention stays flat. For people paid for financial synthesis, the gap between attention and available data keeps widening.
Moshi's product: Six Agents and Billion Data Points
Moshi is a 24/7 quant analyst for institutional traders. It runs a team of Six Agents built in a deterministic manner, watching markets continuously and combining market data with an investor's own material stored in PDFs, emails and subscriptions. Sokhal presented the architecture at an AI Engineer talk published on 2026-10-20, describing it as the first public discussion after a period of quiet work with some of the largest players in trading.
The scale figure she gives is 2.4 billion data points, growing every millisecond. She pairs it with a rhetorical claim that the system learns 300 new things in the four seconds it takes to read that sentence. Those numbers are vendor-reported and come from the founder's own presentation, not from an independent audit. Treat the throughput claim as a description of intended behavior rather than a measured benchmark.
The more useful part of the design is what it does with all that data. Sokhal says new context is worthless without knowing whether it is trustworthy, where it came from, and whether the source deserves attention. In finance, a confident answer that is wrong is worth nothing. She says Moshi measures hallucination on every claim and shows the user the assumptions behind a conclusion, so an investor can inspect the reasoning rather than accept a confident answer.
That mirrors how investors interrogate each other. Ask a trader for an opinion and they rarely give the opinion alone; they walk through how they got there. Moshi is built to supply the same trail.
Pull Tools vs Push Tools: Two Different Finance Problems
Finance tools fall into two categories that solve different problems. Pull tools are the ones you query: you supply the context, you argue with a model frozen at some training cutoff, and you get an answer shaped by what you already thought to ask. Push tools send information to you, and the problem there is volume, which is a data engineering problem rather than a reasoning one.
Sokhal's example of the pull failure is a model that insists it is 2 a.m. and time for bed when it is actually 3 p.m. on a Wednesday. The complaint is not that the model is stupid; it is that prompting a static model makes the user the context provider, and that is a poor use of a trader's attention. Debating an LLM about whether the FOMC meeting is today or tomorrow wastes time that should go into the trade.
The push side has its own failure mode. Flooding a person with more information does not produce more insight, because attention stays flat while volume grows. Her answer is to make the pushed material selective and traceable, so the trader gets a blind spot they would not have caught rather than another feed to skim. Pushed data forces the brain to make connections it would not have made on its own, and the trader can then take a 360-degree view of one event instead of skimming everything.
| Dimension | Pull tools | Push tools |
|---|---|---|
| Direction of data | You query the tool | The tool sends to you |
| Who supplies context | The user | The system |
| Main failure | Frozen model, user does the context work | Volume overwhelms attention |
| Core problem type | Reasoning and context | Data engineering |
| What good looks like | A relevant answer to a well-framed question | A traceable blind spot you would have missed |
The distinction matters for anyone evaluating AI for investors. A chat interface asks the user to do the context work. A system that watches markets continuously and surfaces what changed takes that work on itself, which is a different engineering commitment and a different cost structure.
Why the weekend dashboard breaks in production
A dashboard built over a weekend stops being a product the moment a second person needs to use it. Sokhal lists what that transition actually requires, and each item is a cost the demo never shows.
- Security, because other people's data and credentials are now inside it.
- A database, because state has to survive between sessions.
- Storage for the documents, emails and subscriptions the system reads.
- Reliability, because a desk that depends on the tool cannot have it go down.
- Credits, because running the system continuously costs money every hour.
Pricing in the category has moved accordingly, she says, with vendors charging per API call rather than per tool, which changes the economics of anything that runs continuously. The old model of maintaining a team of quants, data scientists, data engineers and traders was expensive in salaries. The new model of one person plus a few API calls looks cheap until the bill for continuous inference arrives.
That gap is where she locates the real difference between a prototype and infrastructure. A tool that has to run 24/7 across every market, on large data, in a form other people can trust and trace, is infrastructure. A tool one person built and showed a colleague is not, regardless of how good the demo looked.
Her sharpest criticism lands on the staffing assumption behind one-person pods. In finance, the traditional path is sitting next to a trader, absorbing how they think, keeping one ear on the desk and one on everything else. Consistency is what pays off. If every analyst runs alone with a set of agents, the next generation of that tacit knowledge has nowhere to form, and she argues that cost is not being priced in.
Accountability: Why Agents Cannot Own the Trade
Someone has to stand up when a trade goes wrong and say they made the decision, for a stated reason. That person cannot be an agent. Sokhal's point is that the other side of a trade is a counterparty: a client, a physical cargo in commodities, a team of people at a company. Finance runs on named accountability, and an agent that apologizes and self-corrects does not supply it.
She makes the failure concrete with an API key leak. An agent's response is to say it should not have done that and ask you to rotate the key, while a small error compounds. The argument is not that agents are uniquely dangerous; it is that self-correction is not the same as responsibility, and regulated workflows are built around the latter.
That is also why the decision loop stays human in her design. Moshi is positioned as something that makes the trader a better decision-maker, not something that trades. Sokhal puts it as Moshi being the investor on steroids: the same instinct, fed by data that lives in different places. Whether that boundary holds under commercial pressure is a separate question, and she does not claim it is settled.
What the 35% efficiency claim does and does not cover
Moshi's efficiency figure is up to 35% more, which Sokhal frames as reclaimed headspace for conviction rather than a headcount reduction. The number is vendor-reported, presented without a published methodology, sample size or measurement window, so it should be read as a product claim from the founder rather than an independently verified result.
The mechanism she describes is time reallocation. Traders tell her they do not want to spend the day talking to anyone, they want to trade, and prompting a model about whether the FOMC meeting is today or tomorrow burns attention for no return. If the system answers the calendar question and surfaces the blind spot, the saved time goes into the part of the job that requires judgment.
Notice what the claim does not say. It does not say the system improves returns, reduces losses or replaces a quant team's output. Those would require performance evidence that a talk does not supply. The 35% sits in the workflow category, and the honest reading is that a founder is describing where her customers feel relief.
The unsolved question: How Much Context Can a Machine Capture
Sokhal leaves one question open, and it is the interesting one: if real edge lives in a human brain, how much of it can a machine actually capture? She says she does not have the answer, and that whoever finds it will be building a new way of thinking in finance rather than another product.
Her account of how the context forms is worth taking seriously. It comes from decades of working alongside experienced traders, learning which signals mattered, which sources were reliable and how the market responded to similar events. Some of that is written down somewhere; much of it is not. That is why the IP leaves with the person when an investor changes firms or retires.
Traceability is her partial answer. If every claim carries its source and the assumptions behind a conclusion are visible, a user can inspect the reasoning and decide whether to trust it. That makes the system auditable and drillable. It does not make the tacit judgment reproducible, and the gap between those two things is where the open question lives.
For teams evaluating AI for investors, the practical takeaway is to separate what can be checked from what cannot. Sourcing, recency and assumption disclosure are testable properties of a system. Whether it captures an individual investor's instinct is not, at least not yet.
Does Moshi replace a quant team?
No, in Sokhal's framing it replaces the need for every individual to have one. She argues that every firm runs desks of quants, data engineers and data scientists, that running that stack per person is economically unrealistic, and that the product supplies an institutional-grade analyst as a layer while the trader keeps the decision. The pitch is access to institutional-grade analysis for people who cannot staff their own quant function, not the elimination of the people who do that work.
FAQ
Is Moshi available to individual investors?
The product is described as built for institutional traders, and Sokhal says it was developed quietly with some of the largest players in trading before the public discussion. Availability for individuals was not stated in the talk.
Who is Anekha Sokhal?
Anekha Sokhal is the founder and CEO of Moshi. She spent nearly a decade as a derivatives trader before moving into data science and machine learning on a full scholarship, and she describes her work as combining those two backgrounds in the Moshi product. Her machine learning research includes a paper at ICLR on machine learning systems with NASA.
What is Moshi?
Moshi is a 24/7 quant analyst product for institutional traders, built by a company of the same name founded by Anekha Sokhal. It runs Six Agents in a deterministic manner over market data plus an investor's own documents, and it leaves the trading decision with the human user.
How does Moshi handle hallucination?
Sokhal says Moshi measures hallucination on every claim and shows users the assumptions behind a conclusion. That makes a claim traceable to its source, though a vendor's self-measurement is not the same as third-party validation.
What does the 35% efficiency figure mean?
It is a vendor-reported claim, described as up to 35% more efficiency that traders can spend on conviction. No sample size, measurement window or methodology was published alongside it in the talk, so it should not be read as a verified productivity result.
Why keep humans in the decision loop?
Sokhal argues a person must be able to stand up and explain a decision when a trade goes wrong, and that an agent apologizing for an error does not satisfy that requirement. Finance counterparties and regulators deal with named accountability.
What is the risk of one-person trading pods?
The traditional way traders learn is by sitting next to experienced colleagues and absorbing how they think. If every analyst works alone with agents, that pipeline for the next generation of tacit knowledge weakens, and Sokhal says the cost is not being priced in.
What is the open question Sokhal leaves unanswered?
She says nobody has established how much of an investor's edge a machine can capture. Answering it, in her framing, would mean building a new way of thinking in finance rather than another tool.
Why does a weekend dashboard fail in production?
Because sharing it with one more person brings security, a database, storage, reliability and the credits to run it, none of which the demo required. Pricing has also shifted to per API call, so continuous operation carries a recurring cost that a prototype never shows.
What is the difference between pull tools and push tools in finance?
Pull tools are queried by the user, who supplies the context to a frozen model; push tools send data to the user, and their problem is volume. Sokhal treats the first as a reasoning and context problem and the second as a data engineering problem, with different design answers for each.
Turning a Spoken Argument Into a Written One
The core of this talk is a split that is easy to say and hard to build: machines take the grind, people keep the judgment, and the reasoning stays visible enough to defend. Sokhal's version of that idea lives in a 16-minute conference recording, which is a fine format for a room and a poor one for anyone who wants to reread the argument or check a number.
That problem is general. A trader's edge walks out the door because most of it was never written down, and a speaker's argument disappears the same way when the only copy is a video. Turning the recording into text is the cheapest way to keep it inspectable.
If you have similar material sitting in your own videos, whether an interview, a technical walkthrough or a position you have argued on stage, Skala Blog turns that recording into a draft you can edit and publish. The workflow is straightforward: paste a YouTube URL, transcribe the video, and generate an article from the transcript.
The related Crazystack Typescript material and the Bootcamp do Dev Doido are built around the same instinct that spoken technical knowledge deserves a durable written form. Gustavo Dev Doido teaches in that format, and a transcript is where the writing starts.
The talk makes the case that judgment is worth protecting and worth recording. A video only protects it until the transcript is turned into something a reader can argue with.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits