In 2000, shortly after I left MIT, the Clay Mathematics Institute, down the road from us in Cambridge, put a $1 million prize on each of seven unsolved problems in mathematics. One was the Navier-Stokes problem, which OpenAI says an experimental model has solved.

Navier-Stokes equations describe how fluids behave, from air moving over a wing to blood moving through an artery. Engineers and physicists have used the equations for generations. What they and their colleagues in mathematics have not been able to prove is whether the solutions to Navier-Stokes always remain physically reasonable in three dimensions.

Solving that problem would be an extraordinary mathematical achievement. The dispute around the claimed solution poses a more immediate business question: What happens to your methods, calculations and ideas after you give them to an LLM?

RIAs should use AI where it improves the work. Sitting out has a cost, too: firms lose productivity and the firsthand experience their teams need as these tools enter research, planning, documentation and client service. The practical question is how to use them while deciding what the model can see, what its developer can retain and whether any of it can train or improve the model.

What the Navier-Stokes Dispute Means for RIAs

Over the past year, NYU professor Tristan Buckmaster and Anthropic's Levent Alpöge have been working on problems related to Navier-Stokes. They developed new methods along the way and used several AI models, including GPT-5.6 Sol through Codex, as part of that work. Buckmaster says he and Alpöge put drafts of their research into Codex as they worked. When OpenAI told him its model had produced a proof, he asked whether the model had been trained on, or had access to, those Codex sessions.

OpenAI says an investigation confirmed that Buckmaster's Codex prompts could not have influenced its system, including through training. The RIA question does not depend on deciding who is right in this dispute. Before a firm shares its own work with an LLM, it should know what its contract allows the LLM developer to do with prompts, outputs, feedback and other interactions, including whether any of that material can train or improve the model.

The researchers' prompts contained the reasoning, methods and insights they were developing as they worked. By putting drafts into Codex, they gave the system access to the methodology they were using to attack the problem. For a business, the valuable asset may be the method itself.

An RIA can expose the same kind of value when it asks an LLM to work through a tax calculation, encode an investment methodology into an agent, analyze an internal planning model or automate a process refined across thousands of client interactions. Those methods help distinguish one firm from another. If an AI system can learn from them, the firm should know who can use that learning later.

This month, Anthropic launched Claude for Financial Advisors, designed to connect Claude with systems and workflows used by advisory firms. Anthropic says its commercial products include controls around customer data, model training, auditability and human approval. An RIA should inspect the whole chain: the foundation model, the application built on top of it, the systems it connects to and the contracts between them. Each layer can determine what gets sent, who retains it and whether it can improve a model.

That review should cover two assets: client data and the firm's own know-how, meaning its calculations, models, workflows and methods.

How Datalign Protects What We Put Into LLMs

Andy Berkheimer, our Chief Technology Officer, built large-scale software and AI at YouTube, Google and Meta. Lucas Seibert spent a decade at Amazon working on Alexa and generative AI after earlier building the original AI for Call of Duty. Sarah Campbell, our Chief Scientist, led AI research for Alexa and later built production generative AI and multi-agent systems at Deloitte. Along with myself and many others at Datalign, our senior leaders have been working in AI for decades.

At Datalign, we follow one rule: send an external model only what it needs to perform a specific task.

Our own models, algorithms and methods stay outside the external LLMs we use. The same applies to models and information shared by our RIA partners. An LLM can participate in a product without seeing the process that makes the product valuable.

Consumer data has a separate rule. We use only LLMs whose contracts prohibit the data we send from training the underlying model. If a developer will not provide those terms, we do not use it.

We also limit the sources Halo, Datalign's AI platform for consumers, can use to formulate financial answers. It draws from a defined set, including Datalign's own financial-education content, rather than asking a model to construct financial guidance from the open internet or general training data. It checks answers against those sources before a person sees them. Samuel Carter, our Chief Compliance Officer, reviews Datalign's personal-finance content. He is a securities lawyer who previously served as in-house counsel and a compliance officer at an institutional broker-dealer and as a state securities regulator.

We disclose our use of AI in our regulatory filings as a registered investment adviser. Datalign's Form ADV Part 2A says our platform uses proprietary AI tools to analyze consumer and participating-advisor information, generate and review educational and marketing content, respond to certain communications and support the service. Our privacy policy says Datalign keeps personal information only as long as required and does not sell or share it in exchange for money or other value.

Six Questions RIAs Should Ask Before Connecting an LLM or AI Application

Before an RIA connects client data or firm methods to an LLM or AI application, ask:

1. What client or firm information does the application send to an external LLM?

2. Does that include calculations, methods or workflows the firm wants to keep proprietary?

3. Which companies receive or retain it?

4. Can any of them use it to train or improve a model?

5. How long is it retained, and why?

6. Has the team built systems that protect both client data and the firm's own methods?

A capable prototype can now be built on a foundation model in days. The harder work is deciding what the model can see, what gets stored and whether the system can learn from it. For an RIA, those choices determine whether AI improves the firm while its own methods remain proprietary.