AI Labs Will Trap You or Replace You - Dillon Rolnick, Nous Research

Loading slides…

Read the talk

When Models Get Cheaper, Who Owns the Investment Process?

Dillon Rolnick, CEO of Nous Research, argues that falling model margins give labs a reason to raise prices on firms that depend on their agents or compete with those firms directly. An investment decision illustrates what a firm can keep under its control when it builds on an open framework.

From a talk by Dillon Rolnick

At a glance

Ideas worth remembering

  • Measure the share of business output that depends on an agent platform. Employee adoption and token spend alone can obscure what a change in the provider’s terms would put at risk.

  • Trace one investment decision through its data sources, routine code, analysts’ opinions and model judgment. Check which parts the firm could carry forward if it changed models or left the provider.

  • Separate the work that encodes the firm’s expertise from general agent maintenance. A framework can supply improvements while the firm controls its process, deployment and choice of model.

Falling token prices complicate a trillion-dollar bet

Dillon Rolnick opens with a question about recovering the cost of building AI. He puts individual labs’ infrastructure and research bets in the hundreds of billions of dollars and the industry’s combined investment at perhaps trillions. Consider the customer on the other side of that bet. An engineer curious about a new model can move among Astra, Fable, Opus and Kimi with little attachment to the company serving the request. Each lab has paid to develop a capability that its customer feels increasingly free to shop around for.

Rolnick points to three trends in charts that had circulated that week. The average price of a token was falling, total spending on tokens was surging, and the infrastructure costs associated with serving them had stayed roughly flat. Growing demand and shrinking margins can coexist. Customers can buy far more tokens even as competition narrows the premium a lab can charge for a particularly capable model. In Rolnick’s reading, a broadly available baseline of intelligence changes where a lab can expect to make its money.

0:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

The agent layer gives a lab leverage over business output

Rolnick sees two broad responses. A lab can move to another layer of the stack, either down toward infrastructure or up toward the user. Or it can win the race to what he calls godlike intelligence and monetize the lead. He spends most of his time on the routes that bring a lab into a customer’s business. Owning the interface puts the lab in a position to retain the user. Developing a sufficiently capable system gives the lab a reason to pursue the user’s business itself.

A two-item numbered list appears in white over a blurred blue image of a figure seated cross-legged.Open full source frame
Compete at a different layer / Build god

Agents make the first route especially attractive. Rolnick considers consumers fickle, while businesses buy large volumes of tokens for work that generates economic value. He recalls an earlier speaker mentioning 5 percent of a workforce engaging with agents. He asks executives to consider a different measure two or three years out. What share of the company’s output might be agentic? His own forecast is 70 percent, which he considers conservative. The two percentages describe different things. A head count of agent users tells an executive little about how much production depends on their tools.

Rolnick likens the resulting dependency to someone else having the power to organize a strike. If an outside platform coordinates most of a firm’s output, its owner can change the terms once the firm has embedded it. A rival’s cheaper tokens offer little immediate relief to a company whose work runs through the incumbent. The possibility of interruption gives the provider room to raise prices.

1:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:53 · section reference included

A lab that can solve a market has a reason to enter it

Take the second route and suppose a lab achieves the intelligence it set out to build. It now has a choice about a lucrative problem. It can sell intelligence at a narrowing margin to a customer who captures the return, or it can direct that intelligence at the problem and collect the return itself. Rolnick believes labs view financial markets as tractable enough to make the latter course tempting.

Domain knowledge constrains that ambition today. A lab may believe it can solve markets and still lack the context, know-how and experience that a financial firm brings to an investment decision. Rolnick points to labs hiring investment banking analysts and partnering with some of the world’s largest firms. Those moves give labs access to expertise they lack. His prediction is that they will eventually pursue the financial opportunity directly. For a firm considering a partnership today, that possibility matters before a competing product exists.

3:55
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:55 · section reference included

The cost of leaving grows as the organization learns

Embedding a product and raising prices once customers depend on it is a familiar strategy. What changes, Rolnick argues, is the scale. He believes tens of percent of the world’s economy could be at stake. For the executives and engineers choosing a platform today, dependence grows as their colleagues learn how to get good results, turn that knowledge into routines, and build systems around those routines. A later move has to account for the know-how the firm has built into those systems.

That is why Rolnick advocates choosing an open foundation while the processes are still taking shape. He recommends Hermes Agent, then addresses his own company’s incentives. Nous Research, he says, began as an effort to resist concentrated control of a technology its founders considered exceptionally consequential. He describes a staff of science-fiction fans for whom the risk was easy to picture. His joke that Skynet was closed source makes the founders’ preference clear. The business question for Nous is whether it can earn a living serving customers who retain the power to leave.

5:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:18 · section reference included

Nous reproduced a Grok bots interface in a weekend

When xAI released Grok bots, the Nous team liked how intuitive the interface was. Rolnick says they pointed Hermes Agent at it, reproduced the interface over a weekend and released the result as an open-source feature of Hermes Agent. The concrete comparison is the time it took another team to deliver a user experience it admired. Rolnick gives no performance comparison between the underlying systems.

He contrasts the weekend’s work with the economics of building SaaS. Assembling open-source libraries into a polished proprietary product used to require enough coordinated engineering labor to protect the result for a time. Agents also reduce the labor required by motivated open-source developers working in the other direction. A closed supplier can still introduce a desirable feature. Rolnick expects the interval during which it alone can offer that feature to shrink. He sees more durable advantages in network effects, reliability and reputation.

7:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:57 · section reference included

Distillation puts pressure on a model’s lead

Rolnick sees a parallel in model development. In distillation, a capable model supplies outputs that serve as supplementary training data for another model. In his account, the student can approach the teacher’s capability because much of the teacher’s advantage comes from data quality and training scale. A competitor with sufficient scale can learn from the teacher’s examples even without reproducing a proprietary architectural invention.

He describes two responses from labs. They characterize unauthorized distillation as theft, and they train models to decline assistance with training other models, a restriction they describe as safety. Rolnick challenges both framings. In particular, he sees a tension between claiming ownership over learning from a model’s outputs and having trained that model on people’s written and creative work. His argument concerns the incentives behind those positions. He offers neither a legal ruling nor a benchmark of student performance.

9:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:13 · section reference included

An investment decision shows what the agent coordinates

Rolnick recalls an earlier presentation about the inputs a financial firm brings to an investment decision. They include data sources, routine code, people’s opinions and judgment. An agent framework lets an application combine deterministic operations with reasoning whose outcome can vary. Routine code can process an input in a repeatable way. The application can then bring that result together with an analyst’s view when the decision calls for judgment.

This is what Rolnick means by comparing an agent to an operating system or an application framework. A personal assistant exposes the interaction a user has with it. The framework also determines how code, evidence and reasoning fit together underneath. His architectural test is whether the firm can swap the model while retaining control of that arrangement. A new model may reach a different judgment. The firm should still decide which sources and opinions it receives and where that judgment enters the investment process.

Five phrases stand out in white boxes against a blue paragraph. They highlight using local models, switching models mid-conversation, running offline, reading the prompt, and exporting the agent.Open full source frame
What you should be able to do with your agent
11:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:06 · section reference included

Keep the investment process and inherit the framework improvements

Ownership creates a maintenance question. A financial firm has reasons to develop its investment process that do little to justify spending its engineering effort keeping a general-purpose agent framework current. Rolnick’s proposed division of labor lets the firm build increasingly sophisticated applications from code, data and reasoning while inheriting improvements to the common machinery. The firm can concentrate on the expertise that made it successful as the state of the art in agents changes.

Hermes Agent is Nous’s offer on those terms. Rolnick says customers can choose where and how to run it, modify it, and keep using it if they end their relationship with Nous. He wants it to become a standard and describes it as the most used and downloaded open-source agent at the time, with far more downloads than some closed competitors. He gives no counts or measurement method. The commercial proposition is more concrete. Customers who hold the software themselves can still pay Nous for support, safety and help operating at scale. Nous then has to earn their continued business through the quality of that work.

A network of glowing points branches above a small figure. The Hermes statement is centered, with GitHub stars and repository rank at lower left and user and clone figures at lower right.Open full source frame
Hermes is the open standard for agents.
12:50
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:50 · section reference included

Transcript download
Full transcript

Loading transcript…