2025 is the Year of Evals! Just like 2024, and 2023, and …

How can you monitor an AI application if you cannot measure its behavior? For John Dickerson, that is the common foundation of monitoring and evaluation: observability depends on measurements, and evaluation supplies ways to make them.

Dickerson approaches the problem from six years as co-founder and chief scientist of Arthur AI, spanning observability, evaluation and security across traditional machine learning, generative AI and agents. At the time of this talk, he had become CEO of Mozilla.ai, whose mission was to support open-source AI tooling and give the open-source community a voice in AI’s direction. Here, however, his subject is the evaluation market—and his prediction that its long-anticipated growth is finally arriving.

Dickerson identifies three pressures that brought evaluation beyond the technical organization and onto the executive agenda:

• Accessible capabilities. ChatGPT gave CEOs, CFOs and chief information security officers a direct experience of AI. They did not need to understand its internals to see possible applications.
• Constrained budgets. In his account, late-2022 recession fears froze ordinary enterprise spending just as leaders were setting budgets for the following year. Discretionary exceptions remained possible, and generative AI became an executive-sponsored exception.
• Systems that act. AI was beginning to act for people and teams, rather than merely supply inputs to another application. That change made its behavior more consequential—and harder for decision-makers to overlook.

Together, he argues, these pressures explain the growing demand seen by evaluation vendors.

An agent can make decisions and carry out several steps toward an action, either autonomously or with human involvement. Dickerson explicitly retains people in the loop as a good choice for many systems. But deployment across enterprises, smaller businesses and personal projects means evaluation is already an operational concern, not something to postpone until full autonomy arrives.

The contrast with earlier ML applications is organizational as well as technical. A model might produce a number that a larger application consumes. The application’s complexity then obscures the model’s contribution: technical teams know the output matters, but the evaluation problem remains inside the CIO’s organization. An agent acting on someone’s behalf makes that responsibility much more visible.

Before ChatGPT’s November 30, 2022 launch, data-science teams already used statistical methods to monitor ML systems. The difficulty was connecting those measurements to downstream business KPIs: dollars saved, dollars earned or another outcome that could justify a purchase. In Dickerson’s sales experience, executive enthusiasm about AI’s return on investment often left the actual buying conversation concentrated in the CIO’s organization.

This was an established market. He traces an early generation through H2O, Algorithmia and Seldon; specialist vendors through companies such as WhyLabs, Arize, Arthur, Galileo and Fiddler; and broader platforms through Snowflake, Databricks, Datadog and the major cloud providers. Later entrants such as Braintrust extended that landscape. The missing ingredient was not awareness among practitioners.

In purchasing discussions, model behavior could still lose to security, latency or other familiar technology concerns. Buyers acknowledged that monitoring mattered without making it the immediate priority.

Dickerson recalls a recurring pitch-deck prediction from the mid-2010s onward: this would be the year a CEO lost their job over an ML failure. He says he knew of no such case. That observation captures the gap between a technically credible risk and one that senior executives personally felt accountable for.

He illustrates the scale question with Jamie Dimon’s letter in JPMorgan Chase’s 2021 annual report, released in April 2022. Dickerson describes its consumer-business example as $100 million invested in AI/ML from 2017 through 2021, which he considers small relative to JPMorgan’s size.

Editor: The cited amount covers consumer fraud-risk systems using AI, ML and other technology initiatives; it is not JPMorgan’s total AI spending. The letter separately describes annual AI investment in the hundreds of millions of dollars. That broader figure materially qualifies the spending comparison. See the 2021 shareholder letter.

The next part of Dickerson’s account begins with enterprise budget setting in October and November 2022. He connects recession fears to frozen or reduced IT budgets for 2023. Without ChatGPT’s arrival, he thinks those constraints would have left much less room for new IT projects.

The combination was bittersweet: ordinary spending was constrained, but a small discretionary project could attract extraordinary attention. His image for that concentration is the Eye of Sauron turning toward one project. Generative AI became that project.

The timing mattered. ChatGPT arrived just before the holiday breaks, with a single web interface that let nontechnical executives experience the technology themselves. Dickerson recalls Arthur and its Series A investor, Index Ventures, hosting an event at NeurIPS on launch night. Even that room of machine-learning specialists became absorbed in trying it.

Playful writing requests made the capabilities accessible beyond that audience: have Eminem rap like Taylor Swift, or ask for poetry in a rapper’s voice. These were demonstrations people could understand immediately, without a technical presentation.

In Dickerson’s account, that experience unlocked discretionary money for CEO-sponsored generative AI work. The resulting 2023 enterprise experiments grew inside a wider environment of austerity. This is his explanation of the adoption pattern he observed, rather than a claim that every enterprise allocated its budget the same way.

By 2024, Dickerson was seeing generative AI applications enter production, primarily for internal use. Examples included chat applications and hiring tools. Deployment brought a different set of stakeholders into the conversation, asking about return on investment, governance, risk, compliance and brand consequences.

Those questions require more than an impressive demonstration. A quantitative estimate of risk requires evaluation. Once an application is used in real work, its behavior becomes relevant to people outside the machine-learning team and the CIO’s office.

His sequence is experiments in 2023, production in 2024, then shipping and scaling in 2025. Later budgets could earmark substantial spending for AI instead of relying on exceptions to a freeze. He reads rising usage and provider revenues as signs of that transition.

Improving models, open-source participation and investment from venture capital and large technology companies reinforce the momentum in his account. But the third pressure remains essential: systems are moving toward autonomy. The evaluation demand comes not just from more AI usage, but from AI taking on more responsibility.

Dickerson’s short description of an agent includes perceiving an environment, learning, abstracting and generalizing, then reasoning and acting. He contrasts that reasoning-and-action capability with the earlier picture of a model supplying an output to a larger application. The action might occur in a virtual environment or in a cyber-physical system, where software interacts with the physical world.

That expands both complexity and risk. Evaluation therefore needs a connection to the outcomes a buyer cares about: mitigating risk, increasing revenue or reducing losses. An isolated model score is not the end of that explanation. The measurement must help establish what the system’s behavior means for the business.

Executive alignment does not mean everyone wants the same measurement. Dickerson maps evaluation to several distinct responsibilities:

The market thesis depends on these budget-owning roles arriving at a shared need to understand AI behavior, even though their reasons and purchasing processes differ.

As vendors move into agentic and multi-agent monitoring, Dickerson emphasizes the scope of the measurement: monitor the whole system, not only the model used by one particular agent. The system is what performs the work. Measuring one component does not cover the behavior of the larger process in which it participates.

He also offers a future check on his growth prediction. A mid-April article in The Information had reported leaked revenue figures for evaluation startups, including Weights & Biases, Galileo and Braintrust. Dickerson says those figures lagged by roughly six to eight months. Based on conversations with peers, he expected newer figures to show a stronger business picture.

His proposed test was to look at reporting in early 2026 and see whether evaluation-startup revenue still appeared to lag. That remains a forecast within the talk: executive attention, budgets and agent deployment are his reasons to expect growth, not a substitute for observing it.

Before the questions, Dickerson returns briefly to Mozilla.ai. He distinguishes it from the evaluation vendors discussed throughout the talk and introduces any-agent, presented at the time as an open-source, non-monetized project.

His analogy is LiteLLM for multi-agent systems: a unified interface over multiple agent frameworks. The suggested use is straightforward—try different frameworks through that shared interface. The coda concerns making framework experimentation easier, rather than prescribing a deployment architecture.

An audience question makes the evaluation problem concrete: suppose a multi-agent system produces a discounted cash flow spreadsheet, a valuation based on discounting future cash flows to their present value. Who determines whether it did the work correctly?

The question is about domain expertise. Familiar evaluation methods for structured ML outputs do not automatically settle quality judgments when a workflow interprets unstructured information and performs specialist work. Producing something that resembles a human analyst’s output is not sufficient evidence of correctness.

Dickerson first points to the limits of persona-based agents, invoking his Nature Machine Intelligence work on the risks of using language models as substitutes for human participants. Prompting a model to act as a farmer in Ohio in their mid-forties may have value, but it does not perfectly reproduce that person’s perspective. A role description cannot simply stand in for the expertise or experience the evaluation needs.

His practical answer is to put experts alongside the system. Dickerson recalls a leaked Mercor spreadsheet listing expert rates around $50, $100 and $200 per hour, and offers Google, Meta and large banks as examples of customers. The point of the anecdote is the willingness to pay for expensive human validation in lockstep with agent work.

He likens the agent to an intern who may eventually change the expert’s job, then backs away from a simple prediction of replacement. For the immediate evaluation decision, the stakes matter: when an incorrect result could cause substantial financial losses or serious professional consequences, costly expert review can be justified. What happens over the following five years, as those expert judgments become data incorporated into the systems, remains an open question in his answer.

That investment can also create competitive value. Dickerson places particular weight on dataset creation and environment creation: building the task-specific material and setting in which a system can be evaluated. A strong DCF evaluation environment can become an asset that competitors lack. In this view, expert validation is not only an operating expense for checking today’s outputs; building the evaluation capability can be a capital investment.

The final question asks when evaluations will be driven primarily by generative AI or language models. Dickerson does not give a date. Instead, he points to a practice already in use: LLM-as-a-judge, where a language model assesses another system’s output. Adoption is happening despite limitations in the judgments.

He refers to his ICLR 2025 work on differences between LLM and human judgments, Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking, while discussing qualities such as conciseness and helpfulness.

Editor: In the ICLR 2025 study, the tested judges favored stylistic features, and their preferences did not track the study’s concrete safety, knowledge and instruction-following measures. That finding concerns the evaluated judges and benchmark setting, rather than every judge or task.

Persona prompting offers an imperfect substitute for a human judge and can partly ease the problem of creating evaluation data. Dickerson describes it as a practical crutch: it makes judgments easier to obtain, but does not establish that they are reliable.

That brings the answer back to the domain-expertise question. The judge’s outputs still need validation against human judgment, with attention to whether the evaluation is moving in a biased direction. Automating the assessment does not remove the need to check the assessor.