Embodied Intelligence: Teaching agentic AI and robots to think and move like us

In the ongoing race to scale generative AI, one truth has hardened into strategic consensus: large language models are no longer the endgame. The next leap, agentic AI, requires models that not only interpret the world but move through it, make judgments, adapt, and take autonomous action. This is not a shift in capabilities. It is a shift in infrastructure.

While tech giants double down on model size and compute, a quieter foundation is being laid elsewhere. The companies reshaping AI’s training stack—Cogito Tech, Turing, Snorkel AI, and Surge AI—are emerging not just as data vendors, but as research accelerators. They are building the data supply chains, human-AI collaboration loops, and evaluative systems needed to teach models how to behave, not just answer.

The transformation is not hypothetical. It is already underway. And it is being forced into the open by a $15 billion shock to the system: Meta’s acquisition of a 49 percent stake in Scale AI.

Cogito’s Bet on Agentic AI

These companies, especially Cogito Tech, Snorkel AI, and Turing, are shaping the next layer of AI infrastructure. Cogito Tech in particular, founded in 2011, has worked through the autonomous vehicle wave, the geospatial AI cycle, and is now scaling rapidly—with 500% growth over the past 30 months—as demand for agentic systems and structured human feedback accelerates.

For CEO Rohan Agrawal, this isn’t about labor. It’s about engineering aligned behavior.

“AI that acts cannot be trained on disconnected or crowd-sourced labels. It needs context, judgment, and expert feedback aligned to the real-world environments it will operate in. That’s what we build,” he said.

Cogito’s core business is centered on reinforcement learning with human feedback. Unlike legacy suppliers that scale through low-cost labor, Cogito specializes in data pipelines where nuance, interpretability, and traceability define success. It deploys localized teams—German legal annotators for German court models, Filipino clinicians for Southeast Asian health systems, financial experts for regulatory training environments.

That operational structure is backed by DataSum, Cogito’s traceability and governance layer.

“We are focused on traceable, verifiable data. That is how we maintain both performance and trust,” said Agrawal.

DataSum: Building Trust Through Traceability

With more scrutiny around how AI systems are trained, transparency and accountability are becoming foundational requirements. In 2024, Cogito launched DataSum, the certification framework that documents every step of the data process—from workforce composition to tool usage to ethics compliance.

DataSum allows Cogito’s clients to answer questions that are increasingly asked by regulators, investors, and internal safety teams. Where did this data come from? Who labeled it? Under what conditions? And how do we know it reflects the goals we claim?

“You cannot scale agentic AI if you cannot explain how it was trained. DataSum gives our partners that confidence,” Agrawal said.

This need for traceable, auditable training data has only grown in urgency following Meta’s $15 billion acquisition of a 49 percent stake in Scale AI, once considered a neutral backbone of the data labeling industry. The move has prompted several large AI developers—including Google, Microsoft, OpenAI, and xAI—to reevaluate their data strategies and shift work to providers seen as more independent and aligned with their safety and product goals. Cogito, along with a handful of others, is emerging as one of those alternatives.

Meta-Scale: The Collapse of Neutral Ground

Scale AI, long dominant in the human labeling market, took in $870 million in revenue last year by scaling massive crowdsourcing pipelines. It won business from nearly every leading AI lab. But its value was built on being a neutral infrastructure layer. That neutrality collapsed overnight when Meta took nearly half the company and moved Scale’s CEO into its new superintelligence lab.

The response from the rest of the ecosystem was immediate. Google is moving hundreds of millions in data work elsewhere. OpenAI began stepping back months ago. Microsoft and xAI are planning exits. Several of Scale’s own employees expressed concern that Meta would gain visibility into prior client work, even under strict contract firewalls. As one former employee put it bluntly, “They all want to cut Scale off now.”

What’s left is not just a vacuum. It is a new playing field.

The Rise of the Research Accelerators

Turing, Snorkel AI, and Cogito Tech, now sit in a different category: research accelerators. They are not vendors executing outsourced labeling at scale. They are strategic partners embedded inside RL pipelines, multimodal capture systems, and agent training environments.

This is not a shift in branding. It is a shift in function. Research accelerators help labs generate frontier-aligned data: behavioral feedback for LLM agents, high-resolution context for STEM problem-solving, multimodal annotations in 3D environments, and domain-specific RL gyms. This data cannot be crowd-labeled. It requires orchestration, iteration, and human-machine collaboration.

Cogito’s DataSum, for example, provides a granular audit trail of how every dataset was built. It shows who labeled it, what tools were used, how quality was measured, and what ethical standards were applied. In an era of rising regulatory oversight, auditability is no longer optional.

Embodied Intelligence and the Stakes Ahead

As LLMs are embedded into physical systems—autonomous vehicles, hospital diagnostic agents, defense networks—the demand for behavioral fidelity grows. It is no longer enough to predict the next word. Systems must be trained to interpret feedback, learn from mistakes, and operate safely in complex, unpredictable environments.

In this context, the commodity model that Scale pioneered has reached its limit. Labs need partners who can run tight iterations on high-complexity tasks. They need data vendors that operate like product teams, not pipelines.

That is what separates research accelerators from data vendors. Firms like Cogito, Turing, and Snorkel are no longer in the business of selling annotations. They are building the operating systems of AI development.

A Reordered Ecosystem

The Meta-Scale transaction lit the fuse, but the reordering of the ecosystem was already in motion. AI labs no longer want clean data alone. They want strategic collaboration. They want partners who help design test cases, spot hallucination patterns, and create multi-turn simulations. They want neutrality, agility, and infrastructure that can evolve with model design.

As AI systems grow more agentic, the human-AI interface becomes a research problem, not an ops function. And those who can engineer that interface—data vendors who became research accelerators—are now core to the generative AI stack.

This is not a branding shift. It is an architectural one. The next leap in AI capability will not come from model scaling alone. It will come from how and with whom we teach models to behave.

Agrawal summarized it this way: “Data is no longer an input. It’s a control surface. If you want alignment, you have to design for it. And that begins at the training set.”