Please welcome Scaled Cognition, the world’s first agentic pre-trained transformer
Meet Scaled Cognition, the latest addition to Lerer Hippeau’s portfolio.
HQ: New York, NY
Founder: Dan Roth, Dan Klein, and Damon Pender
The problem: A future in which AI agents help individuals and businesses operate is hurtling toward us, but agentic technology needs to progress to help facilitate that future. AI agents take actions in the real world, like processing insurance claims, booking flights, making reservations and appointments, or issuing refunds — simply put, a near-future state in which agents make tasks easier is coming soon. But in order to do that, agentic technology needs to be more accurate, reliable, and consistent than current LLM technology. While certainly powerful, existing LLMs are still fundamentally occupied with decontextualized language scraped from the web. They are built to predict linguistic tokens, not actions, which are at the center of agentic behavior.
Scaled Cognition’s solution: Scaled Cognition is developing a new generation of rational, controllable AI models that can be deployed as domain experts for grounded, real-world applications. The company has introduced a new modeling paradigm, the Agentic Pretrained Transformer, which includes synthetic data generation, action modeling, training innovations, and evaluation solutions, all specialized for agent-based workflows. The company’s first system, APT-1, is designed to optimize around action prediction rather than token or language prediction. The agent system is trained on a synthetic data pipeline that doesn’t exist on the internet and which includes not only conversations but also actions tied to those conversations, allowing it to take actions correctly and aligned to existing policies or instructions. The system is also reinforced by what is known as “self play” in which it learns from agent-to-agent simulations before interacting with real humans. Scaled Cognition has also launched an Agent Builder platform that allows developers to create enterprise-grade agents in less than an hour.
Origin story: Scaled Cognition is the fourth company Dan Roth has founded, and the second with Dan Klein, the company’s CTO and an AI professor at UC Berkeley. The team has been working on agentic technology for decades. Previously, Roth, Klein, and Pender founded Semantic Machines, a pioneer in agentic AI which was acquired by Microsoft in 2018. The team reunited to found Scaled Cognition with the idea that a synthetic data training set could enable an LLM to make the leap from token processing to actions.
Why we’re betting on it: Scaled Cognition’s APT-1 system and Agent Builder platform have been developed on top of a foundation of research breakthroughs that set it apart from its competitors. Standard LLMs are extremely good at one thing: language-based exchange. But the leap to action is tricky. APT-1 operates more deterministically and makes significantly fewer hallucinations, allowing it to outperform on agentic applications and workflows. To do this, the team had to invent completely new synthetic agentic training data that teaches the system how to take natural language conversations and business logic and use them to predict complex actions. Already, the APT-1 system outperforms existing foundation models on the two most challenging agentic benchmarks, Tau-Bench and ComplexFuncBench. APT-1 is a significant step forward in the development of agentic AI, and we couldn’t be more excited to partner with the team at Scaled Cognition as they introduce their agent models to the real world.
Our take: We believe the future will be one in which individuals and enterprises are assisted on a daily basis by agents. Whether you are booking your next doctor’s appointment, filing taxes, or developing new backend code, we think agents will become integral to the processes involved in completing human tasks. Scaled Cognition’s APT-1 system is an important and meaningful advancement toward this future reality.
Further learning: Hear from CTO Dan Klein on Scaled Cognition’s approach
