The Missing Architecture Behind Autonomous AI with Jacob Buckman
In this episode of Machine Minds, we step beyond today’s transformer-dominated AI landscape and into a deeper conversation about what’s missing on the path to truly autonomous, long-horizon intelligence. Jacob Buckman, co-founder and CEO of Manifest AI, joins Greg to explore why current AI systems struggle with long-term reasoning, persistent memory, and extended task execution—and what it will take to unlock the next paradigm.
Jacob’s journey into AI began early, fueled by science fiction, programming, and a fascination with building systems that could do meaningful work autonomously. From studying and conducting research at Carnegie Mellon to working at Google Brain, he watched deep learning unify once-fragmented AI subfields—vision, language, speech—under a single scalable framework. That unification shaped his conviction that the next breakthrough wouldn’t come from incremental tuning, but from rethinking a fundamental architectural bottleneck.
At Manifest AI, Jacob and his team are tackling what they believe is the missing piece: scalable long-context intelligence. Their work centers on replacing transformer attention with a new family of architectures called retention models, designed to compress and retain relevant information over time—rather than repeatedly replaying massive histories. The goal: AI systems that can reason, learn, and work continuously over hours, days, or longer.
In this conversation, Greg and Jacob explore:
Jacob’s path from aspiring scientist to AI researcher and founder—and why curiosity plus first principles thinking matter more than trends
Why today’s large language models excel at short tasks but break down over long horizons
The core limitation of transformer attention—and why “attention is all you need” may no longer hold
How retention architectures unify the strengths of transformers and recurrent neural networks
What it means for an AI system to compress knowledge instead of endlessly appending memory
Why long-term reasoning, iterative problem solving, and true autonomy require architectural change—not orchestration hacks
The misconception that agent orchestration can substitute for unified, persistent intelligence
How long-context models could reshape agents from short-lived “consultants” into persistent, personalized collaborators
The technical challenge of translating theoretical breakthroughs into high-performance GPU kernels
Why Manifest AI is open source—and how their work aims to move the entire field forward
Lessons from unifying AI subfields, the “bitter lesson” of scale, and avoiding ad-hoc solutions that won’t last
Jacob’s view on cost, intelligence density, and why better architectures will increase—not reduce—investment in AI
Advice for founders and researchers: focus relentlessly on the single bottleneck that matters most
If you’re building AI systems, researching foundations of intelligence, or trying to understand what comes after today’s models, this episode offers a rare, deeply reasoned look at where the field may be heading—and why architectural simplicity could unlock far more than brute force scale.
Learn more about Manifest AI: https://manifestai.com
Explore the open-source retention models: pip install retention
Connect with Jacob Buckman on LinkedIn: https://www.linkedin.com/in/jacobbuckman
Connect with Greg Toroosian on Linkedin: https://www.linkedin.com/in/gregtoroosian