A conversation with Manifest AI’s Jacob Buckman on Machine Minds, hosted by Greg Toroosian

If you have spent any time working with today’s AI agents, you have probably run into the same wall over and over. They are brilliant for the first few minutes of a task, and then somewhere around hour two, or even minute twenty, they start losing the plot. They forget what you told them earlier. They repeat steps. They lose track of the goal. It is not a bug you can patch away. According to Manifest AI’s co-founder and CEO, Jacob Buckman, it is a structural problem baked into the very architecture that powers nearly every large language model on the market today.

Jacob sat down with Samson Rose’s founder, Greg Toroosian, on Episode 118 of the Machine Minds podcast to talk about what he calls the “missing piece” of autonomous AI. It is a conversation that goes well beyond the usual hype cycle chatter about bigger models and more parameters. Instead, it digs into something much more fundamental: the way today’s systems are built might be the very thing standing between us and AI that can actually work independently over long stretches of time.

From Science Fiction to Google Brain

Jacob’s path into AI research started the way many great origin stories do: with curiosity sparked by science fiction and a fascination with building systems that could do meaningful work on their own. That curiosity took him to Carnegie Mellon, where he studied and conducted research, and eventually to Google Brain, one of the most influential AI research labs of the last decade.

While at Google Brain, Jacob had a front-row seat to something remarkable. Deep learning was quietly unifying fields that had once operated almost completely separately. Vision, language, and speech recognition all started converging under a single scalable framework. Problems that used to require entirely different toolkits and different research communities were suddenly being solved with the same underlying approach.

That experience shaped a conviction that stuck with him. The next real breakthrough in AI was not going to come from another round of fine-tuning or another clever prompt trick. It was going to come from rethinking a fundamental bottleneck in how these systems are architected. That belief eventually led him to co-found Manifest AI.

The Bottleneck Nobody Wants to Talk About

Here is the part of the conversation that will make a lot of builders sit up straight. For years, the AI world has operated under the mantra that attention is all you need, a reference to the paper that introduced the transformer architecture and effectively launched the current generation of large language models. Transformers are the backbone of everything from chatbots to coding assistants to image generators.

Jacob argues that this mantra may no longer hold. Transformer attention, the mechanism that lets a model weigh the importance of different pieces of information, works beautifully for short tasks. Ask a model to summarize an email or answer a quick question, and attention shines. Ask it to reason through a multi-day project, retain context across dozens of steps, or work continuously on a complex goal, and the wheel starts to come off.

The reason comes down to how attention actually functions. It essentially replays the entire history of a conversation or task every single time it needs to make a decision. As that history grows, the computational cost balloons and the model’s ability to pick out what actually matters starts to degrade. It is like a little like trying to remember an important detail from a meeting three weeks ago by rereading every single email you have sent since then. Technically possible, but wildly inefficient, and eventually unworkable.

Enter Retention Models

This is where Manifest AI’s work comes in. Jacob and his team are building what they call “retention models,” a new family of architectures designed to replace transformer attention. Instead of endlessly appending and replaying massive histories, retention models compress and retain the relevant information as they go.

Think of the difference between someone who takes detailed notes and periodically distills them into a clean summary versus someone who just keeps a running transcript of everything they have ever heard and tries to scan through it whenever they need an answer. The first approach scales, and the second approach collapses under its own weight.

Jacob explained that retention architectures are designed to unify the strengths of transformers and recurrent neural networks, an older style of architecture that processes information sequentially and keeps a running internal state. The goal is a system that gets the best of both worlds: the parallel processing power that made transformers so successful, combined with the persistent, evolving memory that recurrent networks were always better suited for.

The payoff, if this approach works at scale, is significant. AI systems that can reason, learn, and operate continuously over hours, days, or longer, without needing to be reset or without ballooning compute costs every time the task grows.

Why Orchestration Hacks Are Never the Answer

One of the more pointed parts of this episode deals with a popular workaround in the industry right now, which is agent orchestration. A lot of companies are trying to solve the long horizon problem by stitching together multiple AI agents, each handling a small piece of a larger task, with some kind of coordination layer managing the handoffs.

Jacob is skeptical that this approach gets to the root of the issue. He argues that orchestration can paper over some of the symptoms of the problem, but it cannot substitute for a model that has genuine, unified, persistent intelligence backed into its core architecture. Chaining together a bunch of short-lived specialists is a clever workaround, but it is still a workaround. The real fix, in his view, requires architectural change, not more scaffolding on top of a system that was never designed to remember.

This distinction matters a lot for how the AI agent conversation could evolve. Right now, most agents behave like short-term consultants. You bring them in for a specific job, they do it, and then they are gone, with no lasting memory of the relationship. Jacob’s vision is one where long context models could turn agents into something closer to persistent, personalized collaborators, systems that actually build a relationship with the work and the people they support over time, rather than starting from zero every session.

The Hard Part Nobody Sees

It would be easy to treat this as a purely theoretical conversation, but Jacob was candid about just how difficult it is to translate these architectural breakthroughs into something that actually runs fast and efficiently on real hardware. Getting a new architecture to work in a research paper is one challenge. Getting it to run as a high-performance GPU kernel, the low-level code that determines whether a model is fast enough and cheap enough to actually be useful in production, is an entirely different and much harder engineering problem.

This is part of why Manifest AI has chosen to open source their retention model work. Jacob's reasoning is refreshingly straightforward. If this really is the missing architectural piece the field needs, then the fastest way to prove it and improve it is to let the broader research and engineering community pressure test it, build on it, and push it forward together. Anyone curious enough to explore it firsthand can grab the code with a simple pip install. It is a bet on collective progress over closed-door competition.

Cost, Intelligence Density, and the Bitter Lesson

The conversation also ventured into territory that will matter to anyone thinking about where AI investment is headed. Jacob shared his take on what is sometimes called the bitter lesson in AI research, the idea that approaches which scale with more compute tend to win out over clever, hand-tuned, ad hoc solutions in the long run.

His view is that better architectures will not shrink AI investment; they will actually increase it. As systems become more capable of doing useful, sustained work, the economic case for investing in more compute and more sophisticated models only gets stronger. Intelligence density, essentially how much genuine reasoning capability you can pack into a given amount of compute, becomes the metric that matters, and Jacob believes retention-style architectures move that needle in a meaningful direction.

For founders and researchers listening in, his advice was refreshingly simple. Focus relentlessly on the single bottleneck that actually matters. It is tempting to chase every shiny new technique or trend, but the teams that make real progress tend to be the ones willing to sit with one hard, foundational problem long enough to actually solve it.

Why This Episode is Worth Your Time

This conversation is a genuinely rare thing in the current AI content landscape, a deeply reasoned, technically grounded discussion that does not lean on hype to make its point. If you are building AI systems or simply trying to understand what might come after today's transformer-dominated models, this episode offers a clear-eyed look at where the field could be heading, and why architectural simplicity might unlock far more than brute force scale ever could.

Episodes like this one are exactly why Machine Minds has become a go-to listen for people building at the edge of robotics, hard tech, and AI. The show is hosted by Greg Toroosian, founder and CEO of Samson Rose, a boutique retained search firm and trusted talent partner for robotics and hard tech companies. Greg brings a rare vantage point to these conversations. Because Samson Rose works so closely with founders and researchers building the next generation of physical and artificial intelligence, he knows exactly which questions to ask and where the real story is hiding.

Samson Rose was built on the idea that hiring in robotics, hard tech, and AI is fundamentally different from hiring in any other industry. It takes more than a database of resumes to fill these roles well. It takes people who genuinely understand the technology, the pace of the field, and the kind of talent it takes to bring ambitious, physically grounded ideas to life. Greg and his team position themselves as true search partners rather than transactional recruiters, working alongside clients to advise, support, and elevate the hiring process for both companies and candidates. That same curiosity and depth of understanding is what makes Machine Minds such a valuable listen for anyone trying to keep up with where AI and robotics are really headed.

If you are a founder building in robotics or AI and looking to bring on exceptional talent, or if you are a candidate looking to work on some of the most exciting hard tech problems out there, Samson Rose is worth a conversation. And if you want more episodes like this one, featuring the researchers and builders shaping the next era of intelligent systems, Machine Minds is available wherever you get your podcasts.

Deepen Your Understanding of Long-Context AI

Explore More from Machine Minds and Manifest AI:

Listen to the Full Discussion: Head over to Apple Podcasts for The Missing Architecture Behind Autonomous AI with Jacob Buckman here.
Analyze the Technical Architecture: Visit the official Manifest AI website to explore their research on retention models, open-source releases, and technical papers.
Connect with the Speaker: To follow insights on long-context architectures, AI research, and autonomous agents, look up Jacob Buckman on LinkedIn.

Why Your AI Agent Forgets Everything by Lunch: Inside the Architecture Race to Fix Long-Term Memory