The following is from a Radical Ventures AI Founders Masterclass session with Databricks Chief AI Scientist, Jonathan Frankle.
Artificial intelligence continues to scale at extraordinary speed, but the questions shaping its future are no longer about capability alone. As models grow larger and more powerful, attention is shifting toward a different set of constraints: efficiency, evaluation, and the systems required to deploy AI responsibly at enterprise scale.
In this Masterclass conversation, Radical Partner Vin Sachi joined Databricks Chief AI Scientist, Jonathan Frankle, to explore how those constraints are evolving and where meaningful progress is still possible. Drawing on his journey from academic research, to co-founding Mosaic ML, through its $1.3B acquisition, and into a leadership role at Databricks, Frankle offers a clear-eyed view of what actually determines success in applied AI today.
Across the discussion, he reflects on why early efficiency breakthroughs now feel “quaint,” how evaluation has emerged as the field’s missing infrastructure, and why enterprise AI adoption is less constrained by talent than by confidence, measurement, and proof.
From Academia to Entrepreneurship: A Practical Shift
Frankle’s path into entrepreneurship was not premeditated. During his PhD, he had viable options in both academia and industry, and initially resisted the idea of joining a startup in any serious capacity. Mosaic ML began as an advisory role rather than an entrepreneurial endeavor.
What ultimately changed his trajectory was the experience of building. As the company took shape, the feedback loops of entrepreneurship: customer needs, operational constraints, and measurable outcomes, made the work feel immediate in a way academic research rarely does. In hindsight, Frankle frames startups as a powerful vehicle for applied science: a different kind of research environment, one that forces clarity around what actually matters.
When Efficiency Research Becomes “Quaint”
Frankle is notably candid when reflecting on his early work on the Lottery Ticket Hypothesis, which explored whether smaller, pruned neural networks could be trained efficiently from the outset. At the time, the work drew significant attention amid hardware scarcity, rapid model scaling, and a broader effort to understand why deep learning worked as well as it did.
In hindsight, much of that era’s efficiency research now feels “quaint,” Frankle notes. Sparse training proved difficult to operationalize at scale, and identifying so-called “winning tickets” was computationally expensive. Ideas that once felt controversial, particularly the emphasis on empirical validation over theory, have since become widely accepted.
The broader lesson is not that the work was misguided, but that the field moves quickly. Research impact often lies less in preserving specific techniques than in reshaping how practitioners think about constraints, evidence, and trade-offs.
Scaling After Acquisition: Learning to Use Leverage
One of the most instructive moments in the conversation centers on Mosaic ML’s transition into Databricks following its $1.3B acquisition. The shift was not incremental; it was a step change in the operating environment.
Mosaic ML moved from a small, highly fluid organization to becoming part of a multi-thousand-person enterprise with formalized product, sales, finance, and procurement functions. Frankle likens the experience to going from riding a bike to being handed a high-performance sports car: the leverage is enormous, but only if you learn how to drive it.
The early months were challenging. Academic training does not prepare leaders to navigate enterprise processes, large sales organizations, or structured performance systems. Over time, however, learning how to work within that infrastructure unlocked scale and velocity that would have been impossible as a standalone startup.
For founders, the takeaway is clear: scale is not just amplification, it is transformation.
Open Source, Proof, and Scientific Credibility
Frankle also addresses Databricks’ approach to open-source releases, emphasizing that these decisions are rarely ideological. Instead, they are grounded in first principles: enabling customer success and advancing the broader ecosystem.
In many cases, open source serves as a credibility mechanism. In technical fields, bold claims are insufficient. The research community expects evidence, reproducibility, transparency, and results that can be examined independently. Open-sourcing models, code, or weights is often the most direct way to demonstrate that a capability is real.
Frankle is unequivocal on this point: serious scientific communities operate on proof, not rhetoric. In a landscape crowded with hype, receipts matter.
The Real Bottleneck: Evaluation as Missing Infrastructure
Perhaps the strongest theme of the session is Frankle’s argument that enterprise AI struggles are widely misunderstood. He pushes back against the idea that enterprises fail to deploy AI because they lack talent or sophistication. In his experience, many large organizations are deeply informed, technically fluent, and actively engaged with the state of the art.
The real bottleneck, he argues, is evaluation.
Building a demo is easy. Determining whether a system is reliable enough to deploy, especially for complex, agentic workflows is far harder. While researchers often advise enterprises to “build evals,” Frankle notes that the field itself lacks mature methodologies for many real-world tasks.
As a result, companies are being asked to invent evaluation frameworks under tight timelines, with real financial and reputational risk. Until evaluation becomes better-supported infrastructure, deployment will remain slow and uneven, regardless of improvements in model capability.
Customer Success as the Only Durable Strategy
Frankle closes with advice that cuts across technical and organizational boundaries: customer success is the only durable foundation for building AI companies.
Research excellence, technical sophistication, and compelling demos matter, but only insofar as they enable real users to succeed. Sustainable products emerge from deeply understanding customer constraints, defining what “good” means in production, and building systems that can meet that bar consistently.
Frankle’s perspective offers a useful reframing for founders and technical leaders navigating AI’s next phase. The hardest problems are no longer about building models, they are about earning confidence, defining standards, and delivering outcomes that hold up in production.
In an ecosystem increasingly saturated with capability claims, lasting advantage will come from rigor, proof, and a deep understanding of real-world constraints. The companies that succeed won’t be the ones that move fastest in theory, but the ones that move most deliberately in practice.
This post is based on insights from Radical Talks, a podcast from Radical Ventures exploring innovation at the frontier of AI. For more conversations with leaders in AI, subscribe wherever you get your podcasts.