Jonathan Frankle

Jonathan Frankle is Databricks’ chief AI scientist, a MosaicML founding-team member, and coauthor with Michael Carbin of the lottery ticket hypothesis. His work addresses how efficiently neural networks can learn and how to judge their usefulness. In a July 2026 interview, he described his research role as being “the risk budget” for Databricks: pursuing ideas that might become products, while accepting that some experiments fail.[7][2][3]

Frankle was MosaicML’s chief scientist during the May 2023 release of MPT-7B. He explained that building models was necessary to validate the tools customers would use to create their own models. The business centered on those customers’ models, rather than a single flagship model. Research scientist Abhinav Venigalla spearheaded the MPT-7B training run; Latent Space reported that the team trained it on one trillion tokens using 440 GPUs in 9.5 days, at approximately $200,000.[1]

Frankle worked with MosaicML’s data team on the training mixture. In the May 2023 interview, he described two selection criteria: achieve evaluation results at least as good as mC4 English, then prioritize diversity through code, scientific papers, and Wikipedia. He questioned assumptions about what made data valuable, including the claim that adding code necessarily improved language-model reasoning. These were his assessments of the evidence available then, and he acknowledged that the team’s evaluations did not fully measure useful open-ended generation.[1]

During model development, Frankle returned to Georgetown Law, where he had taught, and gave law students the actual candidate datasets, mixtures, and token quantities. He asked them to construct a training-data mix. He recalled that they identified the major trade-offs, including English versus multilingual coverage and whether to include code. The exercise surfaced choices about model behavior; it was not an account of students selecting the final mixture.[1]

Frankle was skeptical that multiple-choice benchmarks or text-overlap scores adequately captured the generation people wanted from language models. Alongside automated evaluations, the team watched answers to a fixed set of prompts during training. One asked for games a three-year-old and a seven-year-old could play together. He found changes in that answer personally more informative than the benchmark scores alone.[1]

For long-context modeling, Frankle explained the team’s preference for exact quadratic attention as a way to avoid the caveats of approximation methods they had not established would work in practice. Venigalla described how FlashAttention, ALiBi, training-stability improvements, and automated hardware-failure recovery supported the models. Frankle emphasized that reliable training had taken incremental work: earlier runs suffered catastrophic loss spikes and required overnight intervention when hardware failed.[1]

The StoryWriter demonstration also prompted a decision about release responsibilities. Frankle recalled briefly changing its license amid concerns about its Books3 training data, then retaining Apache 2.0. He said the team had not appreciated the timing of releasing a creative-writing model during the Hollywood writers’ strike. In hindsight, he might have chosen a different demonstration. He left copyright conclusions to lawyers while arguing for making the models available and letting users decide what they were comfortable using.[1]

Frankle connected training efficiency to the ability to investigate models scientifically: “We need to make it cheap enough that we can actually do science, and science involves running dozens of experiments.” He also argued that open research allowed ideas to circulate between laboratories and warned that reduced openness would slow innovation. These were explicit public commitments in his 2023 interview, alongside an acknowledgment that some company work would remain private.[1]

Before MosaicML’s language-model work, Frankle studied whether much smaller neural networks could train successfully. His lottery-ticket research provides the technical background to his efficiency questions. He completed his MIT computer-science PhD in 2023; that May, he recalled that working at MosaicML had delayed his defense by about two years.[2][1]

Frankle and Michael Carbin first submitted The Lottery Ticket Hypothesis in March 2018; it appeared at ICLR 2019. They proposed that dense, randomly initialized feed-forward networks contain smaller subnetworks—“winning tickets”—that can reach comparable test accuracy in a similar number of training iterations when trained in isolation. The original initial weights matter: the claim concerned a combination of retained connections and favorable initialization, rather than merely a smaller architecture.[3]

The paper reported winning tickets below 10–20% of the original size in several fully connected and convolutional architectures tested on MNIST and CIFAR10. This extended the question beyond shrinking an already trained network: could a small subnetwork also train effectively? The reported results concerned those vision experiments and did not establish equivalent savings for every network or immediate hardware speedups.[3]

With Gintare Karolina Dziugaite, Daniel M. Roy, and Carbin, Frankle investigated why initialization-time iterative magnitude pruning struggled on deeper networks. Their 2019 precursor explored subnetworks obtainable early in training rather than precisely at initialization. The successor, Linear Mode Connectivity and the Lottery Ticket Hypothesis, published at ICML 2020, connected successful subnetworks to stability under stochastic-gradient-descent noise, such as changes in data order and augmentation. In their experiments, full accuracy coincided with that stability: at initialization in small MNIST settings, but early in training for ResNet-50 and Inception-v3 on ImageNet.[4][5]

By May 2023, lottery-ticket-style sparsity was not MosaicML’s focus. Venigalla emphasized the difficulty of accelerating irregularly sparse weights on available hardware. Frankle added that imposing structure to make sparsity faster on standard hardware could reduce model quality. Their distinction was practical: removing weights and obtaining a useful training-speed improvement were separate achievements.[1]

Frankle joined Databricks through its acquisition of MosaicML. In July 2026, he described direct customer contact as essential to choosing research problems, including work on reinforcement-learning algorithms and custom models. His stated measure was concrete: “The true measure of success in AI is: did your customer succeed?” He endorsed avoiding customer lock-in, while crediting Databricks’ longstanding openness efforts to Ali, Matei, Reynold, and other company leaders.[2][7]

Frankle described testing internal Genie answers against knowledge he could verify. He asked about products whose numbers and development history he knew, and asked both an account executive and Genie the same questions about a customer before meeting them. He then asked the salesperson to identify correct answers and errors. These were his everyday evaluation practices, rather than reported controlled benchmark results.[7]

He also argued that giving an agent the same permissions as a human does not ensure comparable behavior. An agent can connect information and act at a scale, or in ways, that the person would never have time to attempt. In the 2026 interview, he called for fundamentally rethinking agent governance and explicitly declined to claim that his team had a complete solution.[7]

Frankle also worked across computing and law. His personal profile records his service as the inaugural Staff Technologist at Georgetown Law’s Center on Privacy and Technology, his contribution to The Perpetual Lineup, a report on police use of face recognition, and his collaboration with Professor Paul Ohm on Computer Programming for Lawyers. He also identifies himself as a participant in developing the OECD AI Principles in 2019.[2]