Jonathan Frankle
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,389 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,389 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jonathan Frankle is an American computer scientist who serves as Chief AI Scientist at Databricks, where he leads the company's Mosaic Research team. He is best known for the lottery ticket hypothesis, a finding about neural network pruning and sparsity that he and his doctoral advisor Michael Carbin introduced in a paper that won a Best Paper award at the 2019 International Conference on Learning Representations (ICLR). Frankle joined Databricks through its 2023 acquisition of MosaicML, the generative AI startup where he was a member of the founding team and its chief scientist, and his research group went on to build the open large language model DBRX. [1][2][3]
| Field | Detail |
|---|---|
| Known for | Lottery ticket hypothesis; neural network sparsity and efficient training |
| Education | BSE and MSE, Princeton; PhD, MIT (2023) |
| Doctoral advisor | Michael Carbin |
| Current role | Chief AI Scientist, Databricks (Mosaic Research) |
| Notable honors | ICLR 2019 Best Paper; 2023 AAAI/ACM SIGAI Doctoral Dissertation Award |
Frankle studied computer science at Princeton University, earning a Bachelor of Science in Engineering (BSE) and a Master of Science in Engineering (MSE). His interests there ranged across programming language theory, distributed systems, and computer security before he moved toward machine learning. [1][10]
He then pursued a PhD in computer science at the Massachusetts Institute of Technology (MIT), working in the Computer Science and Artificial Intelligence Laboratory (CSAIL) under the supervision of Michael Carbin. He completed the doctorate in 2023. His dissertation, "The Lottery Ticket Hypothesis: On Sparse, Trainable Neural Networks," received the 2023 AAAI/ACM SIGAI Doctoral Dissertation Award, given each year for the best doctoral thesis in artificial intelligence. During graduate school he interned at Google Brain and at Facebook AI Research. [2][10]
The lottery ticket hypothesis is the result that established Frankle's reputation in deep learning research. In "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks," first posted in 2018 and presented at ICLR 2019, Frankle and Carbin showed that a randomly initialized dense neural network contains a much smaller subnetwork that can be trained in isolation to reach the same accuracy as the full network in a comparable number of steps. They called these subnetworks "winning tickets." The metaphor gives the hypothesis its name: training a large network resembles buying many lottery tickets at once, only a few of which turn out to be winners. [3]
The winning tickets are found through iterative magnitude pruning. A network is trained, its smallest-magnitude weights are removed, the remaining weights are reset to the values they held at initialization, and the procedure repeats. Using this approach the authors discarded more than 90 percent of the weights in standard image-classification networks while preserving accuracy, evidence that the trainable structure was already present at initialization rather than created during training. The paper won one of the two Best Paper awards at ICLR 2019 and prompted a large body of follow-up work on sparsity and model compression. [3]
Frankle's later research refined and stress-tested the idea. With collaborators he introduced a technique often called weight rewinding, which resets pruned weights to their values from an early point in training rather than to their original initialization. That change allowed the method to scale to larger networks such as deep residual networks, and it linked pruning to questions about the loss landscape and the early phase of training. This line of work shaped how researchers and practitioners think about the relationship between a network's size and its trainability. [2][3]
In 2021 Frankle helped start MosaicML, a startup built to make the training of large neural networks faster and cheaper. The company was founded by chief executive Naveen Rao and Hanlin Tang, who had previously worked together at the AI chip startup Nervana Systems (later acquired by Intel), with Frankle joining as chief scientist and a member of the founding team. MosaicML released open source training software and a family of open large language models called MPT, including MPT-7B and MPT-30B, which were downloaded millions of times and positioned the company as an alternative to closed model providers. [6][8][10]
In June 2023 Databricks, the data and AI company led by Ali Ghodsi, agreed to acquire MosaicML in a deal valued at roughly $1.3 billion, a figure inclusive of retention packages for the team. Databricks announced the definitive agreement on June 26, 2023, and completed the acquisition on July 19, 2023. Frankle became Databricks' Chief AI Scientist, a role sometimes styled chief scientist for neural networks, and continued to lead the research group, which Databricks operates as Mosaic Research. [4][5][6]
In March 2024 that team released DBRX, an open large language model built on a fine-grained mixture of experts architecture. DBRX has about 132 billion total parameters, of which roughly 36 billion are active for any given input, and it uses 16 experts from which it selects 4. It was pretrained on approximately 12 trillion tokens of text and code. At release, Databricks reported that DBRX outperformed open models such as Meta's Llama 2, Mistral's Mixtral, and xAI's Grok-1 across a range of language, programming, and mathematics benchmarks. [7]
As of 2026 Frankle remains Chief AI Scientist at Databricks, leading a Mosaic Research lab of more than 30 scientists whose work spans reinforcement learning, model training, and the evaluation of AI agents. Methods and products associated with his team include TAO, a tuning approach introduced in 2025 that uses reinforcement learning and synthetic data to improve models without large amounts of labeled data, and Agent Bricks, an agent-building framework unveiled at the Databricks Data and AI Summit in June 2025. [13][14][15]
Before his machine learning career, Frankle worked at the intersection of technology and public policy, and he has continued that work alongside his research. He was the inaugural Staff Technologist at the Center on Privacy and Technology at Georgetown University Law Center, where he contributed to "The Perpetual Line-Up," a 2016 report on American law enforcement's use of face recognition that drew wide attention to the technology's accuracy gaps and civil liberties implications. As an adjunct professor of law at Georgetown, he co-developed and taught a course on computer programming for lawyers with Professor Paul Ohm. [1][12]
Frankle has remained engaged in AI policy. He works with the Organisation for Economic Co-operation and Development (OECD) to help implement the OECD AI Principles, the intergovernmental standard on trustworthy AI adopted in 2019, and he regularly engages with lawyers, journalists, and policymakers on issues such as AI governance, copyright, and privacy. [1][11]
He also retains ties to academia. Frankle was named an assistant professor of computer science at Harvard University's John A. Paulson School of Engineering and Applied Sciences and was affiliated with its Kempner Institute for the Study of Natural and Artificial Intelligence, but he chose to build MosaicML rather than take up a conventional faculty post. He remains connected to Harvard and the Kempner Institute as an affiliate and advisor. [6][9]