Dan Biderman

Dan Biderman is an AI and computational neuroscience researcher and Engram co-founder. His collaborative work includes studying the learning–forgetting tradeoff in language-model finetuning and designing cost-efficient collaboration between on-device and cloud models. On his personal website, he describes his research focus as “building language models that dynamically learn from experience.”[2][3][5][6]

Latent Space identified Biderman as Engram CEO in its July 14, 2026 announcement of his Food show appearance. The episode was titled “The AI Memory Problem: Why Long Context Isn’t Enough”. Engram’s June 23, 2026 launch announcement described a strategy of starting with strong pretrained models and spending training compute on users’ context. Its stated goal was “a single training algorithm that can absorb arbitrary amounts of data into a model that gets continually better.” The company also acknowledged that making such learning work at scale over repeated updates remained an open problem.[1][2][4]

Before Engram’s 2026 launch, Biderman coauthored research on two practical questions underlying efficient AI systems: how adaptation affects retained capabilities, and how smaller local models can share work with frontier cloud models.[5][6]

Biderman was the first-listed author of LoRA Learns Less and Forgets Less, submitted in May 2024 and accepted to Transactions on Machine Learning Research that August with Featured Certification. His collaborators were Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, Cody Blakeney, and John P. Cunningham. Biderman describes the work as a collaboration with Databricks Mosaic AI.[5][3]

LoRA saves memory by training low-rank changes to selected weight matrices. Comparing it with full finetuning on programming and mathematics, Biderman and his collaborators found that standard low-rank configurations substantially underperformed full finetuning in the target domains while better preserving the base model’s performance outside them. The team also found that LoRA mitigated forgetting more than weight decay and dropout and helped preserve more diverse generations. These results exposed a tradeoff between acquiring new capabilities and retaining existing ones in the configurations studied.[5]

In February 2025, Biderman coauthored Minions: Cost-efficient Collaboration Between On-device and Cloud Language Models with first-listed author Avanika Narayan, Sabri Eyuboglu, Avner May, Scott Linderman, James Zou, and Christopher Re. The team studied financial, medical, and scientific reasoning over long documents, with a small local model reading the full context and communicating with a frontier cloud model. Simple back-and-forth collaboration struggled because the local model had difficulty following multistep instructions and reasoning over long contexts. Their MinionS protocol instead had the remote model divide tasks into easier subtasks over shorter document chunks, which local models executed in parallel.[6]

In the paper’s evaluations, the simple protocol reduced remote inference costs 30.4-fold but recovered only 87% of frontier-model performance. MinionS reduced those costs 5.7-fold on average while recovering 97.9% of remote-only performance. The reported gains concerned cloud inference costs in the studied tasks; the protocol traded some of the larger savings for higher performance.[6]

Biderman’s earlier work applied machine learning to biological data. He lists a 2018–2024 PhD in computational neuroscience at Columbia University, where he says he worked with John Cunningham and Liam Paninski, following a 2018 MA in cognitive science at Tel Aviv University. He describes Lightning Pose, a deep learning system for tracking animal movement in video, as his main PhD project. His website identifies its publication in Nature Methods in 2024 and describes combining semi-supervised pose estimation, physically informed constraints, deep ensembles, and state-space models to improve tracking.[3]

In his undated personal account, Biderman describes a Stanford postdoctoral appointment co-advised by Christopher Ré and Scott Linderman and summarizes his work as building resource-efficient AI systems and applying them to neuroscience. He also lists co-organizing the Efficient Systems for Foundation Models workshop, most recently at ICML 2025.[3]