Alex Dimakis

Alex Dimakis, also known as Alexandros G. Dimakis, is a UC Berkeley EECS professor and co-founder and CSO of Bespoke Labs. His AI research includes collaborative work on DataComp and DataComp-LM, which make training-data curation a controlled experimental problem, and generative models that learn from corrupted observations. He joined Berkeley as a professor in January 2025.[2][4][3][5][6][7]

Dimakis co-founded Bespoke Labs, which he describes as a startup working on data curation and data-centric AI. He serves as CSO alongside CEO Mahesh Sathiamoorthy. The company develops reinforcement-learning environments and optimization systems for reliable agents.[2][3]

His stated ambition for open reasoning models includes access to their training data. On his personal homepage, he points to DeepSeek-R1's unreleased reasoning data and describes collective work curating the OpenThoughts reasoning dataset and training OpenThinker, “aiming to make the best open-source (open-weights and open-data) reasoning models.” This is an ambition for the work, rather than an achieved ranking.[2]

In 2023, Dimakis co-authored DataComp with a large collaboration including Samir Yitzhak Gadre, Gabriel Ilharco, Georgios Smyrnis, Vaishaal Shankar, and Ludwig Schmidt. The benchmark lets researchers filter a candidate pool of 12.8 billion Common Crawl image-text pairs or curate new data sources, then compare datasets using standardized CLIP training and 38 downstream test sets. That structure makes dataset choices comparable under a common training procedure.[5]

The authors reported that their DataComp-1B baseline trained a CLIP ViT-L/14 model to 79.2% zero-shot ImageNet accuracy, exceeding OpenAI's CLIP ViT-L/14 by 3.7 percentage points with the same training procedure and compute. Dimakis also appeared with Gadre in the DataComp poster-session segment of Latent Space's December 2023 NeurIPS recap.[5][1]

He subsequently co-authored DataComp-LM, first submitted in June 2024, with collaborators including Jeffrey Li, Alex Fang, Georgios Smyrnis, Matt Jordan, Samir Gadre, Ludwig Schmidt, and Vaishaal Shankar. DCLM extends controlled dataset experiments to language models: researchers can test deduplication, filtering, and data mixing at scales from 412 million to 7 billion parameters, using OpenLM pretraining recipes and 53 downstream evaluations.[6]

In the April 2025 revision, the DCLM authors identified model-based filtering as key to their high-quality baseline dataset. They reported a 7-billion-parameter model reaching 64% five-shot MMLU accuracy after training on 2.6 trillion tokens—a 6.6-percentage-point improvement over MAP-Neo with 40% less training compute. These are results for the specified models and evaluation, illustrating the collaboration's emphasis on dataset design.[6]

Before these dataset benchmarks, Dimakis worked on using generative models to reconstruct signals from limited measurements. His 2017 compressed-sensing collaboration provides earlier technical context for his later research on inverse problems and learning from corrupted data.[8][2][7]

With Ashish Bora, Ajil Jalal, and Eric Price, Dimakis co-authored Compressed Sensing using Generative Models, submitted in March 2017. Instead of assuming that a signal is sparse in a chosen basis, the method assumes it lies near the range of a generative model. For an L-Lipschitz generator mapping a k-dimensional latent space into the signal space, the paper gives a recovery guarantee using roughly O(k log L) random Gaussian measurements. In demonstrations using published variational autoencoders and generative adversarial networks, the authors reported needing 5–10 times fewer measurements than Lasso for the same accuracy.[8]

In the NeurIPS 2023 paper Ambient Diffusion, Dimakis collaborated with Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, and Adam Klivans on a related but distinct problem: learning a clean distribution when only corrupted training samples are available. Their method adds further measurement distortion and trains the model to predict the original corrupted image from the further-corrupted version. Under technical conditions on the corruption process, including applicable inpainting and compressed-sensing processes, they prove that the model learns the conditional expectation of the full clean image.[7]

The Ambient Diffusion collaborators reported learning image distributions on CelebA, CIFAR-10, and AFHQ even when every training sample had 90% of its pixels missing. They also reported fine-tuning foundation models on small corrupted datasets, including MRI scans with block corruptions, to learn the clean distribution without memorizing the training set.[7]