Hey, I'm Dan (people also sometimes call me Danny or Daniel). I’m a researcher working on agentic, multimodal, and efficient AI systems among many other things.

My research began at MIT, where I led some of the earliest efforts to study and design more energy-efficient AI systems (“Green AI”) for large models at scale. This work spanned algorithmic efficiency, energy-aware hardware and algorithms, and system-level optimization on MIT’s high-performance supercomputing infrastructure.

Previously, I was a senior researcher scientist at Microsoft working on agentic AI, Copilot and other models, and agentic evaluations, including some of the first comprehensive benchmarks for multimodal computer-use agents. I’ve also worked on AI for scientific applications at places like SandboxAQ (formerly GoogleX), NYU, Twitter, and others. Earlier, I worked across government, international trade, finance, healthcare, drug discovery, gaming, as well as other industries.

My research areas include:

• Agentic AI: building, training, and evaluating agents that reason, plan, and act in realistic environments for general multimodal computer use [1][2][3], social simulation and social media platforms [4][5], agentic and AI safety + security [6], cybersecurity [7], and more.
• Efficient AI: developing new ways to holistically improve the efficiency of large-scale AI models and systems across hardware, systems, algorithms, training, and inference.
  • Hardware, Systems, & Energy Efficiency: studying the compute, energy, and performance tradeoffs of LLMs and other large models at scale for distributed training and inference [8][9], system-wide GPU power-capping at scale [10], the efficiency and incentive structures of AI in society [11], etc.
  • Algorithmic & Model Efficiency: improving training and inference efficiency with new techniques for faster architecture search and ranking [12][13], efficient representation learning [14][15], model compression/sparsification [16], quantization-aware training [17], etc.
• AI for Science: developing ways to accelerate scientific workflows like those in quantum chemistry [18][19], quantum machine learning and computing [20][21], battery/materials research [22][23], etc.