Dmitry Kislyuk

Work

Currently, I am a director of engineering at Pinterest, leading the visual AI team as part of the Advanced Technologies Group. Our team works on visual search products, large-scale representation learning (e.g. multimodal embeddings and VLMs), diffusion models, core computer vision signals (e.g. detection, segmentation), agentic search, and more. I lead a mix of ML research, ML engineering, and product teams. Before that, I was at Stanford University for my masters, and U.C. Berkeley for my undergraduate CS education.

You can reach me at hi@dkislyuk.com.

Not work

Outside of my day-job, I enjoy tinkering on open-ended algorithms (as inspired by the work of Ken Stanley, Jeff Clune, etc.), amateur running, and deep-diving into history {podcasts, books, essays}.

Regarding open-ended algorithms, some of the questions that I ponder include:

• What is the minimum amount of structure needed for evolutionary algorithms to work at modern scale? Deep learning has indicated that human-designed components keep getting more high-level; that hints that classic EA concepts such as mutations, crossover, selection, etc. should not be hardcoded, but instead discovered as part of an open-ended algorithm.

• Is there a place for evolutionary algorithms in the current paradigm of frontier models? Supervised, unsupervised, and reinforcement learning allow for optimization of arbitrary feedforward and recurrent computation graphs at extreme scales, but one problem that still has not been elegantly solved is an automatic design of the computation graphs themselves and the modification of the computation graph at test-time. One viewpoint is that the mega-scale models have so much capacity, that every relevant permutation of a computation circuit to achieve an objective can be discovered inside and selected during test-time. The Hardware Lottery makes it daunting to seriously explore non-uniform, EA-designed computation graphs, but it is an interesting question nevertheless.

• To what extent can agentic systems replace various EA components: for example, can LLM-guided permutations of computation graphs result in a more interesting open-ended computation system?