Gary Bécigneul_
Safety research engineer, RL
One important bottleneck is mutable, forkable, accelerated RL environments, so you can explore many branches in parallel on expensive and slow infrastructure.
Writing
- Notes on building k7dnew a six-part series on building k7d, a Rust VMM that forks a live Kubernetes cluster in ~100 ms and accelerates guest time, starting with Part 1
Now
- Senior AI Researcher, Cisco Systems focusing on scaling loops and RL environments, and teaching AI how to operate networks at the frontier of network safety at scale
- Open Source Contributor
- k7 self-hosted infrastructure for lightweight VM sandboxes that safely execute untrusted code at scale, built on Kata, Firecracker, qemu, Longhorn, Cilium, Kubernetes, and k7d — motivated by AI agents running arbitrary code. #1 on Show HN, featured in Console.dev —
- k7dnew a purpose-built Rust VMM that forks running VMs and live Kubernetes clusters in ~100 ms and accelerates guest time, for RL training and agent evals on byte-identical environments — selected paths verified with Kani and Aeneas→Lean —
- k7 self-hosted infrastructure for lightweight VM sandboxes that safely execute untrusted code at scale, built on Kata, Firecracker, qemu, Longhorn, Cilium, Kubernetes, and k7d — motivated by AI agents running arbitrary code. #1 on Show HN, featured in Console.dev —
Selected
-
across NeurIPS, ICML, ICLR, AAAI, and AISTATS pioneered early work in geometric deep learning — hyperbolic embeddings, Riemannian optimization
- Lead AI Research / Engineering contractor Neural Fabric, Sciemo, LayerLens, YZR
- Max Planck Society Prize, 2020 youngest outstanding PhD across all subjects, 86 institutes
- Supervisor of 6-month full-time MSc theses at ETH Zürich/MIT My alumni include 3 at Google DeepMind, 1 at Meta Superintelligence Labs, 1 at Apple, and an MIT alum/IMO Gold Medalist at Citadel Securities
Background
- Applied ML Scientist, Thomson Reuters building LangGraph agents in the legal domain
- Co-founder, Gematria quant/fintech venture with Louis Abraham; signed an institutional asset manager as a client
- Massachusetts Institute of Technology drug discovery with graph neural networks, leveraging optimal transport and contrastive estimation — funded by DARPA & MLPDS, co-authored with Prof. Tommi Jaakkola, Prof. Regina Barzilay
- Max Planck Institute for Intelligent Systems COVID group testing methods leveraging belief propagation networks and genetic algorithms, co-authored with Prof. Bernhard Schölkopf, Alex Smola
- PhD, Computer Science — ETH Zürich thesis: On the geometry of data representations; under Prof. Thomas Hofmann (ETH), examiners Prof. Gunnar Rätsch (ETH), Prof. Tommi Jaakkola (MIT), Prof. Regina Barzilay (MIT)
- MASt (Part III), Pure Mathematics — University of Cambridge thesis on statistical applications of persistent homology; focused on advanced topics such as Morse theory and Schramm–Loewner evolutions
- Prix Fermat Junior — Mathematics, 2015 awarded once every 2 years in France, for contributions to mathematics made as an undergraduate