engblogs

Designing domain-specific enterprise environments to train and evaluate stateful AI agents in realistic workflows using tools, policies, and grounded evaluation.

MIT appoints Sasha Rakhlin as director of the Statistics and Data Science Center (SDSC), highlighting his interdisciplinary leadership across statistics, machine learning, and artificial intelligence and his mission to advance rigorous foundations, education, and collaboration at MIT.

Alexander Rakhlin named director of MIT's Statistics and Data Science Center (SDSC), uniting statistics, data science, and AI across IDSS, LIDS, and MIT departments with a commitment to interdisciplinary training, rigorous foundations, and AI-enabled scientific discovery.

Private Connect provides a private, auditable, and scalable connectivity layer that removes the security barrier to Agentforce and Data 360 adoption across multi-cloud environments (including Hyperforce), enabling regulated industries to deploy enterprise AI securely and at scale.

Private Connect provides a private, auditable, end-to-end connectivity layer that removes governance barriers and accelerates Agentforce and Data 360 adoption across multi-cloud environments.

Unpacks how Meta's GEM training achieves 2x end-to-end MFU and 4x training FLOPs growth at LLM-scale by co-designing kernels, precision, parallelism, networking, and memory (including JFA, GDPA, BlockAttention, MXFP8, and 5D parallelism) to scale thousands of GPUs for ads foundation models.

Meta's GEM training doubles end-to-end MFU to 20-25% and scales training FLOPs 4x over 12 months through holistic co-design of kernels, precision, parallelism, networking, and memory for an LLM-scale ads foundation model.

Explains how Deep Latent Particles extend self-supervised, object-centric learning to 3D by decomposing RGB-D and voxel data into interpretable particles that drive diffusion-based robotic control and cross-modal understanding.

3D-DLP enables self-supervised, object-centric 3D scene representation learning by decomposing colored RGB-D voxel data into interpretable particles using appearance-aware clustering and a diffusion-based manipulation policy for robust robotic control.

AI-enabled research tooling accelerates discovery across mathematics and theoretical computer science, presenting ten advances—from high-dimensional geometry to lattice cryptography—enabled by AI models and Lean-certified proofs.

Ten advances across geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography and extremal combinatorics, showcased through AI-assisted research, Lean-formalized proofs, and open problems in mathematics and theoretical CS.

Survey of ten advances at the convergence of mathematics and theoretical computer science, highlighting AI-assisted research, formalization in Lean, and progress on long-standing open problems across geometry, coding theory, arithmetic circuit complexity, group theory, quantum complexity, and cryptography.