Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 4 min read

Meta AI Updates: November 13, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Helion Picks Kernel Configs at Compile Time, Beating Hand-Written Triton and CUDA on H100 Without Autotuning

PyTorch published details on November 12 of compile-time heuristics in Helion, its DSL for portable ML kernels, that choose block sizes, num_warps, num_stages and related settings from a kernel’s IR and dataflow instead of running the autotuner. Kernels are classified as matmul, reduction or pointwise, and a liveness analysis estimates register, shared memory and (on Blackwell) tensor memory use per CTA to set the config. Across 96 linear-attention kernel-and-shape cells on H100 the computed config is 1.18x faster than hand-written Triton in Flash Linear Attention on the forward pass and 2.13x faster than Helion’s previous no-autotune default. Across 153 cells from six vLLM reduction kernels it is 1.27x faster than vLLM’s handwritten CUDA/C++ operators. The motivation is tuning time: one paged-attention kernel across 72 shapes previously took 10 hours in quick mode. The analysis adds tens of milliseconds to compilation, and Helion falls back to the old default when extents are not known at compile time. Source

2. EUROPA Consortium Pretrains a 951.9B-Parameter MoE on the JUPITER Supercomputer With TorchTitan

A PyTorch blog post from the EUROPA consortium (Domyn, Fraunhofer IAIS and IIS) describes scaling pretraining of a 951.9B-parameter Mixture-of-Experts model, with 32.8B parameters active per token, 74 layers and 384 routed experts, on the publicly funded JUPITER system for the EuroHPC Frontier AI Grand Challenge. Throughput rose from 90 to about 1,300 tokens per second per GPU on 384 NVIDIA GH200 GPUs (96 nodes) per data-parallel replica, at a sequence length of 65,536 tokens. The main changes were FlexShard placement for the Muon optimizer so each rank owns whole experts and runs Newton-Schulz without communication, running Newton-Schulz on the GPU while parameters stay offloaded (optimizer step from 41.7 s to 2.2 s), offloading activation checkpoints to the Grace CPU, and an unshard_async_op change that frees about 14 GiB of HBM per GPU. The consortium contributed these pieces back to TorchTitan and PyTorch. The post covers training throughput only and reports no model quality results. Source

3. Meta’s Muse Agent Can Now Place Phone Calls to US Businesses for Paid Subscribers

Meta made outbound calling available to all paying Muse subscribers in the United States on November 12, according to TechCrunch, which cites a post on X by chief AI officer Alexandr Wang. The agent phones businesses on a user’s behalf for tasks such as restaurant reservations or calls to insurers and cable providers. The feature had been in limited testing since September. Muse plans start at $20 per month for 500 million tokens per week, with a $100 tier at 3 billion tokens. Wang also said Muse email addresses will roll out to free users soon. Whether Muse discloses to the business that the caller is an AI is [NEEDS VERIFICATION], as the report does not say. Source