Synthesia Enters Live Coaching with "AI Roleplay Sessions"

Synthesia, the AI video generation platform, has launched "AI Roleplay Sessions," an interactive corporate training feature. Going beyond its core video production business, the feature lets employees practice real workplace communication by conversing with AI avatars.

With built-in feedback, scoring, and analytics, companies can now quantitatively measure training effectiveness. AI-powered corporate training is seeing growing demand, particularly in remote work environments, and this move — extending Synthesia's video production expertise into interactive experiences — is drawing attention.

Poolside's Laguna XS.2 / M.1 Come to llama.cpp — Local MoE Options Expand

New models have been added to Poolside's Laguna series, with llama.cpp support advancing through community-driven efforts.

Laguna XS.2**: A MoE model with 33B total parameters and 3B active parameters. Built to run on local machines, it adopts Sliding Window Attention in 30 of its 40 layers to deliver fast inference and a small KV cache

Laguna M.1**: 225B total parameters and 23B active parameters, aimed at larger-scale agentic coding

Both are designed for "agentic coding" and "long-horizon tasks," and they sit in the middle range of the lineup around the Laguna S 2.1 (118B-A8B). With llama.cpp support, the options for experimenting with MoE architectures in local environments have expanded even further.

In addition, the community has shared a proposed chat template fix for the Laguna S 2.1 that forcibly enables `` tags during inference — part of ongoing efforts to squeeze out better performance on engineering tasks.

SkewAdam — Cutting Optimizer Memory for MoE Training by 97.4%

SkewAdam, which has been attracting attention on Hugging Face Daily Papers, proposes a method to dramatically reduce optimizer state memory — the biggest bottleneck in MoE (Mixture-of-Experts) training.

The mechanism is simple: it hierarchically assigns different precisions based on the gradient statistics of each component — the dense backbone, the experts, and the router (tiered state allocation). As a result, optimizer state memory is cut by 97.4%, from 50.6 GB to 1.29 GB, while overall peak training memory is compressed from 81.4 GB to 31.3 GB.

The authors state that this makes it possible to train multi-billion-parameter MoE models within a standard 40 GB memory budget, establishing a scalable training strategy for deeper architectures without sacrificing validation perplexity.

Running Kimi 2.7 (204 GB) on a Single DGX Spark — VRAM and Disk Cache Strategy

The LocalLLaMA community has shared a method for efficiently running Kimi K2.7 Code — a model with 1 TB total parameters and 32B active parameters — on a single DGX Spark.

The key is a strategy that treats VRAM as the top few layers of the disk cache and keeps MoE expert weights on the CUDA compute path. Specifically, it leverages the following combination:

`GGML_CUDA_ENABLE_UNIFIED_MEMORY`: provides a unified address space across GPU and CPU

`GGML_OP_OFFLOAD_MIN_BATCH`: forces GPU computation even for small batches

Regex-based CPU offload configuration for experts (the final layer is kept in VRAM)

With this configuration, it achieves 340 pp/s prompt processing and 9.6 tg/s token generation. Improved token generation has also been confirmed on a dual-3090 setup, making this a useful reference for memory-hierarchy optimization approaches that take advantage of MoE characteristics.

Codeberg Proposes Terms of Service Update Banning LLM Data Extraction

At Codeberg, the Git hosting platform, an extension to the terms of service banning data extraction by LLMs (LLM-extrusions) has been submitted as a pull request.

Platforms remain divided over the use of open source code as training data. Codeberg's community-driven move reflects growing wariness of large-scale data extraction that runs counter to creators' intent, and it is drawing attention as one indicator of where the open source ecosystem may be heading.

"Post-hoc Design" for Runaway AI Agents — How to Roll Back Contaminated Memory

An article on Qiita highlighting the importance of "post-hoc design (rollback)" in dealing with runaway AI agents has been attracting attention.

Safety measures for AI agents tend to focus on "preemptive defense" (how to stop them), but when an agent writes misinformation into shared memory, stopping the process does nothing to halt the activities of other agents that have already read that information. The article discusses the importance of architectural design for breaking this "chain of contamination," offering a practical perspective for running agent systems in production.

---