01 / ACCESSIBLE MULTIMODAL AI

Google brings sign-language translation from the lab to Pixel

Google DeepMind introduced SL2T, a multilingual sign-language-to-text model trained on more than 100,000 hours across over 50 sign languages. The model now powers ASL-to-English dictation in Gboard and Live Transcribe on Pixel 11, letting users sign searches, messages, documents, and Gemini requests. An on-device tracker converts video into pose landmarks and discards the original footage before sending coordinates for server-side translation.

02 / SEMICONDUCTOR WORKFORCE

Lam and NY Creates take virtual chip fabrication into universities

Lam Research and NY Creates announced a program intended to train about 3,500 university students in semiconductor process integration over five years. Participating colleges in the Northeast will use Lam's SEMulator3D software to model etch, deposition, lithography, and full fabrication flows without needing a physical cleanroom. NY Creates will coordinate delivery across its workforce-development network and align curricula with regional advanced-manufacturing programs.

03 / EDGE VISION MODELS

Liquid AI targets on-device vision with a 3B-parameter model

Liquid AI released LFM2.5-VL-3B, a vision-language model designed for screens, documents, grounding, multi-image input, and function calling on local hardware. The company says the roughly 3.1 billion-parameter model fits in about 3 GB of memory and reaches 20 tokens per second on a Galaxy S26 Ultra. It ships with support across llama.cpp, MLX, vLLM, SGLang, and ONNX.

04 / AI TRAINING DATA POLICY

Twitch makes creator content available for Amazon AI training by default

Twitch added a setting that allows channel content to train Amazon generative AI models, with creators required to switch it off if they do not want future use. The scope includes streams, VODs, clips, chat, images, and channel text. Twitch says disabling the setting does not affect captions, recommendations, safety tools, or other AI-supported platform features.

05 / EARTH OBSERVATION MODELS

Ai2 opens custom Earth-observation embeddings for downstream analysis

Ai2 added custom embedding computation and export to OlmoEarth Studio. Users can select a geography, time range, encoder size, resolution, and Sentinel imagery, then download int8 Cloud-Optimized GeoTIFFs for similarity search, few-shot segmentation, change detection, or unsupervised exploration. The models, weights, and source code are public, while Studio also supports supervised fine-tuning for applications that need higher task-specific performance.

06 / AGENT PAYMENTS

AWS details an auditable payment stack for autonomous agents

A Solv Labs workflow on Bedrock AgentCore gates each agent payment with policy checks, enclave attestation, risk pricing, and an audit record before settlement, completing transactions in under four seconds.

07 / INFERENCE INFRASTRUCTURE

AWS extends LLM KV cache across GPU replicas

AWS combined HyperPod routing with a shared Curvine NVMe tier, reporting up to 100% cross-pod cache hits and a 2.7-times improvement in time to first token in its test deployment.

08 / SOVEREIGN ENTERPRISE AI

OneAdvanced runs 50-plus agents on UK-sovereign infrastructure

OneAdvanced self-hosted Llama 4 Maverick and Llama Guard 4 on SageMaker and used the Strands SDK to ship more than 50 specialized agents across regulated business workflows.

09 / AI COST GOVERNANCE

Bedrock spending can now be traced to users and applications

AWS showed how CUR 2.0, IAM principal data, Athena, and CUDOS 5.8 can attribute Bedrock model costs to individual callers, teams, projects, and tenants.

10 / AI CODING TOOLS

GitHub highlights voice input and remote sessions in its Copilot app

GitHub's Copilot app guide shows developers connecting local folders or repositories, selecting models and agents, dictating prompts by voice, and moving sessions to the web through remote control.