Anthropic says Chinese AI labs Moonshot, DeepSeek and Alibaba secretly trained models on millions of Claude exchanges, exposing AI industry impact.

Stanford AI Lab's Marin 535B-A23B began pretraining on 18.75T tokens with 11 x GB200 NVL72 for 2.7e24 FLOPs under Percy Liang.

Moonshot Kimi K3 AI model trained on 20,000 NVIDIA NVDA chips via Alibaba cluster, per Bloomberg report on July 31 2026.

According to Paolo Ardoino, Tether AI has unveiled the QVAC Fabric LLM, integrating the world's first cross-platform BitNet LoRA framework. This innovation enables billion-parameter AI model training and inference on consumer GPUs and smartphones, utilizing Vulkan and Metal backends for AMD, Intel, Apple Metal, and mobile GPUs. By combining BitNet's efficient ternary weight compression with LoRA's parameter reduction, QVAC Fabric drastically reduces memory and compute requirements, achieving GPU inference speeds 2 to 11 times faster than CPU while using up to 90% less memory. This advancement signals a breakthrough in local, private AI accessible on consumer devices.

According to the source, Tether's QVAC has launched the BitNet LoRA framework, enabling large-scale AI training on consumer GPUs and smartphones. This development aims to make AI training more accessible and cost-effective, potentially democratizing AI technology and expanding its adoption across diverse user bases.

According to Paolo Ardoino, Tether AI has launched the latest version of QVAC Fabric, introducing the world's first cross-platform BitNet LoRA framework. This innovation enables billion-parameter AI training and inference on consumer-grade GPUs and smartphones, significantly reducing memory usage and computational requirements. By supporting heterogeneous GPU vendors and mobile GPUs, QVAC Fabric allows efficient AI model fine-tuning and inference across diverse hardware, marking a major leap toward accessible and private AI solutions.

According to Paolo Ardoino, Tether AI has launched the groundbreaking QVAC Fabric, incorporating the world's first cross-platform BitNet LoRA framework. This innovation enables billion-parameter AI training and inference on consumer GPUs and smartphones. Leveraging BitNet's memory-efficient one-bit architecture and LoRA's parameter-reduction techniques, QVAC Fabric supports heterogeneous GPU platforms, including AMD, Intel, Apple Metal, and mobile GPUs. This breakthrough dramatically enhances AI accessibility and efficiency, with flagship smartphones achieving up to 11x faster GPU inference compared to CPUs and requiring 90% less memory than traditional models.

According to @DeepLearningAI, a new short course has been launched in collaboration with Google, focusing on building and training a MiniGPT-style large language model (LLM) using JAX, the open-source library behind Gemini. This course enables participants to construct and train a 20M-parameter LLM architecture from scratch, offering valuable insights into advanced AI model development.

According to @MRRydon, a groundbreaking development has unlocked full neural network training capabilities, including backpropagation, directly on Apple's Neural Engine (ANE) without relying on CoreML, Metal, or GPUs. This innovation enables ultra-efficient AI model training on M4 chips, achieving 6.6 TFLOPS per watt and significantly reducing costs and environmental impact. The Zero-Human Company is already leveraging this to advance autonomous AI systems, marking a transformative shift in AI training from cloud-based GPUs to local Mac devices.

According to @FuSheng_0306, robotics entrepreneurship is surging because it is easier to raise funding and technically less demanding at the prototype stage. According to @FuSheng_0306, a startup in Shenzhen can assemble a humanoid robot for roughly 2 million RMB, add branding, and secure investment if the unit can walk a few steps. According to @FuSheng_0306, foundation model ventures face far higher thresholds, including around 100 billion parameters, substantial GPU purchases, months of training, and strong benchmark rankings, with weak scores making fundraising difficult. According to @FuSheng_0306, these fundraising dynamics currently favor tangible humanoid robot prototypes over foundation model efforts.

According to @AveryChing, large-scale AI training on KGeN IO using audio and video datasets is generating significant on-chain revenue on Aptos, placing the chain 9th by app revenue; source: @AveryChing on X. For traders, the reported strength in Aptos app revenue tied to AI workloads highlights APT as a token to monitor for ecosystem usage and sentiment shifts; source: @AveryChing on X.