Here are today's news for 24th of October about IBM, Anthropic, Google and Meta. Let's dive into the details. On October 20, 2026, llm, d contributors announced version 1.0 at PyTorch Conference North America, consolidating an open, source, Kubernetes, native stack for distributed large, language, model inference. IBM, Red Hat, Google Cloud, CoreWeave and NVIDIA are among the project’s contributors. The release adds priority, and fairness, aware controls for sharing capacity across tenants, router hot standby and rolling upgrades, coordinated updates for separate prompt, processing and response, generation replicas, and a stable v1 configuration API. Session, aware routing and turn, priority scheduling aim to reuse context across agent conversations rather than repeatedly processing it. In a reported test, llm, d served up to 3,000 concurrent coding agents with zero preemptions and peaked above 6.6 million output tokens per minute. On October 20, Anthropic said an internal Claude research model designed minibinders, small proteins intended to attach to targets, against five scorpion neurotoxins. Over 64 hours, it generated 318,484 designs; Adaptyv Bio and Twist Bioscience tested 211, of which 17 met Anthropic’s binding criteria. The hits targeted four toxins from scorpions found in Mexico, North Africa and the Middle East; none qualified against the Brazilian toxin Ts1. The strongest two, aimed at Mexico’s Cn2 toxin, showed single, digit nanomolar binding affinity. This is a notable step for AI, assisted drug discovery: a model helped move from computational designs to laboratory, tested candidates, potentially giving researchers new starting points for antivenom work. But binding is not neutralization. Anthropic has not shown that the proteins block toxins in cells or protect animals, and says it has not produced an antivenom. On October 20, 2026, Google and Meta announced TorchTPU, an open, source PyTorch backend for Google Cloud TPUs, at PyTorch Conference North America. It makes TPU a native PyTorch device, aiming to let developers run existing workloads with minimal code changes rather than porting them to a separate framework. The project was tested across about 5,000 Hugging Face models and integrates with Transformers v5, Diffusers, TorchTitan, vLLM and SGLang; in Transformers, `device_map="tpu"` selects the hardware. Google reports TorchTitan training Qwen 3 8B at up to 52% model FLOPs utilization across as many as 8,192 Ironwood TPU v7x chips. That is a company, reported training result, not a direct price or speed comparison with Nvidia GPUs. The release could lower switching costs for teams already using PyTorch and widen Google’s appeal to AI developers, while strengthening competition among accelerator platforms. On Oct. 22, Bloomberg reported that OpenAI was seeking about $30 billion in fresh funding at a $1.25 trillion valuation, down $150 billion from the $1.4 trillion discussed earlier. CNBC separately reported the round could reach $30 billion and would be led by United Arab Emirates investors alongside Coatue. The proposed figure is pre, money, before the new capital, and negotiations were ongoing, with terms subject to change; OpenAI reportedly aimed to secure investor signatures by the end of the following week. The financing is described as an opportunistic bridge ahead of a potential IPO, not a completed deal. A $1.25 trillion valuation would put OpenAI among the most valuable private companies and intensify pressure on rivals, including Anthropic, as both pursue capital and public, market ambitions. The scale also reflects the costs of competing in frontier AI: model development and deployment require substantial computing investment. On October 12, Meta introduced Muse Wonder, a new model for its Muse personal AI agent, which launched September 8. Meta says Wonder is its most advanced model for digital assistants, built to improve agents’ performance on practical tasks such as booking dinner reservations and checking availability. Access to Wonder through the Muse app requires a paid plan: Power costs $20 a month and Maximum $100. Meta AI chief Alexandr Wang described the model, codenamed Watermelon, as designed to make agents better at personal and professional work; he said API access would be priced “meaningfully below” leading OpenAI and Anthropic models. A more powerful version, Muse Wonder Full, is limited to developers using Muse Code, with plans at $50 and $200 monthly. The launch extends Meta’s push from its free, usage, limited agent, which can shop, plan travel and handle email, toward paid, more capable AI services. That's all for today's AI news briefing. Have a great day!