Issue W45/25: OpenAI's Trillion-Dollar Compute Plan and Agentic GPT-5, Moonshot AI Ships Kimi Linear - November 3, 2025
Last week, OpenAI restructured into a Public Benefit Corporation with a trillion-dollar compute roadmap and a private beta for its agentic GPT-5, while Moonshot AI released its Kimi Linear long-context model and Cursor launched its fast Composer-1 coding agent.
General News
OpenAI has restructured into a Public Benefit Corporation (PBC), altering its deal with Microsoft to gain more autonomy in exchange for a ~$250B Azure commitment. CEO Sam Altman outlined an ambitious roadmap including a $1.4T compute spending plan and a goal to build an automated AI researcher by 2028. In funding news, Poolside raised $1B at a $12B valuation, former X.AI researcher Eric Zelikman secured $1B for a new venture, and voice AI company Cartesia announced a $100M Series C. Elsewhere, a judge allowed a copyright lawsuit from George R.R. Martin against OpenAI to proceed, and Perplexity AI launched a patent research agent and other new finance features.
Google DeepMind
Google is partnering with Jio in India to roll out Google AI Pro plans, including Gemini 2.5 Pro, to eligible users. The company also introduced discounts for its Gemini API, with a 50% reduction for Batch API use and 90% for context caching. Google AI Studio now supports logs and dataset exports for evaluations. On the product front, the Veo 3.1 video model was updated, and an early access program for Gemini for Home was launched in the U.S.
Foundation Labs & Model Updates
Moonshot AI released its Kimi Linear model and technical report, a hybrid architecture using Kimi Delta Attention (KDA) that achieves up to a 75% KV cache reduction and 6x decoding throughput. The company also shipped a new terminal-native Kimi CLI and "Kimi For Coding" experience. OpenAI began a private beta for Aardvark, a GPT-5-powered "agentic security researcher" that finds and fixes code vulnerabilities, and also open-sourced gpt-oss-safeguard, a pair of safety classification models. In agentic coding, Cursor launched Cursor 2.0 with its fast, in-house Composer-1 model, while Cognition released SWE-1.5 (Windsurf), a model co-designed with Cerebras hardware for high-speed coding. Additionally, the Minimax M2 model is gaining traction for its strong coding performance and is now free to try. Anthropic published research on "signs of introspection in LLMs", and Cartesia launched Sonic-3, an SSM-based text-to-speech model.
AI Developer Topics
Hugging Face released the "Smol Training Playbook", a comprehensive 200+ page guide covering the entire LLM training pipeline. In agent frameworks, LangChain shipped a new Deep Agents CLI and a no-code Agent Builder in LangSmith, while VS Code introduced an Agent Sessions view to manage agents. The vLLM project released Sleep Mode for fast, zero-reload model switching in multi-model serving environments. Meanwhile, Confluent is partnering with vector databases like Weaviate and Qdrant to enable event-driven streaming agents.
Research News
A new paper sparked discussion by arguing that switching from BF16 to FP16 for RL fine-tuning can substantially reduce numerical divergence between training and inference. In agent research, the Agent Data Protocol (ADP) was introduced as a unified standard for SFT datasets. New benchmarks like Toolathlon and ScaleAI's Remote Labor Index revealed that top agents still struggle with complex tool use and real-world tasks. Liquid AI released LFM2-ColBERT-350M, a multilingual late-interaction retriever, while an analysis from Epoch AI suggested that open-weight models are catching up to closed SOTA in just ~3.5 months.
Others Topics
OpenAI reported that over 1 million users discuss suicide on ChatGPT weekly, sparking debate amid an ongoing lawsuit concerning the platform's safety protocols. On a lighter note, YouTuber PewDiePie showcased his custom 10x4090 local AI lab for running and fine-tuning large models.