## News / Update
A packed week of releases and industry moves: Google DeepMind launched Veo 3.1 with major upgrades for AI video and broad availability via the Gemini API, Video Arena, and Hugging Face. OpenAI expanded ChatGPT Go to 89 countries and reportedly hit 800 million users and $13 billion in annual revenue, underscoring the high costs of scaling despite rapid growth. Google’s site teased “Gemini 3.0 Pro” as its “smartest model yet.” NVIDIA began shipping DGX Sparks desktop AI systems, while Meta broke ground on a 1GW AI data center in Texas. Azure’s local MCP server is now globally available for offline DevOps, Salesforce rolled out Agentforce 360 to streamline agent deployment, and AMD, Cohere, and Oracle teamed up on enterprise AI. LangChain teased new product reveals at its Boston anniversary event, and Yupp added instant Solana-based cash outs. On decentralized AI, dphnAI and the Dolphin Network opened GPU contribution programs with token incentives for dataset generation and verified inference. Academic and community updates included QuackIR accepted to EMNLP 2025, calls for participation from KAUST, Mila, and Princeton, and OpenArt’s global AI music video competition.
## New Tools
Developers gained several new building blocks: retrieve-dspy introduced flexible, open-source pipelines for compound retrieval; LlamaAgents made it easy to spin up production-ready document extraction agents with schemas, validation, and confidence scoring; and a GEPA + DSPy tool provided verifiable PII removal from incident reports. Amp opened its agentic coding platform for free use, relying on ads and efficient open models to deliver professional-grade code generation at no cost.
## LLMs
Anthropic’s Claude Haiku 4.5 led the model news with double the speed and one-third the cost of its predecessor, matching or exceeding Sonnet 4 in coding and computer-use tasks and integrating across popular tools (including GitHub Copilot). With DSPy optimization, Haiku 4.5 posted strong results on NYT Connections at low cost and time, and user reports showed robust agentic coding performance. Agent systems also moved forward: Shanghai AI Lab’s MUSE paired with Gemini 2.5 set a new SOTA on real-world tasks using memory-based “learn on the job” strategies. Efficiency and context handling dominated research headlines: GPT-5-mini surpassed a larger GPT-5 on long-context problems via iterative code exploration; Recursive Language Models promise effectively unbounded context by recursive decomposition; and “thinking tokens” highlight how modern models allocate extra compute implicitly for harder queries. On-device gains surprised as quantized Qwen3-Next-80B outperformed bf16 on Apple’s M3 Ultra in early tests. GLM-4.6 posted explosive early adoption and topped open web dev benchmarks. Vision-language capabilities advanced with Sonnet 4.5’s strong OCR and immediate MLX-VLM support for Qwen3-VL. Training and evaluation methods evolved, with NVIDIA rewarding informative reasoning chains, Meta’s ETD (Encode-Think-Decode) improving reasoning via recursive training, and new work stressing the difficulty of fair agent benchmarks and proposing information-theoretic fixes for process reward modeling. Safety research progressed with the MALT dataset for studying reward hacking.
## Features
AI assistants are becoming more agentic and integrated. GitHub Copilot now tackles end-to-end tasks, fixes vulnerabilities, connects tools, and submits PRs. ClickUp added a Codegen agent that chats through bug fixes and turns them into one-click pull requests, while also generating code directly from notes, tasks, and whiteboards. Azure’s local MCP server enables fully offline DevOps workflows for compliance-sensitive teams. Veo 3.1 introduced finer creative controls, audio support, scene extensions, multi-reference guidance, and improved editing across Gemini, Video Arena, and Hugging Face Apps, with demonstrated fluency even in Japanese prompts. NotebookLM now turns dense arXiv papers into conversational explanations, and LangChain shipped built-in guardrails like PII redaction and human-in-the-loop options. NVIDIA contributed a patch that boosts llama.cpp generation speeds by up to 40% on DGX Sparks. Salesforce’s Agentforce 360 refined the instruction-to-deployment path for AI agents. Community work also broadened access to vision-language models with day-one MLX-VLM support for Qwen3-VL.
## Tutorials & Guides
Builders received strong hands-on resources: a complete walkthrough for creating an AI voice transcription app using Next.js, AI SDKs, and Together AI; Stanford’s CS336 practical video deep dive into Karpathy’s nanochat (tokenization, architecture, GPU efficiency, scaling); and comprehensive end-to-end robotics tutorials with ready-to-run code via LeRobotHF and Hugging Face. A workshop in Madrid highlighted DSPy for automatic prompt optimization, and new Nanochat demos let practitioners jump into pretraining, fine-tuning, or RL without starting from scratch.
## Showcases & Demos
Interactive and low-cost demos stole the show: ChatGPT Apps ran classic Doom in-browser, signaling the platform’s growing support for rich, interactive experiences. Veo 3.1 was stress-tested in public arenas and HF Apps, where creators showcased smoother roleplay and multilingual prompts with improved video fidelity. The nanochat stack expanded into multimodality for under $10 using a SigLIP ViT projection, with stepwise checkpoints available across training stages. Developers demonstrated Claude’s code subagents delivering high-quality, parallelized code and rapidly producing interactive web apps from VS Code. On the research front, HivergeAI’s algorithm set a new CIFAR-10 training speed record (1.99s on a single A100), underscoring how aggressive optimization can still unlock major performance gains.
## Discussions & Ideas
The community debated the pace and shape of progress: forecasts for AGI by 2027 were deemed premature, even as benchmarks continue to improve steadily. Multiple perspectives framed OpenAI’s Sora 2 as a human-in-the-loop “social” system that could make real-world data collection a participatory learning engine. Developers flagged how GPU export restrictions may discourage advanced kernel innovation, while intense competition between OpenAI and Anthropic is accelerating improvements in coding agents. Fresh research insights suggested a single sentence can substantially boost ChatGPT’s creativity, and “verbalized sampling” clarified how diversity persists beyond sampling tweaks. Methodological advances spanned domains: a simple ColBERT architectural tweak yielded broad retrieval gains; representation autoencoders showed high-dimensional diffusion is practical; Spatial Forcing improved robots’ 3D understanding and training efficiency; Open-YOLO 3D set a new standard for open-vocabulary 3D instance segmentation; and Apple’s DeepMMSearch-R1 aims to strengthen multimodal web retrieval.
