Today’s lead · X
Nvidia CPO switches enter mass production for Vera Rubin-scale AI factories
Nvidia executive Gilad Shainer said co-packaged optics has moved into mass production, with Nvidia-and-partner switches already delivered to close customers and deployed inside Nvidia. For AI infrastructure builders, the signal is that scale-up networking for Vera Rubin-era systems is moving from roadmap to deployment.
Top signals
1 more
Nvidia Nemotron 3 Nano Omni brings video, audio, image, and text understanding to a 31B MoE
Nvidia’s Nemotron 3 Nano Omni is described as a 31B-parameter Mamba2-Transformer hybrid MoE with about 3B active parameters per token and support for video, audio, image, and text inputs. It targets enterprise multimodal workflows such as document intelligence, OCR, video and speech analysis, GUI automation, and agentic applications.
Tools & repos
4 selectedlyogavin/airllm
AirLLM is a GitHub repository for running 70B LLM inference on a single 4GB GPU, useful for builders testing large-model inference under tight hardware constraints.
Termexo
Termexo is a local Windows workspace for Claude Code and Codex that lets developers arrange PTY terminals, resume native sessions, receive approval notifications, and switch Claude-compatible model profiles.
Bolcho AI
Bolcho AI is a voice AI platform for building, deploying, and scaling multilingual phone and web agents for India, with native language support and bring-your-own LLM, STT, and TTS flexibility.
Lumichats
Lumichats is a desktop version of a browser-based AI coding/work assistant, positioned as a Claude Code alternative for people who avoid the terminal.
Blogs worth your time
2 reads
LangChain: How to rein in coding-agent spend
LangChain argues that coding-agent cost overruns are increasingly caused by fragmented visibility across Claude Code, Codex, Cursor, Copilot, Pi, and OpenCode. The post lays out a workflow for tracing, normalizing, optimizing, and governing spend with LangSmith, Engine, and LLM Gateway.
Two Minute Papers explains a parkour AI that combines imitation and goal-directed learning
The video walks through a research approach for virtual parkour characters that learns from a small amount of human motion data while also training to solve new obstacle courses. The key builder takeaway is the trade-off between human-like motion and task success in embodied AI systems.
Community discussions
3 threadsBittensor subnet controversy turns on model verification, not reputation
The thread argues that a fake 120B model on a Bittensor subnet is a stress test for decentralized model verification. The tension is clear: validators initially rewarded a plagiarized checkpoint, but open weights let independent analysis catch the issue quickly.
AI-personalized software sharpens the case for open-source devtools
The discussion argues that coding agents change the return on customizing software: they can make local modifications and help keep those changes rebased on upstream. The debate centers on whether closed-source devtools become less attractive when users increasingly expect agent-driven personalization.
Agent-readable pricing pages become a new distribution surface
The thread frames AI-agent readability as the next layer of distribution after SEO. The practical tension: product pages optimized for human UI can become invisible to agents if key facts are trapped in JavaScript widgets or images.
Funding & acquisitions
0 movesNothing material cleared the editorial bar today.
Bengaluru radar
3 events
Bits n Atoms: Build with OpenAI team
A Bengaluru in-person builder session with the OpenAI team, Together Fund, and The Product Folks, focused on the new 5.6 series, latest multimodal models, and practical production patterns for AI products.
Build AI Systems Enterprises Trust
A Bengaluru in-person session for senior engineers, technical leads, and solutions architects on taking agentic systems from sandbox demos to reliable enterprise production on Google Cloud.
Impact Lab Bengaluru: Superhuman Lab
A two-day Bengaluru hackathon bringing AI engineers, hardware tinkerers, artists, product designers, persons with disabilities, and caregivers together to prototype assistive and human-augmentation technology.