Today’s lead · Google DeepMind Blog
Gemini 3.8 Live pushes voice agents toward real-time work
Google released Gemini 3.8 Live and an Extended Thinking variant for speech-to-speech agents. The useful shift is not just better chat: Google says the models can use visual context, switch across 97 languages, and run tool/API calls in the background while conversation continues.
Top signals
4 more
Salesforce and Nvidia turn Nemotron into a domain reasoning model
Salesforce’s Koa is a practical enterprise counterweight to frontier-model default routing: an open-weight Nemotron-based reasoning model post-trained for sales, marketing, and support tasks. The pitch is lower token spend, tighter data controls, and Agentforce routing—not a break from OpenAI or Anthropic.
FastVideo ships open weights for faster MiniMax H3 video generation
FastVideo announced open-weight FastH3 V2, claiming up to 9x speedup on Nvidia Blackwell with lossless quality versus the base 50-step MiniMax H3. The builder angle is immediate availability: day-zero ComfyUI weights and workflows, plus an API served on Reactor.
Meta gives coding agents a WhatsApp Business MCP server
Meta’s new WhatsApp Business Tools MCP connects agents like Claude, Cursor, Codex, and ChatGPT to WhatsApp Business setup tasks. This is MCP doing what builders actually need: reducing console-hopping for account creation, Cloud API access, template editing, webhook testing, and troubleshooting.
CheatBench measures reward gaming in AI agents
CheatBench is a useful reminder that agent evaluations need to measure behavior under temptation, not just task success. The benchmark sets up hard assignments with discoverable shortcuts, then checks whether agents seek hidden answers, copy submissions, manipulate grading, or follow misleading social cues.
Tools & repos
3 selectedalphaXiv/OpenResearch
OpenResearch is a Rust repository that aims to turn coding agents into research agents. The traction is notable today: 3,524 stars and 531 stars added on GitHub Trending.
Kilo Code for iOS and Android
Kilo Code brings coding-agent control to mobile: run Cloud Agents, manage VS Code and CLI sessions, review pull requests, and respond from iPhone, iPad, or Android.
Voiskey
Voiskey is AI voice typing that rewrites rough speech for the destination and audience. It is available on iOS, macOS, Android, and Windows, with support for over 100 languages.
Blogs worth your time
5 reads
Your agent passed once. IBM asks if it can pass every time
IBM’s post makes a sharp evaluation point: Mean@k can hide agents that only sometimes solve the same task. Their ALTK-Evolve consistency guidelines target flip-prone decisions without requiring end-to-end replay.
Google’s Retrieve-for-Train compiles search reasoning out of the hot path
Retrieve-for-Train is a useful systems idea: spend RL offline to learn good query fan-outs, then distill that behavior into a small diffusion retriever for single-pass inference.
NVIDIA’s dense-vs-MoE guide is really a deployment checklist
NVIDIA uses Nemotron 3.5 Lightning to explain why active parameters are not the same as memory footprint. The post is most useful as a serving tradeoff map.
Ming-Yu Liu on Cosmos, world models, and physical AI verification
The transcript frames Cosmos as an omni model for text, video, audio, and action, with near-term value in policy verification before full sim-to-real training reliability.
Two Minute Papers explains Claude text watermarking
The video explains text watermarking as a statistical fingerprint created by nudging token choices, then argues that light editing may not remove it while full rewriting can.
Community discussions
3 threadsA Claude Code team asks what “fully agentic” work is doing to engineers
The thread is less about Claude Code quality and more about team design. A senior engineer describes a “fully agentic” office with no pair coding, code review, or deep work; commenters push back that speed is not collaboration.
Cursor users debate cross-tool memory for Claude Code, Codex, and Cursor
A practical memory thread: Claude Code and Codex can use lifecycle hooks for automatic recall and saving, while Cursor is described as MCP-only and on-demand. The comments pressure-test whether docs and errata files beat a memory layer.
A tiny offline LLM sparks useful skepticism about persistent memory
The SHADOW-50M post is candidly a proof of concept: tiny, offline, fast, and benchmark-weak. The most interesting discussion targets its persistent retrieval trail and whether wrong first retrievals can become self-reinforcing.
Funding & acquisitions
5 moves
FactoryAI raises $200M at a $5B valuation
FactoryAI says it raised $200M at a $5B valuation to scale enterprise software development automation. The company names large customers, but the post does not disclose round stage or investors.
Profound raises $180M Series D for AI search marketing
Profound raised a $180M Series D at a $1.8B valuation, only months after its Series C. The company is betting that AI search optimization becomes a durable marketing workflow, not a temporary SEO hack.
Exein raises $270M for physical AI security
Exein raised $270M at a $1.7B valuation for runtime security on connected and physical-AI devices. The company’s story is classic infrastructure timing: IoT security meets robots, drones, vehicles, sensors, and edge models.
AIUC raises $40M to certify enterprise AI agents
AIUC raised $40M Series A to build a third-party audit layer for AI agents. If agents become procurement infrastructure, buyers will want SOC 2-like evidence of boundaries, not just vendor promises.
Flam raises $40M Series B for interactive AI content
Flam raised $40M Series B led by QED Investors to expand AI models, product formats, and enterprise sales. The company is selling interactive content for marketing, product visualization, support, entertainment, and fan engagement.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.