Today’s lead · x.com
OpenAI says it built its first AI chip with Broadcom
OpenAI announced Jalapeño, its first in-house AI chip, saying it was designed with Broadcom for the LLM workloads behind ChatGPT, Codex, the API, and future agentic products. The company did not disclose specs, deployment timing, or whether Jalapeño targets training, inference, or both, but the move pushes OpenAI further into full-stack infrastructure as model serving costs become a strategic constraint.
Top signals
11 more
Cordyceps CI/CD pattern puts 300+ major GitHub repos at risk
Novee Security described Cordyceps, a CI/CD workflow weakness where untrusted pull-request data crosses into privileged workflows, allowing unauthenticated users to run code, steal credentials, or gain repository control. Its scan of about 30,000 high-impact repositories found more than 300 fully exploitable cases, with examples involving Microsoft Azure Sentinel, Google’s adk-samples, Apache Doris, Cloudflare Workers SDK, and Python’s Black. The uncomfortable agent-era angle is that AI-generated workflow changes can reproduce these trust-boundary bugs at scale while each individual YAML step still appears to be “working as designed.”
Figma and Genspark push design tools closer to runnable code
Figma’s latest update adds code layers on the collaborative canvas, repo cloning, code-to-design flow extraction, animation and 3D transform support, AI shader fills, prompt-created plugins, and assistant skills that can use context from tools like Notion, Granola, Excel, and GitHub. Genspark separately launched Genspark Design, powered by Claude Opus 4.7, for generating UI prototypes, videos, HTML animations, and posters, with uploaded Figma files/design systems for brand reuse and one-click code generation through Genspark Code.
Notion lets teams assign tasks to Claude and Cursor
Notion introduced External Agents for Claude and Cursor, letting teams @-mention agents, assign work from shared boards, and watch task execution inside the same workspace humans already use. The product pattern is notable because agents become visible collaborators with shared task state rather than private side-channel chats.
Agent tools get real phones and lightweight browsers
MobAI launched tooling that lets Claude Code, Codex, Cursor, OpenCode, Aider, Cline, and other agents drive iOS or Android apps on real devices or simulators via MCP, Python harness, CLI, or HTTP API, with log reading, plain-English test generation, replayable .mob scripts, and demo video capture. Separately, wb is an open-source macOS 26+ browser CLI for agents: an 876KB-to-under-2MB native binary with persistent sessions, compact JSON, screenshots, clicks/forms, JavaScript eval, and live preview, but no Linux, Windows, or older macOS support.
Ray Serve LLM on GKE claims 5x throughput and 8x latency gains
Google Cloud and Anyscale say Ray Serve LLM on Google Kubernetes Engine now reaches up to 5x higher throughput and 8x lower latency through HAProxy-based internal routing, direct token streaming from model replicas, and a v2 Ray executor backend for vLLM. The benchmarks used Gemma 4 E2B on GKE clusters with Google Cloud A4 VMs powered by NVIDIA HGX B200 systems, aiming to close the gap between Ray’s developer ergonomics and native vLLM serving performance.
ParallelKernelBench tests where LLMs fail on multi-GPU kernels
Together AI researchers introduced ParallelKernelBench, a benchmark of 87 multi-GPU kernel problems drawn from real systems including Megatron-LM, DeepSpeed, DeepEP, TensorRT-LLM, and NeMo-RL. The point is narrower and more useful than generic coding scores: models that can write fast single-GPU kernels still struggle with distributed synchronization, communication, and performance constraints.
Vishal Sikka’s Hang Ten raises $32M to automate enterprise software work
Former Infosys CEO Vishal Sikka launched Hang Ten Systems with a $32 million seed round led by Mayfield, plus Aramco Ventures and angels, to use AI-driven development and automation for building, modifying, and operating enterprise software. The company says Siemens Gamesa Renewable Energy and Fresenius are early customers, positioning Hang Ten against traditional IT services economics that scale mostly with headcount.
OpenMontage open-sources an agentic video production system
Roundtable announced OpenMontage as an open-source agentic video production system that turns Claude or Cursor into a video studio workflow. The project claims videos can be produced for under $1 each, making it worth watching as video generation moves from single prompts toward orchestrated editing, scripting, and assembly pipelines.
OptimAI ships an MCP server for live web, social, and onchain context
OptimAI introduced OptimAI Search MCP, an integration layer for agents in Claude, Cursor, Copilot, and Codex to access live public web, social, and onchain signals. The company says the backend processes hundreds to thousands of real-time signals in parallel through a decentralized network, framing search as agent infrastructure rather than a standalone destination.
SemiAnalysis flags a GB300 NVL72 firmware reboot bug
SemiAnalysis says NVIDIA’s GB300 NVL72 racks have a firmware bug that requires a rack reboot every 66.5 days. The report is a reminder that AI infrastructure reliability depends not only on accelerator supply and networking, but also on firmware maturity and operational playbooks for very large GPU systems.
MoEngage buys Aampe to assign AI agents to individual customers
MoEngage acquired San Francisco-based Aampe in an all-cash deal reportedly worth tens of millions of dollars, adding technology that creates a dedicated AI agent for each customer instead of relying on broad marketing segments and rule-based campaigns. Aampe has more than 30 customers, grew ARR 150% over the past year, and will add about 20 employees to MoEngage, which says it serves over 1,350 brands across 75 countries.
Tools & repos
3 selectedcalesthio/OpenMontage
World's first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio.
ZhuLinsen/daily_stock_analysis
LLM 驱动的多市场股票智能分析系统：多源行情、实时新闻、决策看板与自动推送，支持零成本定时运行。 LLM-powered multi-market stock analysis system with multi-source market data, real-time news, decision dashboard, automated notifications, and cost-free scheduled runs.
apple/container
A tool for creating and running Linux containers using lightweight virtual machines on a Mac. It is written in Swift, and optimized for Apple silicon.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Funding & acquisitions
5 movesTencent EdgeOne Makers
Ship AI agents like web apps, in minutes.
Propane
Automatic customer context for product teams and agents
Crewdle AI
Use every business AI tool without every subscription
Stripe.Directory
New way for you & agents to search for businesses on Stripe
Customer Relationship Agents by Clarify
The M in CRM shouldn't be you
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.