Sep 16 · AI Daily

Plasma AI has launched Radio, offering a shared chat room for agents from different makers like Claude, Codex, and Grok, saving people from manually copying outputs between tools. Users create a channel and share a link, and any agent that can fetch the URL can join, read the same conversation, and reply to humans and other agents. Channel logs are fully auditable, and creation requires no signup, directly tackling the pain point of manual routing in multi-agent workflows.

Meta plans to launch smart glasses without cameras this autumn, code-named Luna, seen as a response to privacy and safety scrutiny over recording features on smart glasses. The glasses have six built-in microphones for voice access to the Meta AI chatbot and the consumer AI agent Muse, with a one-touch Meta AI wake button on the frame side. They may launch at next week's Connect conference and begin shipping in October.

Google is bringing several new features to NotebookLM: real-time voice conversation notes supporting nearly 100 languages, mobile lecture recording, interactive study guides with custom quizzes, and shareable 60-second video overviews. Voice recording rolls out to all users starting next week, supporting English and select other languages. Eligible college students can also claim one free year of Google AI, highlighting Google's push into learning scenarios.

Luma AI has released Camera Angles, turning a single photo into an entire set of shots: after picking an angle, you can retrieve the same subject with intact details from every direction. It solves the common request of needing a back view after a shoot, marking another step for image generation from an output tool to a controllable shooting tool.

Insta360 has released the Mic Pro Tencent Meeting Edition AI recording lapel mic for 698 yuan, integrating Tencent Meeting's recording transcription and minutes capabilities. Recordings automatically sync to the app to generate professional minutes. Weighing 19g, it features a three-mic array for 5-meter pickup, up to 9.5 hours of single-session recording, and 32GB of storage for over 1,000 hours of audio, serving as a localized example of AI meeting hardware tied to office software.

A New York Times and Siena College poll of 1,503 likely voters in early September showed that 61% oppose building data centers for AI technology, with only 14% in strong support. Among opponents, 56% favor placing limits on them and 38% lean toward an outright ban, with environmental and water use concerns being the most common reason at 32%. However, the issue ranks below 1% in importance for midterm voting topics, showing a clear gap between public opinion and voting weight.

Mark Zuckerberg posted about the debate over slowing down capability progress to let alignment catch up, arguing that trust and alignment are becoming the most important abilities distinguishing agents and models, and labs that ignore alignment will fall behind. He listed Meta's steps: delaying Muse's release for months for safety, bringing in independent evaluators for MSL across multiple domains, and committing most compute to serving users rather than recursive self-improvement, while noting he does not support forcing everyone to slow down together.

Rohan Paul, citing a December 2025 interview with Dario Amodei, stated that Anthropic's annualized revenue grew from about $9 billion at the end of 2025 to over $65 billion by July 2026, a growth rate of over 7x. Amodei noted that revenue grew 10x each year over the past three years, though even so he does not believe next year's figures can be extrapolated to $100 billion. This figure serves as a key benchmark for the commercialization speed of leading labs.

Arena has updated the Image-to-WebDev leaderboard and added four new evaluation models: GPT-6 Astra (Max) took first place with 1,733 points, leading GPT-5.6 Sol (xHigh) by 129 points. Claude Fable 5.1 (Max) ranked second with 1,710 points, Muse Spark 1.3 (Max) fourth at 1,645 points, and GLM-5.3-Flash tenth at 1,588 points, offering a fresh reference for model selection on image-to-website tasks.

Communities affected by heavy industry, such as Grays Ferry in Philadelphia, have launched the No Data Centers in Philly campaign, with over a hundred people rallying at City Hall to demand a halt to data center construction. Although the city has no official plan yet, officials have narrowed down two candidate sites. The AI data center boom is colliding directly with urban communities bearing scars from heavy industry.

Kevin Bai from the Anthropic Applied AI team gave an FDE 101 talk, explaining the frontline deployment engineer role: Palantir used the FDE model to reach the highest average contract value for a public SaaS company at $4 million. The prerequisite for establishing an FDE team is having a reusable platform, and it only fits scenarios selling complex products to non-technical buyers, offering a reference for delivery organization design at AI companies.

The author and Niklas used Codex and Claude over about a month to reverse engineer the AGX firmware ABI for the M4 Mac Mini and MacBook Neo, writing a Linux GPU driver that achieves full OpenGL ES 3.0 compatibility. Minecraft runs at 200fps, and WebGL in Chrome and Firefox composites properly, demonstrating the upper efficiency limit of LLM-assisted system-level reverse engineering.

LlamaIndex published an analysis on the relationship between accuracy and automation rates in OCR document processing, pointing out that routing gating rather than reading capability determines straight-through processing rates. Example calculations show that a 20-field invoice at 97% accuracy per field has only about 54% fully correct documents, while a pipeline with 94% accuracy but per-field confidence scores can automatically pass 92% of documents, whereas a pipeline with 97% accuracy but no signals has a 0% straight-through rate, offering useful reference for production document automation design.

SemiAnalysis published an article pushing back against the popular narrative that moratoria are killing US data center construction. Based on a project-level model covering over 6,000 facilities, they estimate that only about 2.3 GW of planned capacity has actually been delayed by local moratoria and executive orders in New York. Compared against opposition in public polls, this shows that policy friction has limited impact on actual construction.

Vercel COO Jeanne DeWitt Grosser said in an interview with The Information that the company's inbound sales development has achieved 90% automation, shrinking the team from 10 people to 1.25, with AI sales development agents costing only a few thousand dollars a year. This is a first-hand practical case quantifying the staff replacement effect and return on investment of AI agents in sales pipelines.

NVIDIA CEO Jensen Huang said at a Salesforce event that the AI industry does not need new laws or regulations for safety, and the conflict between safety and speed is a false dichotomy. He believes companies can rely on market mechanisms, controlling release cadence based on product maturity. Market competition, user choice, and business reputation will push companies to improve product safety, echoing remarks made by Altman on the same day and outlining a unified stance among leading makers on regulation.

Google announced at SEMICON Taiwan 2026 that its TPU interconnect architecture now supports up to 1 million TPU chips in a unified computing cluster. CTO Amin Vahdat stated that AI computing bottlenecks are shifting from a shortage of chips to a shortage of power, and argued that agent interactions will bring nearly infinite demand for computing power, setting the coordinates for the next stage of competition in AI infrastructure scale.

OpenAI CEO Sam Altman said at a Salesforce event in San Francisco that AI companies can safely research and develop AI through industry self-regulation without causing major harm to the public. He argued for establishing safeguards first through industry collaboration, independent evaluation, and stricter safety processes alongside government oversight. OpenAI previously delayed frontier reinforcement learning training, shifting some compute toward monitoring, safety guardrails, and risk assessment.

On September 15, Google introduced two native speech-to-speech dialogue models: 3.8 Live aimed at scaled deployment and cost optimization, and 3.8 Live Extended Thinking aimed at high-complexity tasks with enhanced multi-step reasoning. Both are now available on the Gemini Live API and Google AI Studio, supporting 97 languages. Artificial Analysis evaluations show the Extended Thinking version topped the Speech to Speech Index with 82.6 points, and ranked first on Tau Voice with 68.6%, becoming the new benchmark for production-grade voice agents.

IFM released the open-source 7B dense small model K2-Horizon-7B, scoring 21 on AA Bench, close to the 22 scored by Qwen3.6-27B. The authors stated that it outperformed GPT-5 and DeepSeek-V4 in BrowseComp tests, with similarly impressive scores on SWE-bench Verified and Terminal Bench. A parameter count of 7B approaching 27B levels has direct implications for local deployment and inference costs.

Periodic Labs built a high-throughput materials lab in Menlo Park, creating a closed loop between experiments and models where the lab generates new data and models learn from it to guide the next experiment. The team used only 1,300 H200 chips plus several months of experimental data for mid-training and RL on an open-source model to produce Neon, which outperformed GPT-5.6 Sol and GPT-6 Astra on their analysis benchmarks. Their initial research focuses on superconductors, magnets, and semiconductor materials, demonstrating the value of the low-compute plus real-world experimental data route.

During Jensen Huang's appearance at Dreamforce, Salesforce released its first CRM reasoning model, Koa, post-trained on NVIDIA Nemotron 3 Super using training data derived from proprietary synthetic datasets of enterprise CRM deployments spanning nearly three decades. Combining industry-proprietary data with open-source base models marks a signature move for CRM vendors building their own vertical reasoning models.

The FastVideo team, alongside nuvalab and NVIDIA FastGen, released FastH3 V2 open weights, achieving lossless quality compared to the base 50-step MiniMax H3, with up to 9x acceleration on NVIDIA Blackwell. It provides Day 0 ComfyUI weights and workflows alongside the reactorworld API, with Omni-ref currently in training, marking an important step forward in cutting video generation inference costs.