🤖 AI HOT Daily · Sep 16, 2026
Google DeepMind ships Gemini 3.8 Live and 3.8 Live Extended Thinking — near-real-time voice models for voice agents and complex work; Shengshu unveils Vidu S2 with dual Avatar (real-time digital-character interaction) and Editing (real-time video-stream editing) models, plus VR-headset spatial video explorations; StepFun's StepAudio 3 family — Realtime, ASR, TTS, Gen, and Music — seizes multiple No.1 slots on Artificial Analysis; Google's language tech now spans 300+ languages reaching 86% of humanity, alongside TranslateGemma, a Gemini-trained lightweight open translation model (55 languages, offline-capable); Claude for Small Business adds 43 workflows and 27 integrations (Shopify, Salesforce, Stripe, Gusto…), passing 900K installs since May — approval-mode by default and free training included; Perplexity builds CobbleDB to replace AWS DynamoDB for rapid web fetching — two engineers plus hundreds of always-on Computer agents in two months, saving up to $100M a year; Pragmatic Engineer visits OpenAI and finds Codex + ChatGPT Work underpinning nearly all work since ~a month ago — a true agentic software factory; Anthropic and OpenAI pitch coordinated frontier slowdowns with an antitrust exemption (Altman and Musk agree) as Cohere's CEO and others question the real motive; 404 Media exposes 'Project Lily' — reviewers earning $50+/hr scrutinize anonymized real chats for relevance, AI-speak, and sycophancy; Arena's Image-to-WebDev leaderboard puts GPT-6 Astra first at 1733 (129 clear of GPT-5.6 Sol), Claude Fable 5.1 second at 1710; Artificial Analysis's Speech-to-Speech Index crowns GPT-Live-1 (Astra backend, medium reasoning) at 81.5, a hair over Grok Voice Think Fast 2.0 High's 81.3; Trail of Bits calls 1Password's AI-patching benchmark misleading, blaming four experimental choices for the inflated 26% clean-fix number; and Vercel shrinks inbound sales from 10 people to 1.25 FTE — 90% automated, with AI SDR agents costing a mere few thousand dollars a year.
1. Gemini 3.8 Live and 3.8 Live Extended Thinking
Two near-real-time voice models from DeepMind aimed at voice agents and harder tasks — conversation, at thinking speed.
2. Vidu S2: Avatar + Editing, Space Videos in View
Avatar powers real-time digital-character interaction, Editing wields live video-stream edits — with spatial video gen/editing for VR headsets in exploration.
3. StepFun's StepAudio 3 Tops Voice Charts Worldwide
Five speech models — Realtime, ASR, TTS, Gen, Music — dock on the open platform, several ranking No.1 on Artificial Analysis.
4. Google: 300+ Languages and an Open TranslateGemma
Language tech now reaches 86% of humanity; TranslateGemma — a Gemini-trained, 55-language, offline-capable open lightweight — joins it.
5. Claude for Small Business: 43 Workflows, 27 Integrations
Now reaching Shopify, Salesforce, Stripe, Gusto and more; 900K+ installs since May with approval-mode by default and free training — safety first, speed second.
6. Perplexity Swaps DynamoDB for CobbleDB, Saves Up to $100M/yr
Two engineers plus hundreds of always-running Computer agents built the key-value store for fast web fetching in just two months.
7. Inside OpenAI: A Codex-Powered Software Factory
Pragmatic Engineer talks to seven engineers and leads: since about a month ago, Codex and ChatGPT Work underpin nearly all of the company.
8. Anthropic & OpenAI Push a Coordinated Slowdown — Motives Questioned
Amodei wants a coordinated, government-blessed pace with an antitrust exemption — Altman and Musk nod along, skeptics smell something else.
9. 'Project Lily': OpenAI Staff Read Your Chats to Improve the Model
Reviewers paid $50+/hr examine anonymized real conversations, scoring relevance, AI-speak, and sycophancy.
10. Astra Tops Image-to-WebDev at 1733 Points
Astra leads GPT-5.6 Sol by 129 points, with Claude Fable 5.1 second at 1710.
11. GPT-Live-1 Leads the Speech-to-Speech Index at 81.5
Astra backend at medium reasoning edges past Grok Voice Think Fast 2.0 High's 81.3 by a hair.
12. Trail of Bits: 1Password's AI-Patch Benchmark Misleads
The 26% clean-fix claim is distorted by four design choices — prompted bad fixes, 36% of trials blocking compiles, uneven reasoning settings, and more.
13. Vercel Cuts Inbound Sales 10x: AI SDRs Cost Thousands a Year
With 90% of inbound SDR work automated, staffing drops from 10 to 1.25; the COO details it in The Information.
Sources and verification
This legacy briefing did not retain its original source URLs. Verify safety, financial, policy, and breaking-news claims with authoritative primary sources before acting on them.