2026-09-16 · WednesdayPublished daily at 08:00
AI Daily
- 1Google DeepMind releases Gemini 3.8 Live and 3.8 Live Extended ThinkingGoogle DeepMind:Blog
- 2ShengShu Technology releases Vidu S2: featuring Avatar and Editing models, exploring spatial videoWeChat official account: Shengshu Technology
- 3
01Model Releases/Updates
Google DeepMind releases Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind released two near-real-time voice conversation models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, focused on voice agents and complex task execution.
Original textShengShu Technology releases Vidu S2: featuring Avatar and Editing models, exploring spatial video
ShengShu Technology officially releases Vidu S2, including Vidu S2-Avatar for real-time interaction with digital characters and Vidu S2-Editing for real-time editing of video streams, while exploring real-time spatial video generation and editing for VR headsets.
Original textStepFun released the StepAudio 3 series of large speech models, with several ranking first globally on the Artificial Analysis leaderboard.
StepFun has released the StepAudio 3 series, comprising five models: Realtime, ASR, TTS, Gen, and Music. They are now available on the StepFun Open Platform.
Original text02Product Releases/Updates
Google released multilingual AI achievements including TranslateGemma, with language technology covering more than 300 languages.
Google announced that its language technology now supports more than 300 languages and covers 86% of the global population. It also released TranslateGemma, a lightweight open-source translation model trained on Gemini that supports 55 languages and can run offline.
Original textClaude for Small Business adds 43 workflows and 27 integrations, and launches a free training program
Anthropic has added 43 workflows and 27 integrations to Claude for Small Business, covering Shopify, Salesforce, Stripe, Gusto, and more. The product has been installed more than 900,000 times since launching in May. Workflows default to approval mode, so users must confirm every message sent, post published, or payment made. This fall, the company will hold free workshops in 10 U.S. cities, and 14 integration partners will each host a free webinar between late September and November.
Original text03Industry News
Perplexity developed CobbleDB in-house to replace AWS DynamoDB, potentially saving up to $100 million per year.
Perplexity CEO Aravind Srinivas announced CobbleDB, an in-house key-value database that replaces AWS DynamoDB for fast web-content scraping and could save up to $100 million a year after migration. Two engineers and hundreds of continuously running Computer agents built the core infrastructure in two months. According to the company’s research, batch-read latency for hot storage is about 5x lower than with DynamoDB across all distributions (P50 fell from 31.4ms to 5.60ms).
Original text04Tips and perspectives
Gergely Orosz visits OpenAI: a Codex-powered agentic software factory
Gergely Orosz visited OpenAI’s headquarters and interviewed seven engineers and engineering leaders. He found that since around January, Codex and ChatGPT Work have become the foundation for nearly all work at the company.
Original textAnthropic and OpenAI proposed coordinating a slowdown in frontier AI development, while critics including the CEO of Cohere questioned their true motives.
Anthropic CEO Dario Amodei called on the industry and governments to coordinate a slowdown in frontier AI development and seek antitrust exemptions. Sam Altman and Elon Musk agreed.
Original text404 Media exposes OpenAI’s Lily program: manually reviewing ChatGPT chat records to improve the model
404 Media revealed an OpenAI project internally codenamed Project Lily, in which prompt reviewers earning more than $50 per hour examine anonymized records of real user chats and judge whether the replies are on-topic and whether they use AI-like phrasing or a sycophantic tone.
Original textArena updated its Image-to-WebDev leaderboard: GPT-6 Astra took first place with a score of 1733.
Arena updated its Image-to-WebDev leaderboard, adding four newly evaluated models. GPT-6 Astra (Max) ranked first with 1733 points, 129 ahead of GPT-5.6 Sol (xHigh). Claude Fable 5.1 (Max) came second with 1710 points, Muse Spark 1.3 (Max) ranked fourth with 1645, and GLM-5.3-Flash ranked tenth with 1588.
Original textArtificial Analysis evaluation: GPT-Live-1 tops the Speech to Speech Index with a score of 81.5
Artificial Analysis has released the Speech to Speech Index. OpenAI’s GPT-Live-1 ranked first with a score of 81.5 (Astra backend, medium reasoning effort), ahead of Grok Voice Think Fast 2.0 High at 81.3. The Sol backend configuration came third with a score of 80.1.
Original textTrail of Bits criticized 1Password's AI patch benchmark as misleading and released two patch-verification Agent skills.
Trail of Bits criticized 1Password’s FLAWED report, published on August 6, saying its 26% AI clean-fix rate was distorted by four experimental design choices. These included prompts deliberately instructing agents to apply incorrect fixes, which made up 22% of the data; 36% of trials that prohibited compilation tests; and differences in reasoning-effort settings.
Original textVercel cuts its inbound sales team from 10 people to 1.25, while AI sales development agents cost only a few thousand dollars per year
Vercel COO Jeanne DeWitt Grosser said in an interview with The Information that the company has automated 90% of its inbound sales development, reducing the team from 10 people to 1.25.
Original text