Google DeepMind Releases Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind releases Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two near-realtime voice conversation models focused on voice agents and complex task execution.
Model Releases & Updates
Google DeepMind Releases Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind releases Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two near-realtime voice conversation models focused on voice agents and complex task execution.
Source: Google DeepMind：Blog（RSS）
Shengshu Technology Releases Vidu S2: Avatar and Editing Dual Models, Exploring Spatial Video
Shengshu Technology officially releases Vidu S2, including Vidu S2-Avatar for real-time digital character interaction and Vidu S2-Editing for real-time video stream editing, exploring real-time spatial video generation and editing for VR headsets.
Source: WeChat：生数科技（Vidu·Video）
StepFun Releases StepAudio 3 Series Voice Models, Multiple Ranking #1 on Artificial Analysis
StepFun releases StepAudio 3 series, including five models: Realtime, ASR, TTS, Gen, and Music, now available on the StepFun open platform.
Source: WeChat：阶跃星辰（Step）
Product Releases & Updates
Google Releases TranslateGemma and More Multilingual AI成果, Language Tech Covers 300+ Languages
Google announces its language technology now supports over 300 languages, covering 86% of the global population, and releases TranslateGemma, a lightweight open-source translation model (trained on Gemini, supports 55 languages, runs offline).
Source: Google Blog：AI（RSS）
Claude for Small Business Adds 43 Workflows and 27 Integrations, Launches Free Training Program
Anthropic adds 43 workflows and 27 integrations to Claude for Small Business, covering Shopify, Salesforce, Stripe, Gusto, and more. The product has over 900,000 installations since its May launch. Workflows default to approval mode, requiring user confirmation for all sends, posts, or payments. Free workshops will be held in 10 US cities this fall, with 14 integration partners each hosting a free webinar from late September through November.
Source: Claude：Blog
Industry News
Perplexity Builds CobbleDB to Replace AWS DynamoDB, Could Save Up to $100M Per Year
Perplexity CEO Aravind Srinivas announces the self-developed key-value database CobbleDB to replace AWS DynamoDB for fast web content scraping. After migration, it could save up to $100 million per year. The project was completed by two engineers and hundreds of continuously running Computer agents in two months. Official research shows hot storage batch read latency dropped approximately 5x compared to DynamoDB's full distribution (P50 from 31.4ms to 5.60ms).
Source: X：Aravind Srinivas（Perplexity CEO）
Insights & Perspectives
Gergely Orosz Visits OpenAI: Codex-Driven Agent Software Factory
Gergely Orosz visits OpenAI headquarters and interviews seven engineers and engineering leaders, finding that since approximately one month ago, Codex and ChatGPT Work have become the foundation for nearly all of the company's work.
Source: Pragmatic Engineer（RSS）
Anthropic and OpenAI Propose Coordinating Slowdown in Frontier AI Development, Cohere CEO and Critics Question True Motives
Anthropic CEO Dario Amodei calls for industry and government coordination to slow frontier AI development, seeking antitrust exemptions. Sam Altman and Elon Musk agree.
Source: The Decoder：AI News（RSS）
404 Media Exposes OpenAI Project Lily: Human Review of ChatGPT Chat Logs to Optimize Models
404 Media reveals OpenAI's internal Project Lily, where hourly contractors over $50 review anonymized real user chat logs, evaluating whether responses are relevant and free of AI-speak and sycophancy.
Source: IT之家（RSS）
Arena Updates Image-to-WebDev Rankings: GPT-6 Astra Tops with 1733 Points
Arena updates the Image-to-WebDev leaderboard with four new evaluation models. GPT-6 Astra (Max) ranks first with 1733 points, 129 points ahead of GPT-5.6 Sol (xHigh). Claude Fable 5.1 (Max) ranks second with 1710 points, Muse Spark 1.3 (Max) ranks fourth with 1645 points, and GLM-5.3-Flash ranks tenth with 1588 points.
Source: X：Arena
Artificial Analysis: GPT-Live-1 Tops Speech to Speech Index with 81.5 Points
Artificial Analysis releases the Speech to Speech Index. OpenAI's GPT-Live-1 ranks #1 with 81.5 points (Astra backend, medium reasoning intensity), surpassing Grok Voice Think Fast 2.0 High's 81.3. Sol backend configuration scores 80.1, ranking third.
Source: X：Artificial Analysis
Trail of Bits Criticizes 1Password's AI Patch Benchmark as Misleading, Releases Two Patches to Verify Agent Skills
Trail of Bits publishes an article criticizing 1Password's FLAWED report from August 6, stating its 26% AI clean fix rate is distorted by four experimental design choices, including prompts that deliberately instructed agents to apply wrong fixes accounting for 22% of data, 36% of trials prohibiting compilation tests, and varying reasoning effort settings.
Source: Trail of Bits：AI Security Research
Vercel Reduces Inbound Sales Team from 10 to 1.25 People, AI SDR Agents Cost Only Thousands Per Year
Vercel COO Jeanne DeWitt Grosser says in a The Information interview that the company's inbound sales development has achieved 90% automation, with the team shrinking from 10 people to 1.25 people.
Source: Tomer Tunguz Blog（VC Analysis）
【Hacker News Top Posts】
Keywords：AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning" OR "large language model" OR AGI OR "generative AI" OR AI agent OR RAG OR diffusion OR transformer
Data Source：hnrss.org | Filtered from Hacker News
1. Show HN: Plurnk (Yet Another AI Harness)
A self-described "AI harness" with ANTLR grammar, omnipatterns via treesitters, MCP2 tools, AG-UI client, A2A agent-to-agent, and a lean 200MB footprint that runs on weak hardware with multiple agents.
2. Show HN: Agenttik – work on multiple projects in parallel with AI agents
Tool for managing multiple AI agent projects in parallel, distributing work across different agents and branches, with subscription support for Claude Code, OpenAI, and other providers.
3. Has anyone made any money/successes making SaaS's with AI?
Discussion thread about struggling to monetize AI-powered side projects, with the author sharing several attempts including homework help, drawing tools, and pet adoption games.
4. Gemini 3.8 Live and 3.8 Live Extended Thinking
Google DeepMind's new near-realtime voice conversation models for voice agents and complex task execution. (261 points, 177 comments)
Originally published on WeChat Official Account 「比特财商」.
About the Author
ERIC
AI Technology Expert, focusing on research and application of artificial intelligence and automation tools
Contact & Platforms