# 📆 ThursdAI - highest signal weekly AI news show - recorded LIVE Source: https://www.youtube.com/watch?v=lkeMsTYnJ4k Recap page: https://rapidrecap.app/video/lkeMsTYnJ4k Generated: 2025-09-04T18:01:40.031+00:00 --- ## Quick Overview The "ThursdAI" episode for September 4th highlighted significant advancements in AI, including the release of News Research's Hermes 4 models and Husky Holdem benchmark, Tencent's Hunan MT translation model, and the Upperus multilingual LLM from Switzerland, alongside major corporate news like Anthropic's $13 billion funding and OpenAI's reported $10 billion share sale. **Key Points:** - News Research released Hermes 4, featuring 4.5B and 7B models fine-tuned on Llama 3.1 and a 14B model fine-tuned on Qwen 3, aiming for neutral alignment and showing strong performance on EQBench and their new Husky Holdem poker bot evaluation benchmark. - The Upperus 8B and 70B models from Switzerland were open-sourced, trained on 15 trillion tokens across 1,800 languages with an Apache 2.0 license, representing a significant European contribution to open-source multilingual AI. - Tencent open-sourced Hunan MT, a 7B machine translation model that won WMT 2025 awards, showcasing competitive performance against larger models and highlighting its utility for data work and specific translation tasks. - Major AI companies saw significant financial activity: Anthropic raised $13 billion in Series F funding at a $183 billion post-money valuation, while OpenAI reportedly secured a $10 billion funding round through a sale of employee shares, valuing the company at $500 billion. - OpenAI also released GPT real-time with new API features including remote MCP tool support and image input, and acquired the companies Statsig and Alex (an AI coding agent for Xcode), bolstering its developer tools and AI capabilities. - Apple released Fast VLM 7B, a vision-language model emphasizing speed with 85x faster time to first token compared to peers, generating significant buzz on Hugging Face. - The episode featured discussions on the rapid development of coding agents, with Grok Fast Code (formerly Sonic) becoming a top performer on OpenRouter, and Codex seeing a 10x increase in usage following recent improvements. **Context:** This episode of "ThursdAI," a weekly AI news show, features hosts Alex Fov, Yan Pelleg, and LDJ, with guests Roger Jin and Besh Kumar from News Research discussing their latest model releases and benchmarks, and Quindla Kramer joining later to talk about GPT real-time. The show covers a wide range of AI news, from open-source model releases and company funding rounds to new benchmarks and API updates, reflecting the fast-paced nature of the AI industry. ## Detailed Analysis The "ThursdAI" September 4th episode provided a comprehensive overview of the week's AI news. News Research launched Hermes 4, a series of fine-tuned models (4.5B/7B on Llama 3.1, 14B on Qwen 3) focused on neutral alignment, demonstrating strength in EQBench and their novel Husky Holdem poker bot evaluation. Switzerland contributed the Upperus 8B and 70B multilingual LLMs, trained on 15 trillion tokens across 1,800 languages under an Apache 2.0 license. Tencent's Hunan MT, a 7B translation model, won WMT 2025 awards and showed competitive performance. In corporate news, Anthropic secured $13 billion in Series F funding, and OpenAI reportedly raised $10 billion via employee share sales, with both companies achieving massive valuations. OpenAI also enhanced its GPT real-time API and acquired Statsig and Alex. Apple released the speedy Fast VLM 7B. Coding AI saw advancements with Grok Fast Code topping OpenRouter charts and Codex usage surging. Discussions also touched on the competitive landscape of coding agents and the potential of companion models. ### Open Source AI News - News Research released Hermes 4 (4.5B/7B Llama 3.1, 14B Qwen 3) focused on neutral alignment, strong EQBench performance, and the Husky Holdem poker bot benchmark - Upperus 8B/70B from Switzerland launched, trained on 15T tokens across 1800 languages under Apache 2.0 license - Tencent open-sourced Hunan MT, a 7B WMT 2025 winning translation model competitive with larger LLMs - Kimiko Kamik2 released minor model updates with coding improvements. ### Big Companies & LLMs - Anthropic raised $13 billion (Series F) at $183B valuation - OpenAI reportedly raised $10 billion (employee share sale) at $500B valuation - OpenAI released GPT real-time API updates (tool support, image input) and acquired Statsig and Alex (coding agent) - Apple released Fast VLM 7B, emphasizing speed with 85x faster time to first token. ### Coding Agents & Performance - Grok Fast Code (Sonic) is now taking 50% of traffic on OpenRouter, noted as very fast - Codex usage is up 10x in two weeks, with improvements noted and potential for Codex 2 hinted at - Statsig and Alex were acquired by OpenAI, with Alex being an AI coding agent for Xcode. ### Key Interviews & Discussions - Roger Jin and Besh Kumar from News Research detailed Hermes 4's development, data mix (5M samples, 56B tokens) using techniques like 'data forge', and its 'soulful' nature, making it ideal for RL and companion use cases - They also explained the Husky Holdem benchmark, where LLMs code poker bots, testing coding and strategic understanding, with models like Gemini 2.5 Pro, Qwen 3, and Hermes 4 competing - Yan Pelleg and LDJ discussed the rapid rise of coding agents and the perceived improvements in Codex and Grok Fast Code. ### Corporate & Partnership News - Corweave acquired OpenPipe, integrating its RL training platform and bringing founders Kyle and David Corbett to the Corweave/Weights & Biases team for a full-stack infrastructure solution.