# ThursdAI - Jan 8 - NVIDIA's Vera Rubin 5x Leap, XAI's $20B Raise, Ralph Wiggum Takes Over Coding Source: https://www.youtube.com/watch?v=_8zf5BtF-Ts Recap page: https://rapidrecap.app/video/_8zf5BtF-Ts Generated: 2026-01-09T00:35:26.732+00:00 --- ## Quick Overview The ThursdAI January 8th episode covered major AI news including NVIDIA's announcement of the Vera Rubin platform featuring new server GPUs delivering five times inference performance over Blackwell, XAI raising $20 billion at a $230 billion valuation, the viral explosion of the Ralph Wiggum coding technique popularized by Ryan, and Wolfram officially joining the Weights & Biases evangelism team. **Key Points:** - NVIDIA announced the Vera Rubin platform in full production at CES, delivering five times inference performance over Blackwell, with six new chips featuring 88 Arm cores and significant interconnect bandwidth improvements. - XAI announced a Series E raise of another $20 billion at a $230 billion valuation, bringing massive GPU resources to Grok. - Ryan's article detailing the Ralph Wiggum coding technique, created by Jeff Huntley, achieved 1.1 million views, leading to Ryan shipping three new features using the technique concurrently. - Wolfram officially joined the Weights & Biases evangelism team, focusing on evaluation (eval) and promising to bring goodness related to it to the show. - OpenAI launched GPT Health, a privacy-first space for personalized health conversations connected with electronic health records, currently available via a waitlist. - Google integrated Gemini 3 into Gmail for 3 billion users, powering features like AI overuse, smart replies, and an AI inbox. - Upstage released Solar Open, a 102 billion parameter Mixture of Experts (MoE) model trained on 19.7 trillion tokens, which also features notable Korean language optimization. **Context:** The hosts Alex Volkov, Ryan, and Wolfram welcomed co-hosts and guests, including LDJ and Nisten, for the first ThursdAI show of the new year on January 8th, marking their return from a short break. The show immediately highlighted the massive influx of AI news over the holidays, focusing on major announcements from CES, significant corporate funding rounds, and the viral success of new coding methodologies. ## Detailed Analysis The broadcast covered extensive AI developments starting with open source releases, where Upstage released Solar Open, a 102B parameter MoE model trained on nearly 20 trillion tokens with specific Korean language optimization, and Miro Mind AI launched Miro Thinker 1.5, a 30B parameter open source search agent using 'interactive scaling' that outperforms larger models on specific benchmarks, highlighting the importance of agent harnesses. Additionally, ZAI, the maker of GLM models, went public, raising $558 million, and Nous Research released Nous Coder 14B, achieving a 7% jump on the Live Code Bench in four days via RL training, with results tracked via Weights & Biases Reports. In big company news, NVIDIA stole the spotlight at CES by announcing the Vera Rubin platform is in full production, delivering a massive five-times inference performance leap over Blackwell with significantly improved bandwidth and efficiency, while XAI secured a staggering $20 billion raise. Other major updates included Amazon launching Alexa Plus on the web, OpenAI opening a waitlist for GPT Health, and Google integrating Gemini 3 features into Gmail for 3 billion users. The segment on coding tools focused heavily on Ralph Wiggum, a technique that went viral due to Ryan's article, emphasizing that smaller, specialized models combined with excellent harnesses can beat larger generic models. ### Open Source Model Releases - Solar Open, a 102B parameter MOE from Upstage, trained on 19.7T tokens, showed strong benchmarks and Korean optimization - Miro Thinker 1.5, a 30B search agent, uses interactive scaling to beat trillion-parameter models on agent search benchmarks - Liquid AI released LFM 2.5, tiny on-device foundation models with text, vision, and audio support running efficiently on consumer hardware like 82 tokens/sec on Snapdragon Gen 4. ### Corporate and Funding News - XAI raised $20 billion at a $230 billion valuation, bringing GPUs to Grok - ZAI, maker of GLM models, IPO'd on the Hong Kong stock exchange, raising $558 million - Amazon launched Alexa Plus on the web for Prime members. ### NVIDIA Vera Rubin Platform Deep Dive - Vera Rubin is the next generation of AI processors, delivering 5x inference performance over Blackwell, with 75% fewer GPUs needed for 10-trillion parameter MoE training - The platform includes six chip arrays, Vera CPUs with 88 Arm cores, and high-bandwidth NVLink interconnects. ### Viral Coding Techniques and Agent Harnesses - Ryan's article on the Ralph Wiggum coding technique went viral with 1.1 million views, emphasizing the power of clever orchestration - The discussion stressed that smaller models with great harnesses can outperform larger models, making harnesses critical for employing open-source and smaller models affordably. ### Audio and Health AI Updates - OpenAI launched GPT Health, a waitlist for personalized health conversations integrated with EHRs - Doc launched its first US pilot allowing AI to autonomously prescribe medication - LFM 2.5 audio model supports unified audio input/output, running locally on CPU/RAM with 8x faster audio token processing than the standard ASR-LLM-TTS pipeline. ### Notable Model Performance - Nous Research's Nous Coder 14B achieved a 7% jump on Live Code Bench in four days using RL training, with results tracked via a detailed Weights & Biases Report dashboard.