# 🎂 ThursdAI 3rd Birthday: Karpathy's Auto Researcher, Fly Brain Upload, OpenClaw in China & more Source: https://www.youtube.com/watch?v=a2D_yLEyYqw Recap page: https://rapidrecap.app/video/a2D_yLEyYqw Generated: 2026-03-13T02:03:16.477+00:00 --- ## Quick Overview The third anniversary episode of ThursdAI celebrated three years of tracking AI developments since the launch of GPT-4, highlighting massive cost efficiency improvements in LLMs, the emergence of autonomous AI researchers like Karpathy's Auto Researcher, the simulation of a complete fruit fly brain with 91% behavioral accuracy by Eon Systems, and the overwhelming adoption of OpenCLAW in China, signaling significant acceleration toward advanced AI capabilities. **Key Points:** - The podcast celebrated its third anniversary, marking the start date near GPT-4's launch on March 14th, 2023, noting context window costs have drastically decreased, with one host spending a billion tokens in 24 hours without hitting a $200 limit. - Andrej Karpathy released an autonomous AI agent researcher that achieved an 11% speedup in GPT-2 training time by autonomously modifying training code, which hosts view as a 'mini singularity moment' showing scaling improvements. - Eon Systems uploaded the first complete fruit fly brain connectome into a physics simulation, achieving 91% behavioral accuracy with 140,000 neurons and 50 million synapses, prompting comparisons to uploaded intelligence concepts like the Netflix show Pantheon. - OpenCLAW is seeing massive adoption in China, with 50% of tracked instances running there, leading to people lining up for installation subsidies, though the central government later warned against security risks. - Nvidia announced a $26 billion commitment over five years to build the world's best open-source models, coinciding with the launch of their 120-billion-parameter Nemotron 3 Super model featuring a 1 million context window. - XAI quietly released Grok 4.20 via API, which features a 2 million token context window but is notably 10 times more expensive per million tokens than its predecessor, 4.1. - Google launched Gemini Embeddings 2, natively supporting multimodal data including text, images, and video within the same vector space, enabling advanced semantic sorting and querying. **Context:** The hosts, Alex Volkov, Wolfram Ravenwolf, Nisten Tahiraj, and Ryan Carson, celebrated the third anniversary of their weekly live show, ThursdAI, which began covering AI news around the time of the GPT-4 release in March 2023. The discussion reflects on the rapid evolution of AI over three years, contrasting early models with current capabilities, and focuses heavily on recent 'singularity updates' suggesting accelerated progress in autonomous research and biological simulation. ## Detailed Analysis The episode heavily featured 'Singularity Updates,' starting with Andrej Karpathy's Auto Researcher, an autonomous agent that ran 700 experiments, discarded ineffective ones, and achieved an 11% reduction in GPT-2 training time by modifying the training code itself, which hosts suggest is a major breakthrough showing scaling improvements that generalize to larger models. This was immediately followed by the announcement that Eon Systems simulated a complete fruit fly brain with 91% behavioral accuracy, a development seen as extremely significant for uploaded intelligence concepts. Furthermore, OpenCLAW has become a cultural phenomenon in China, where local governments subsidize installations, leading to thousands lining up, although the central government issued security warnings against the software. In product news, Nvidia released Nemotron 3 Super (120B parameters, 1M context) and committed $26 billion to open-source models, while Google released natively multimodal Gemini Embeddings 2. The hosts emphasized that despite the hype surrounding OpenCLAW, the technology remains difficult to implement effectively, echoing early challenges seen with Linux adoption. ### Anniversary Reflection - Podcast began covering AI post-GPT-4 launch on March 14, 2023 - Context window pricing has dropped significantly, enabling hosts to use billions of tokens cheaply - Open source models are now competitive with early proprietary releases running locally. ### Singularity Updates - Karpathy's Auto Researcher achieved an 11% speedup in GPT-2 training by autonomously modifying training code, showing scaling improvements - Eon Systems simulated a fruit fly brain with 91% behavioral accuracy, connecting 140,000 neurons and 50 million synapses. ### OpenCLAW Phenomenon - OpenCLAW has dominated adoption in China, with 50% of tracked instances originating there; local governments subsidize installation - The adoption is driven by its utility as an always-on assistant, leading to a 'mimetic velocity' where people widely adopt the concept. ### Big Companies & APIs - XAI quietly released Grok 4.20 API with a 2 million token context window but at a 10x higher cost than 4.1 - Google launched Gemini Embeddings 2, natively supporting text, images, and video in a unified vector space. ### Open Source & Tools - Nvidia launched Nemotron 3 Super (120B MoE, 1M context) and committed $26B to open source over five years - The open standard ACP for AI agent integration is gaining traction over the dying MCP standard - News Research's Hermes agent continues to gain interest. ### Agent Engineering Insights - Ryan Carson reports his agents are 'on fire' using the open-source orchestration tool Symphony - LDJ noted that Claude is demonstrating new capabilities like generating 'YouTube poop style videos' using external tools like EPIfg.