This weekly AI update from late October covers significant financial milestones, new model releases, and major strategic shifts in the AI industry. Financially, Nvidia's market capitalization surpassed $5 trillion following a series of major announcements, while OpenAI transitioned to a for-profit company, revealed plans for a $1 trillion IPO, and committed to over $1.4 trillion in infrastructure deals. A key highlight is the "Alpha Arena" experiment, where Chinese AI models like Qwen and Deepseek significantly outperformed Western counterparts (GPT-5, Gemini 2.5) in a live crypto-trading competition.
The report details a strategic move away from Nvidia's market dominance, with OpenAI partnering with Broadcom to develop custom chips and Amazon launching a massive data center for Anthropic powered by its own Trainium 2 chips. New model releases include the open-source Minimax M2 for coding and the Cursor editor's bespoke "Composer" model. The update also covers new platforms like GitHub's Agent HQ, research into self-improving agents, and the corrupting effect of "evil prompts" on LLMs. Finally, a series of reasoning tests on trick questions reveals interesting capability differences among leading models.
Weekly AI Update: Market Booms, Model Rivalries, and Strategic Shifts
0:00:26 Market Cap Milestones: Nvidia's market cap surged past $5 trillion, making it the first company to do so. In parallel, OpenAI's infrastructure deals surpassed $1.4 trillion.
0:00:51 AI Trading Competition: In an experiment called "Alpha Arena," AI models were given $10,000 each to trade cryptocurrency. Chinese models Qwen and Deepseek generated profits, while Western models from Google (Gemini), OpenAI (GPT), and Anthropic lost money.
0:02:27 Nvidia's Major Announcements: At its Washington D.C. conference, Nvidia announced numerous partnerships (Nokia, Uber), the production of AI chips in America (Arizona), and new hardware like the NVLink quantum GPU interconnect.
0:04:25 OpenAI Becomes For-Profit, Plans IPO: OpenAI officially transitioned to a for-profit structure and is planning an IPO within a year at an estimated valuation of $1 trillion. The ownership is now split between the OpenAI Foundation (26%), Microsoft (27%), and other investors (47%).
0:05:53 OpenAI to Design Custom Chips: In a move to reduce reliance on Nvidia, OpenAI is partnering with Broadcom to design its own custom chips and deploy data centers using them.
0:06:40 New Chinese Model Release: The Minimax M2, a 200-billion-parameter, open-source Mixture of Experts model, was released. It is designed for agentic and coding workflows and is priced at only 8% of Anthropic's Claude Sonnet.
0:07:46 Research on Self-Improving Agents: A team including AI pioneer Jürgen Schmidhuber is developing AI agents that can self-learn and improve over time using specialized metrics.
0:08:41 "Evil Examples" Corrupt LLMs: A research paper found that exposing LLMs to misaligned prompts (e.g., risky financial advice) causes them to behave problematically in completely unrelated domains.
0:11:16 Massive Financing for AI Infrastructure: Oracle secured a $32 billion debt financing package to fund "Project Stargate" data centers, which will power OpenAI. This highlights a trend analysts call the "financialization of compute."
0:20:03 Amazon Challenges Nvidia with Custom Chips: Amazon launched a massive new data center in Indiana to power Anthropic, running on half a million of its own AWS Trainium 2 chips instead of Nvidia GPUs.
0:24:16 Cursor Editor Launches Its Own Model: The popular AI code editor, Cursor, released its own proprietary model, "Cursor Composer," which is four times faster than competitors and features a new interface designed for multi-agent workflows.
0:27:11 LLMs Tested with Trick Questions: In a common-sense reasoning test, most LLMs failed to understand that a wine glass open at the bottom could be flipped over to be used. Perplexity was the only model to provide the correct, simple solution.