#231 - Claude Cowork, Anthropic $10B, Deep Delta Learning
Wednesday, 21 January 2026 · 2 min read · Listen to the episode ↗
The podcast episode highlights Anthropic's launch of Cowork, an AI-driven cloud desktop app offering task automation, amidst security concerns and a significant pricing shift reflecting a move toward selling labor. It also discusses Google's beta testing of Gemini, integrating personal intelligence features to streamline user interaction, and touches on the growing demand for AI technologies from companies like NVIDIA, alongside challenges in AI model training and geopolitical implications affecting chip exports.
Anthropic has launched Cowork, a cloud desktop app that automates tasks like organizing files and creating spreadsheets without requiring programming skills. While it offers significant efficiency gains, security concerns arise from its combination of local and web access, prompting Anthropic to use sandboxed virtual machines to safeguard user systems. The pricing for Anthropic's Clawed Max tier has shifted dramatically, now between $100 and $200 a month, reflecting a trend in AI from selling intelligence to selling labor, particularly in B2B contexts.
In related developments, Google is beta testing a personal intelligence feature within its Gemini platform, which integrates with various Google services to enhance user interaction while addressing over-personalization risks. The podcast highlights the shift from traditional search engines to AI-driven chat interfaces, with Google Gemini introducing features like email summarization and an AI inbox, though concerns about the AI's ability to determine message importance persist.
Salesforce has unveiled an AI-powered Slackbot that can draft emails and schedule meetings, indicating a broader trend of AI integration into business applications. Anthropic's recent $10 billion funding round has increased its valuation to $350 billion, with plans for an IPO by late 2026. Meanwhile, XAI has raised $20 billion, bolstered by collaborations with the U.S. Department of Defense.
NVIDIA is experiencing overwhelming demand for its H200 AI chips, with significant orders exceeding current inventory. OpenAI has signed a $10 billion deal with Cerebras for compute, aiming to diversify its sources and reduce latency. The podcast also discusses challenges in AI training and inference, including liquidity issues faced by Coreweave and the valuation of LM Arena at $1.7 billion after a successful funding round.
The conversation touches on Nemetron Cascade, which aims to improve reinforcement learning for general-purpose reasoning models, and the challenges of catastrophic forgetting in AI training. The introduction of deep Delta learning is noted for its potential to enhance neural network architectures. Recursive language models developed at MIT are discussed for their ability to process long prompts effectively.
The podcast also addresses the role of positional embeddings in pre-trained Large Language Models (LLMs) and Anthropic's work on "Constitutional classifiers plus plus" to tackle alignment issues in LLMs. The effectiveness of these classifiers in reducing inappropriate refusals is highlighted, alongside concerns about chip export controls affecting AI training capabilities in China.
Jake Sullivan criticizes past decisions to lift AI chip export controls, arguing they undermine U.S. competitiveness against China. The discussion emphasizes the need for coherent policies amid political rhetoric and the influence of financial interests in shaping AI-related decisions.
This summary was generated from the episode transcript and can contain mistakes.