#231 - Claude Cowork, Anthropic $10B, Deep Delta Learning | Last Week in AI

The podcast episode delivers a comprehensive overview of recent developments in AI, spanning tools, business deals, open-source projects, and research advancements. Anthropic's new Co-work tool, an all-purpose desktop agent, is highlighted for its potential to revolutionize computer interaction, alongside security considerations using sandboxed virtual machines. Gemini's introduction of personal intelligence, connecting to users' Google data, sparks discussion about over-personalization and Google's evolving approach to AI integration. Substantial funding rounds for Anthropic and XAI signal continued investment in the AI landscape, while Nvidia grapples with supply chain challenges amid high demand from China. Research highlights include Nemetron Cascade's approach to mitigating catastrophic forgetting and DeepSeek's MHC for improving training stability, showcasing the ongoing efforts to refine and scale AI models.

Outlines

Part 1: Industry News, Agents, and Product Updates

Part 2: AI Business, Funding, and Infrastructure

Part 3: Research, Architectures, and Training Methods

Part 4: Memory, Context, and Scaling

Part 5: Safety, Policy, and Global Competition

Sign in to continue reading, translating and more.

Continue

#231 - Claude Cowork, Anthropic $10B, Deep Delta Learning

Last Week in AI

Part 1: Industry News, Agents, and Product Updates

00:00Podcast Introduction: AI News, Deep-Seek Paper, and Scaling Memory

Podcast Introduction: AI News, Deep-Seek Paper, and Scaling Memory

00:46Anthropic's Co-work: A Desktop Agent for File Management and Automation

Anthropic's Co-work: A Desktop Agent for File Management and Automation

06:14Anthropic's Pricing Shift: Selling Labor with Co-work at $100-$200/Month

Anthropic's Pricing Shift: Selling Labor with Co-work at $100-$200/Month

09:45Gemini's Personal Intelligence: Connecting to Google Apps for Personalized Reasoning

Gemini's Personal Intelligence: Connecting to Google Apps for Personalized Reasoning

10:54Google's Shift: Experimentation and Shipping with Gemini's Personalization

Google's Shift: Experimentation and Shipping with Gemini's Personalization

12:45Google Removes Flawed AI Health Summaries Amidst Safety Concerns

Google Removes Flawed AI Health Summaries Amidst Safety Concerns

14:46AI Overviews Become Standard in Google Search: A Shift in User Behavior

AI Overviews Become Standard in Google Search: A Shift in User Behavior

16:14Gemini Expands Within Gmail: New Features for Email Management

Gemini Expands Within Gmail: New Features for Email Management

18:11Slackbot as an AI Agent: Salesforce Integrates AI for Information Retrieval and Task Automation

Slackbot as an AI Agent: Salesforce Integrates AI for Information Retrieval and Task Automation

Part 2: AI Business, Funding, and Infrastructure

20:11Anthropic's $10 Billion Funding Round: Valuation Doubles to $350 Billion

Anthropic's $10 Billion Funding Round: Valuation Doubles to $350 Billion

22:25XAI Raises $20 Billion: Valued at $230 Billion with Sovereign Wealth Fund Support

XAI Raises $20 Billion: Valued at $230 Billion with Sovereign Wealth Fund Support

23:30XAI's Data Advantage and Nvidia's Supply Chain Miracle

XAI's Data Advantage and Nvidia's Supply Chain Miracle

25:39Nvidia's H200 Supply Chain: Packaging as the Rate Limiter

Nvidia's H200 Supply Chain: Packaging as the Rate Limiter

27:50China's Rush to Buy Nvidia Chips Amidst Export Control Uncertainty

China's Rush to Buy Nvidia Chips Amidst Export Control Uncertainty

29:26OpenAI's $10 Billion Compute Deal with Cerebrus for Inference

OpenAI's $10 Billion Compute Deal with Cerebrus for Inference

30:29Cerebrus' Inference Platform and Delayed IPO Amidst Challenges

Cerebrus' Inference Platform and Delayed IPO Amidst Challenges

31:38CoreWeave Amends Credit Agreements: Liquidity Bridge Due to Hardware Delays

CoreWeave Amends Credit Agreements: Liquidity Bridge Due to Hardware Delays

34:30LLM Arena Valued at $1.7 Billion After $150 Million Series A Funding

LLM Arena Valued at $1.7 Billion After $150 Million Series A Funding

Part 3: Research, Architectures, and Training Methods

35:54Nemetron Cascade: Scaling Reinforcement Learning to Prevent Catastrophic Forgetting

Nemetron Cascade: Scaling Reinforcement Learning to Prevent Catastrophic Forgetting

40:04Nvidia's Open Source Push and the Advantages of RL over Supervised Fine-Tuning

Nvidia's Open Source Push and the Advantages of RL over Supervised Fine-Tuning

43:15DeepSeek's MHC: Manifold Constraint Hyperconnections for Training Stability

DeepSeek's MHC: Manifold Constraint Hyperconnections for Training Stability

49:53iQuest Coder V1: A Coding Model with a Multi-Pass Workflow

iQuest Coder V1: A Coding Model with a Multi-Pass Workflow

53:33iQuest Coder V1's Performance and the Transformer Mamba Hybrid

iQuest Coder V1's Performance and the Transformer Mamba Hybrid

57:23Falcon H1R-7B: A Reasoning Model with Hybrid Transformer and Mamba2 Architecture

Falcon H1R-7B: A Reasoning Model with Hybrid Transformer and Mamba2 Architecture

1:01:14Deep Delta Learning: A More Efficient Way of Designing Data Flow

Deep Delta Learning: A More Efficient Way of Designing Data Flow

1:02:26Deep Delta Learning: Overcoming Limitations in Neural Networks

Deep Delta Learning: Overcoming Limitations in Neural Networks

1:04:42Deep Delta Learning: Enabling Deletion and Flipping of Internal Logic

Deep Delta Learning: Enabling Deletion and Flipping of Internal Logic

Part 4: Memory, Context, and Scaling

1:07:42Recursive Language Models: Scaling to Arbitrarily Long Prompts

Recursive Language Models: Scaling to Arbitrarily Long Prompts

1:09:48Recursive Language Models: Dynamic Peeking and Programmable Search

Recursive Language Models: Dynamic Peeking and Programmable Search

1:13:00Conditional Memory: Augmenting Transformers with External Memory

Conditional Memory: Augmenting Transformers with External Memory

1:14:44Conditional Memory: Looking Up Chunks of Words in a Massive Table

Conditional Memory: Looking Up Chunks of Words in a Massive Table

1:18:54Extending Context by Dropping Positional Embeddings

Extending Context by Dropping Positional Embeddings

1:20:17Positional Embeddings: Rotating Vectors and Telling Time

Positional Embeddings: Rotating Vectors and Telling Time

1:23:04Removing Positional Embeddings: Recalibrating and Generalizing

Removing Positional Embeddings: Recalibrating and Generalizing

1:25:34Learning Positional Information: Starting with and Dropping Embeddings

Learning Positional Information: Starting with and Dropping Embeddings

Part 5: Safety, Policy, and Global Competition

1:26:21Constitutional Classifiers: Defenses Against Universal Jailbreaks

Constitutional Classifiers: Defenses Against Universal Jailbreaks

1:27:56Two-Stage Classifier Cascade: Efficient Production-Grade Defenses