Dylan Patel: $300B OpenAI deal, Codex vs Claude Code, Gemini 3.0

Key Concepts

• AI Model Competition: OpenAI vs. Anthropic, Codex vs. Claude Code.
• Compute Infrastructure: Importance of compute power, Oracle's role, Microsoft's potential regret.
• AGI/ASI: Artificial General Intelligence and Artificial Superintelligence, timelines, and societal impact.
• Tool Use & Agency: AI's ability to use tools, extend context, and perform tasks autonomously.
• Multimodality: Training AI on various data types (text, images, video).
• AI Startup Strategy: Bootstrapping, sales focus, targeting low-hanging fruit.
• Reasoning Effort: The amount of computational resources and time an AI model dedicates to processing a query.
• MCP (Meta-Context Protocol): A hypothetical standard for accessing and sharing data across different applications and platforms.
• Computer Use: The ability of an AI to interact with and control computer systems and applications.
• Sparse Attention: A mechanism for AI models to selectively focus on the most relevant information in a large context.

OpenAI vs. Anthropic: A Battle for AI Supremacy

• Revenue Growth: Anthropic's revenue growth rate suggests it could surpass OpenAI by 2027. This is a significant point, highlighting Anthropic's rapid progress despite OpenAI's head start.
• Model Strengths:
  • OpenAI (GPT-5): Stronger in general chat, but initially behind in coding and computer use. Making strides to catch up with Codex improvements.
  • Anthropic (Claude): Excels in coding, computer use, and productive tasks.
• Company Culture: Anthropic is described as a "cult" with a unified vision focused on AGI, particularly on solving software engineering and computer use to accelerate AI development. OpenAI has broader goals, including search and consumer monetization, which may dilute its focus.
• Leadership: Anthropic's seven co-founders, with diverse technical expertise, foster a more dynamic environment compared to OpenAI's leadership structure.
• Coding Market Focus: OpenAI is actively trying to capture the software engineer market, recognizing its high revenue potential. Codex is improving rapidly, with daily updates to the CLI and extension, making it competitive with Claude Code.

Compute Infrastructure and the Oracle Deal

• Unprecedented Deal: The OpenAI-Oracle deal is described as potentially the largest contract ever announced between two companies, involving many gigawatts of compute.
• Oracle's Guidance: Oracle's unprecedented four-year revenue guidance following the deal demonstrates its confidence and visibility into future earnings.
• Microsoft's Regret: Microsoft may regret not securing the Oracle deal, as it hesitated due to concerns about OpenAI's ability to pay for the massive compute requirements (estimated at $300 billion).
• Hyperscaler Independence: OpenAI benefits from its relationship with Oracle, as it is not tied to hyperscalers that directly compete with its AI services.

Codex vs. Claude Code: The Coding Wars

• Recent Improvements to Codex: Codex has significantly improved in the last month, becoming a strong competitor to Claude Code, especially the Codex CLI and extension.
• User Preferences: While the speaker's company still prefers Claude Code, many users are switching to Codex due to its increased power.
• Reasoning Effort: GPT-5 with high reasoning effort is considered significantly smarter (at least 10 IQ points) than Opus 4.1, but Opus may be stronger if given the same reasoning capacity.
• Model Certainty: Anthropic models tend to jump to conclusions more readily, while Codex with GPT-5 is more uncertain and struggles to provide clear conclusions.
• Workflow Integration: Claude Code is easier to use and better at explaining bugs, while Codex is more powerful for complex debugging.
• Hybrid Approach: The speaker recommends using both Codex and Claude Code, leveraging their respective strengths for different tasks.

The Future of AI: Tool Use, Context, and Generalization

• Tool Use: OpenAI is better at tool use than Anthropic, enabling it to solve problems by calling up Python code or other tools.
• Context Extension: Extending the context window is crucial for AI to maintain memory and perform complex tasks.
• Environments: Training AI in diverse environments (e.g., competition code vs. Salesforce dev environment) will lead to greater divergence in model capabilities.
• Generalization: The key is to move beyond specialization and achieve generalization, enabling AI to apply skills learned in one domain (e.g., coding) to others (e.g., medicine, writing).
• Computer Use: The ability of AI to interact with and control computer systems is essential for automating tasks and improving efficiency.

The Role of Data and Multimodality

• Data Advantage: Google has a significant advantage in multimodality due to its access to vast amounts of image and video data, particularly on YouTube.
• Data Filtering: The challenge is to filter and index this data effectively, as training on all of it is not feasible.
• Multimodal Gains: Significant gains can still be achieved in pre-training by leveraging multimodal data, but this requires a substantial increase in compute power.

XAI and Elon Musk's Position

• Catching Up: XAI has caught up quickly but has not yet taken a leadership position in most areas, except potentially in empathy and emotional AI.
• Funding Needs: XAI needs to raise more money to build Colossus 2, which may be challenging given that it is still behind competitors.

Hype vs. Reality: A Balanced Perspective

• CEO Hype: CEOs tend to overhype AI capabilities to create urgency and attract investment.
• AGI Timeline: The speaker believes that AGI/ASI is likely 2030+, but the definition is nebulous.
• Transformational Change Without AGI: Significant societal transformation can occur through AI automation even without achieving AGI.

Advice for AI Startups

• Bootstrapping: Avoid raising excessive funding, as productivity tools enable building application-layer companies with minimal capital.
• Sales Focus: Founders must be proficient in sales and have a clear vision.
• Low-Hanging Fruit: Target easily achievable, revenue-generating opportunities rather than tackling complex, regulated industries.
• Fail Fast: Embrace experimentation and rapid iteration to learn quickly from failures.

Mobile vs. Desktop: The Future of AI Interaction

• Computer Use on Desktop: Computer use is easier to implement on desktop computers due to the open nature of the operating system.
• Mobile Challenges: Accessing and integrating data across apps on mobile devices is more challenging due to platform restrictions (particularly on iOS).
• Apple's Potential: Apple has the potential to implement computer use on mobile by running virtual iPhones in the cloud, but this remains to be seen.
• MCP Limitations: While MCPs could facilitate data sharing, companies may be reluctant to provide access to proprietary data, especially for sensitive information like internal communications.

Conclusion

The AI landscape is rapidly evolving, with intense competition between OpenAI and Anthropic. While OpenAI currently leads in some areas, Anthropic's unique culture, focus on coding and computer use, and rapid revenue growth position it as a formidable contender. The future of AI hinges on advancements in tool use, context extension, multimodality, and the ability to generalize skills across domains. For AI startups, a focus on bootstrapping, sales, and targeting low-hanging fruit is crucial for success. The integration of AI with computer systems and mobile devices presents both opportunities and challenges, with the potential to transform how we interact with technology and automate tasks across various industries.