The AI bubble is popping

Summary

The discussion argues that AI progress is accelerating even as technology stocks fall. The speakers distinguish financial-market performance from technological adoption: falling share prices may reflect the high cost of data centers, limited GPU and electricity supplies, and borrowing costs—not necessarily weakening demand for AI.

They compare today’s AI industry with the dot-com era. In their view, major companies across the AI supply chain are already generating substantial revenue, while the returns for businesses using AI are still emerging. These market and investment comments are the speakers’ opinions, not financial advice.

Models and Development Workflows

• The speakers expect several major model releases soon, including versions referred to in the subtitles as Fable 5.5, GPT 6.2 Astra, and Google’s Carbon. They expect models to keep improving and reject the idea that AI progress has peaked.
• David says Opus 5.5 is his primary coding model, used in roughly 95% of his sessions because he considers it both capable and fast. For substantial or risky changes, he asks a GPT model—such as Astra at high reasoning effort—to review the work. He says these reviews commonly identify issues that the primary model then fixes.
• He argues that developers should use frontier models throughout software development, at minimum to test and review code for bugs and security problems. In his view, using AI for only a small portion of development limits productivity gains. He says his own software output is far greater than it was two years ago.
• He also describes stronger models as enabling work beyond coding, particularly in visual design, motion graphics, animation, video editing, websites, and advertising. For example, he says he produces launch videos frequently and creates custom graphics from natural-language descriptions.

Business Strategy, Jobs, and AI Agents

• The proposed business “moat” is speed of adaptation: responding quickly to customer needs, market changes, and security vulnerabilities. He argues that slow companies can be overtaken, while adaptable companies can change products and marketing rapidly.
• Because AI capabilities are changing quickly, he advises against rigid long-term roadmaps. Instead, businesses should keep learning, respond to customers, and build products that benefit from increasingly capable models.
• He expects agents to handle a growing share of routine work, but argues that automation will also create new roles and tasks—especially coordinating agents, designing agent workflows, and managing “software factories.” He attributes the continued presence of some seemingly outdated jobs and businesses to institutional inertia.
• Businesses that add friction, charge for information AI can provide cheaply, or rely on customer lock-in may be vulnerable. He believes products that genuinely serve users and work well with agents are better positioned.
• He favors using both frontier cloud models and self-hosted models rather than treating them as mutually exclusive. Frontier models are recommended for demanding work; local models may suit tasks where privacy or cost matters.

Cybersecurity and Cloudroom

The speakers expect AI to increase both the scale of cyberattacks and the need for automated defense. They recommend using AI agents to repeatedly test one’s own code and systems, including by simulating attacks, rather than waiting for attackers to find weaknesses. They stress that this means defensive testing of systems the user owns or is authorized to assess.

David presents Cloudroom as a free, open-source tool for organizing agent workflows. Described features include:

• Preset, one-click agent cycles, including templates for defensive cybersecurity work.
• Scanning agent sessions from tools such as Hermes, Codex, Claude Code, Pi, Cursor, and others.
• Analyzing sessions to identify recurring work patterns and help users create agent cycles tailored to their tasks.
• Designing cycles so that routine, well-understood work can run autonomously, while uncertain or higher-risk work waits for human approval.

The discussion also mentions Deep API, a previous startup described as providing agents with a unified API for tasks such as data gathering, web search, and email. David says its usage increased as customers adopted more agents, presenting it as an example of a product that benefits from AI adoption.

San Francisco and the Business Environment

David describes San Francisco as a concentrated hub for AI talent, startups, and investment, while criticizing the city’s safety, housing costs, limited construction, and regulation. He contrasts the relative ease of starting and operating a company in the United States with the administrative burden he associates with parts of Europe. His personal plan is to maintain a team in Europe while establishing an office in San Francisco.

Practical Recommendations

For the next 90 days, David recommends:

1. Use at least one leading frontier-model subscription and learn how to work effectively with it.
2. Improve AI skills through regular use and structured learning.
3. Review existing software—especially work written manually or with weaker models—with a frontier model, focusing on authentication, APIs, backend security, and other vulnerabilities.
4. Reduce unnecessary software dependencies and strengthen cybersecurity.
5. Stay adaptable rather than assuming that current job roles, products, or business models will remain unchanged.

Reviews, Guides, and References

• Model-review workflow: Use a primary coding model, then ask another strong model to review larger or riskier changes.
• Cybersecurity guide and tool demonstration: Cloudroom’s agent-cycle presets and session scanning are presented as ways to start automating defensive review.
• External references: The speakers recommend an Olin podcast about AI data-center financing and a Sequoia article by Sean Maguire about SpaceX. Neither is reviewed in detail in the subtitles.

Main Speakers and Sources

• David Ondrej: Main commentator on AI models, software development, business strategy, cybersecurity, and Cloudroom.
• Interviewer: Asks questions and prompts the discussion.
• Other sources mentioned: An Olin podcast; a Sequoia article by Sean Maguire; comments attributed to Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman; and examples involving Linus Torvalds and Tim Sweeney.