Key Concepts
AI Agents, Reasoning, Multimodality, Model Capabilities, Cursor for X, AI Leapfrog Effect, Co-pilots, Automation, Defensibility, Execution as a Moat.
Capabilities: Reasoning, Agents, and Multimodality
The speaker, Sarah Goa from Conviction, emphasizes that reasoning is a crucial new vector for scaling intelligence. While labs are excited about the increased compute, the focus should be on the new capabilities unlocked, such as transparent high-stakes decisions and sequential problem-solving.
Agents, defined as software that plans, includes AI, takes ownership of a task, and holds a goal in memory, are seeing a surge in development. The number of agent startups has increased by 50% in the last year, indicating real-world applications are emerging.
Multimodality is also progressing rapidly, with companies like HeyGen, ElevenLabs, and Midjourney achieving significant revenue (over $50 million ARR). These companies demonstrate the ability to create realistic voice and video content, including clones with gestures and expressions that reflect emotion. The speaker believes that voice will be the first modality to see widespread adoption in business workflows, but as other modalities become more controllable and less costly, they will all become prevalent.
Model Capabilities and the Competitive Landscape
The speaker dismisses the idea of a "data wall" or the "end of AI summer," asserting that capability improvements are ongoing across the model layer. The market for model capabilities is becoming more competitive, with Sam Altman stating that "last year's model is a commodity."
Data from Open Router shows that Claude has cut into OpenAI's market share, and Google has made a strong comeback with Gemini. New players like SSI and Thinking Machines are also entering the market with orthogonal technical approaches. DeepSeek is releasing competitive base and reasoning models at a fraction of the training cost. The speaker advises planning for a multimodel world and utilizing tools like Open Router or inference platforms like Base 10.
Application Layer: Cursor for X
The success of Cursor, a code editor with AI assistance, is highlighted as a prime example of a killer application. Cursor achieved $1 million to $100 million ARR in 12 months with half a million developers and zero salespeople initially. Cognition, another company focused on AI-powered coding assistance, is also seeing success. Windsurf's acquisition by OpenAI for $3 billion further validates the value in this space.
The speaker analyzes why code was the first domain to see such success:
- Code is structured text with logic.
- Much of coding is sophisticated boilerplate.
- Code has deterministic validation.
- Researchers believe code is crucial for AGI.
- Engineers built tools for engineers, understanding the workflow intimately.
The speaker proposes the concept of "Cursor for X," applying the same principles to other industries. This involves understanding the workflow intimately, redesigning it from first principles around manipulating models, and building products that show up informed, collecting and packaging context automatically.
The speaker emphasizes that the prompt is a bug, not a feature, and the best AI products should feel like mind-reading.
AI Leapfrog Effect and Co-pilots vs. Full Automation
The speaker notes a counterintuitive trend: the most conservative, low-tech industries are adopting AI fastest, which she calls the "AI leapfrog effect." Examples include Sierra (customer service), Harvey (legal), and Open Evidence (medical research). These companies demonstrate that AI is essential for staying competitive in these industries.
While many are excited about full automation, the speaker argues that co-pilots are still underrated. She uses the Iron Man analogy: the suit augments Tony Stark, allowing him to do amazing things, but he is still in control. Human tolerance for failure decreases as latency increases, so building great augmentation is often the best path forward.
Defensibility and Execution
The speaker argues that execution is the moat in AI. Cursor did not invent code completion or the underlying models, but they out-executed their competitors on every dimension, shipping a great experience faster and capturing the hearts and minds of developers. In contrast, Jasper, despite having first-mover advantage and raising $125 million, was quickly surpassed by ChatBT because its first product was simply a series of prompts and a text box.
The speaker concludes that the opportunity in AI is early and massive. The game board keeps getting shaken up with new model releases and capability breakthroughs, creating opportunities to win again and again. She encourages the audience to be the translators for the rest of the world and build something revolutionary.
AI summaries can miss context or contain errors. Check important details against the original video.