From: allin

The landscape of AI development is seeing an emergence of “vertical AI” startups, which are distinct from the well-known general-purpose Large Language Models (LLMs) such as OpenAI’s GPT-4, Google’s Gemini, and Anthropic’s Claude [00:29:00], [00:29:09], [00:29:11]. While general-purpose LLMs are trained on vast internet data to answer questions on almost anything [00:29:13], vertical AI companies are specializing in specific job titles or societal roles [00:29:30].

Examples of Vertical AI Applications

Several startups are pioneering this vertical approach:

Devon: The AI Software Engineer

A notable debut in this space is Devon, a tool from the startup Cognition, touted as an AI software engineer [00:29:54], [00:29:56]. Demos of Devon show it capable of fixing bugs in real-time, fine-tuning AI models, and building applications end-to-end [00:30:03], [00:30:06], [00:30:08]. While not confirmed, speculation suggests Devon might be built on OpenAI’s GPT-4 [00:30:10]. Its CEO stated that Devon was created by enhancing an existing LLM with improved reasoning and long-term planning capabilities [00:30:16], [00:30:18], [00:30:20]. On coding benchmarks, Devon reportedly performs significantly better than generic language models [00:30:24], [00:30:32], [00:30:34].

The Evolution to AI Agents and Automation

Chamath Palihapitiya finds this development “so powerful” and “incredible” given the rapid week-over-week progress [00:30:51], [00:30:54], [00:30:55]. He suggests that what were once “impenetrable job types” for the average person, like being a developer, will become “like a command line interface where you just kind of describe in English what you want to do” [00:31:09], [00:31:12], [00:31:18], [00:31:22]. This automation will broaden the accessibility of these tools [00:31:30].

David Sacks notes that autonomous coding is a core and obvious use case for LLMs because code is text and can be debugged via compilers, theoretically leading to high accuracy [00:32:30], [00:32:38], [00:32:40]. However, he points out that while Devon represents an “agent-first approach” ideal for generating new software projects, it becomes more challenging with existing codebases [00:33:02], [00:33:11]. Other companies, like Sourcegraph with their Cody product, focus on a “context-first” approach for co-piloting within existing codebases [00:33:20], [00:33:26], [00:33:27], [00:33:29]. Regardless of the approach, the overall trend is towards autonomous coding, making human coders significantly more productive [00:33:56], [00:34:05], [00:34:07].

David Friedberg highlights that this evolution moves from “co-pilots” that assist developers to full-fledged “agents” that act as the primary “pilot” [00:32:06], [00:32:10]. The next step is the concept of an “AI conductor[00:36:30].

“I think what we’re seeing in this era is everyone’s taking a vertical human and creating a vertical version of a human um in the AI era” [00:35:17], [00:35:19], [00:35:21]. — David Friedberg [00:35:17]

These specialized tools will emerge as specific vertical tools that people use to automate and scale tasks [00:35:44], [00:35:46]. The value will accrue to the company that offers the best “lawyer service” or “accounting service,” etc., due to fine-tuning and unique data [00:36:00], [00:36:03].

Implications for Jobs and Company Structure

This shift has significant implications for job roles and the structure of companies:

  • Initially, AI acts as a co-pilot, then a peer, and eventually, a “conductor” that coordinates multiple AI agents (e.g., AI lawyer, AI accountant, AI developers, AI designer) [00:36:12], [00:36:26], [00:36:30], [00:36:39].
  • This could lead to millions of companies with just “one person and then a whole layer of software and conductors and agents and bots” [00:38:03], [00:38:05], [00:38:07].
  • Human expertise and creativity will still be essential for thinking through architecture, process, and ensuring AI agents perform their jobs [00:37:34], [00:37:37], [00:37:38].
  • This technology creates “extraordinary leverage for people and organizations,” increasing economic productivity [00:37:45], [00:37:46], [00:37:48].
  • Instead of job losses, people may “level up” their roles [00:37:50], [00:37:52].
  • Operational expenses for companies are projected to be cut in half [00:37:54].

This era of vertical AI promises to democratize the ability to build and operate complex systems, fostering a new wave of solo entrepreneurs or small teams capable of creating substantial businesses [00:38:35], [00:39:02].