# Behind Latent Space, ft AIEi Shanghai preview — with swyx and Galen Gao

## Executive summary

This technical discussion, centered around the AIEi Shanghai conference, explores the evolving definition of the AI Engineer role. Key themes include the shift from traditional ML research to practical 'harness' development, the threat posed by coding agents to mid-tier Software as a Service (SaaS) businesses, and the importance of adopting advanced interaction models (like full-duplex communication). The speakers emphasize that successful AI development requires a strong focus on Go-To-Market (GTM) strategy and community building, rather than solely on technical capability.

## Key takeaways

- The AI Engineer Role: Beyond ML Research: The AI Engineer is defined as the individual contributor who integrates AI into traditional software code (a 'harness'), requiring strong software engineering skills rather than advanced math or supercomputing expertise. This role is crucial for building practical, working systems.
- SaaS Vulnerability to Coding Agents: Mid-tier SaaS businesses are highly vulnerable to being cloned by coding agents. A $40,000 SAS can potentially be replicated by a skilled agent in a week, forcing business owners to pivot toward becoming agent builders or specialized data vendors.
- Advanced AI Interaction Models: The future of AI interaction moves beyond simple chat windows (LLM paradigm) toward full-duplex, continual learning models, where communication is constant, and the AI can adjust in real-time, similar to human conversation.
- The Importance of GTM and Taste: Successful product development requires integrating Go-To-Market (GTM) strategy early in the process. 'Taste' is defined as the ability to quickly and cheaply narrow down a vast set of possibilities to the single, obvious best solution.

## Technical details

- AI Engineering Architecture: The modern AI Engineer focuses on building 'harnesses'—integrating AI into existing traditional software code—rather than solely performing advanced ML research (e.g., pre-training, manifold hyperconnections).
- Agent Capabilities and SaaS Threat: Coding agents (like Codex, Cloud, Coding, Devon) can click through every screen of a software and replicate its functionality independently, posing an existential threat to mid-tier SaaS platforms.
- System One vs. System Two Models: Diego advocates for 'System One' models, which are integrated into the path of code (like an `if` statement in a normal programming language), rather than relying on the traditional, isolated chatbot window (LLM paradigm).
- Data and Knowledge Infrastructure: The concept of creating a comprehensive, transcribable, and cataloged dataset (like the AIE World's Fair data set) allows for training LLMs and building specialized knowledge bases (e.g., a 'Wikipedia of AI').

## Practical implications

- For product owners, prioritize building systems that are open-source and highly customizable, rather than relying on proprietary, closed-loop SaaS solutions.
- Focus development efforts on creating 'harnesses' that integrate AI capabilities into existing business workflows, as this is the most immediate and valuable application for AI Engineers.
- When advising startups, emphasize the need for early GTM integration and market validation alongside technical development to ensure product-market fit.
- Build engineers should explore developing internal, customized software solutions for niche business problems, leveraging open-source components and coding agents to bypass expensive, rigid enterprise software.

## Topics

AI Engineering, Software Architecture, SaaS Business Models, Generative AI, Community Building, Go-To-Market Strategy, AIEi Shanghai, The Rise of the AI Engineer, Kill My SaaS, Latent Space

Source: https://www.youtube.com/watch?v=QmDLDy3ZSNE