# The Sweet Side of the Bitter Lesson — Thariq Shihipar, Independent

## Executive summary

This talk argues that the future of software development is not defined by the 'bitter lesson' (alien AI forcing a single workflow), but by a 'Renaissance' where human creativity and control are paramount. The speaker introduces a methodology that separates core engine code from human-readable specifications (a custom scripting language), allowing developers to maintain granular control over complex AI-generated features, moving beyond simple rapid prototyping.

## Key takeaways

- AI is not limited to single paradigms (e.g., Bash or React).: The initial fear that models would force developers into a single language (like CLI/Bash or React) has proven false; models can now generate code in HTML, React, Swift, and Figma proprietary JSON formats.
- The focus must shift from 'jagged frontier' to 'smooth interior'.: While capability increases (the jagged frontier) are visible in benchmarks, the more important aspect is the breadth and quality of tasks the models can handle (the smooth interior), allowing for high-quality, Renaissance-worthy software.
- Achieving control through spec-driven development.: The speaker's method involves maintaining two codebases: the engine code and a 'spec' folder containing a custom scripting language that is readable and linted, allowing the developer to tune complex AI-generated features (e.g., tuning a 'rock move').
- The goal is a 'Renaissance' of software.: The ultimate aim is not just functional software, but high-quality software where every marginal bit of creativity and utility is economically valuable, requiring the developer to be an expert in their domain.

## Technical details

- AI Output Diversity: Models can generate similar results across various formats, including HTML, React, Swift, and the Figma proprietary JSON format, challenging the notion that one technology will dominate.
- LLM Knowledge Visualization: The speaker references a visualization (using Claude models) where inputting coordinates (longitude/latitude) allows the model to plot a map, demonstrating the general knowledge base of models like Opus 5.5.
- Custom Scripting Language: The speaker has created a custom scripting language for a fighting game, which is used to define and tune character abilities (e.g., 'rock move'), allowing for granular control over engine logic without needing to understand the underlying implementation.

## Practical implications

- Build processes must evolve to incorporate human-readable specifications alongside complex engine code to maintain creative and functional control over AI-generated features.
- Developers should focus on defining high-quality requirements (the 'Renaissance') rather than merely achieving rapid functionality (the 'jagged frontier').
- The use of specialized tools (like custom scripting languages or advanced IDE features) is necessary to bridge the gap between LLM output and high-quality, maintainable software.

## Topics

AI Engineering, Software Architecture, Game Development, Developer Workflow, LLM Tooling, Thariq Shihipar, Anthropic, Claude, Opus 4.6 / Opus 5.5, Cursor

Source: https://www.youtube.com/watch?v=xRJMiaUoD_8
