The Bootstrapped Founder

417: The Best Tech Stack in the Age of AI

October 3, 2025

Key Takeaways Copied to clipboard!

  • Even in the age of AI, the best tech stack remains the one the founder already knows because the capacity to understand, review, and debug AI-generated code is paramount for maintainability and ownership. 
  • While AI coding tools are becoming proficient in many languages due to extensive training data, relying on a technology you don't understand means outsourcing control and potentially building a black box system based on past collective implementation strategies rather than your current needs. 
  • Founders should prioritize 'you-centric' tech choices over 'AI-centric' ones, meaning they should deeply understand at least one programming language with a strong ecosystem, rather than outsourcing foundational knowledge to an AI tool. 

Segments

Original Tech Stack Philosophy
Copied to clipboard!
(00:00:09)
  • Key Takeaway: The initial premise is that the best tech stack is the one a founder already knows, validated by experienced founders avoiding time investment in new, unproven technologies.
  • Summary: The core belief is that the best tech stack is the one already known, resonating strongly with experienced founders. These founders recognize that investing significant time learning new technology often detracts from the primary goal of quickly building a business. This principle is contrasted with the emerging influence of AI on tech stack decisions.
AI’s Influence on Tech Stack
Copied to clipboard!
(00:01:21)
  • Key Takeaway: A counter-argument suggests the best tech stack might be the one the coding AI knows best, which is partially true for non-coders relying on AI output.
  • Summary: A peer suggested the best stack is what the AI coding tool knows most about, especially for those not writing code themselves. However, the speaker argues this is only ’true-ish’ because the quality of AI-generated code still depends on the human’s capacity to review, judge, and debug it.
How AI Models Learn Code
Copied to clipboard!
(00:02:26)
  • Key Takeaway: AI models are token guessers trained on massive public datasets, leading to superior performance in languages popular at the time of training, like JavaScript.
  • Summary: AI models function by guessing the next token based on ingested data from millions of documents, codebases, and forums. Training data primarily comes from publicly available code on platforms like GitHub, meaning the AI is inherently best at languages and frameworks that were most popular historically, such as JavaScript.
Esoteric Languages and AI Limits
Copied to clipboard!
(00:06:03)
  • Key Takeaway: Niche or esoteric languages may have less training data, but the main limitation for AI knowledge is technology created after the model’s baseline training.
  • Summary: While rare languages like Erlang or Smalltalk have historical code available, the AI’s knowledge is limited to its training corpus. New languages or frameworks released after the AI’s last major training update will not be inherently known unless external documentation is ingested.
MCP and On-the-Fly Updates
Copied to clipboard!
(00:08:01)
  • Key Takeaway: Model Context Protocol (MCP) and tool calling integrations allow modern AI systems to ingest documentation on the fly, mitigating the knowledge cutoff problem.
  • Summary: The limitations of static training data are overcome by MCP solutions and tool calling, which enable AI systems to ingest the full documentation for any framework in real-time. This instrumentation allows capable AI coding systems to effectively code in virtually any chosen language or framework, regardless of its historical representation in the base model.
Understanding Code is Crucial
Copied to clipboard!
(00:09:40)
  • Key Takeaway: Using an AI to build in a language you do not understand results in a black box system where you cannot effectively debug, expand, or maintain the code.
  • Summary: If a founder cannot innately understand the Rust backend code generated by an AI, they are building a black box that only works or fails without clear direction for fixes. The capacity to see errors, understand infrastructure, and ensure maintainability remains the responsibility of the instructing person.
Future-Proofing Tech Choices
Copied to clipboard!
(00:12:37)
  • Key Takeaway: Failing to understand AI-generated code means the output reflects past collective strategies, not the founder’s specific vision, leading to potential business stress and costly rewrites.
  • Summary: If you do not understand the code, the AI builds what past users of that technology thought was best, which may not align with your current business goals. Outsourcing this understanding to AI is outsourcing control and ownership, potentially leading to complex Frankenstein codebases or the need to rewrite projects later.