Opus 5.5 is proof that we do need smarter models

Summary

Theo argues that Opus 5.5 changed his view of AI progress. More capable models are not merely making existing coding tasks faster or cheaper; they are expanding what people can attempt—and who can attempt it. He says this benefits professional developers, technically minded non-developers, and eventually people with little technical background.

How model improvements broaden what people can do

Theo describes a progression in AI-assisted work:

• Helping with work a person already understands, such as autocomplete or targeted code changes.
• Customizing tools for personal use, such as creating a browser extension or adapting a library or app.
• Building and sharing software beyond one’s expertise, including projects in languages a person does not know well.
• Modifying or rebuilding existing software, such as reverse-engineering games or combining components that once seemed impractical to work with.

He says model improvements affect these groups differently. In his view, Opus 5.5 makes more ambitious work accessible to technically capable people who have not become professional programmers. He points to people using AI subscriptions to port maps from older games into new ones and share modifications with friends.

Examples of AI-enabled work

• Lawn: Theo says he built an open-source alternative to Frame.io after finding it slow, expensive, and inconvenient for his team’s video-review workflow. He estimates that, at the time of Frame.io’s Adobe acquisition, Lawn covered roughly 70% of its functionality, plus features he says Frame.io still lacks. He reports that a few hundred people use Lawn in video workflows and that others have forked it.
• TypeScript compiler rewrite: Theo describes using AI to rewrite the compiler in Rust. He says it became usable on real projects in under 20 hours, runs two to three times faster than TSGO, works in WebAssembly, and reaches 99.99% accuracy compared with other TypeScript versions. These are his reported results.
• Game reverse engineering and mashups: He cites projects that decompile or port older games and combine them with other games or engines as evidence that AI can help people work with codebases they might otherwise be unable to understand.

Implications for developers

Theo’s broader argument is that years of working with established tools and practices can constrain developers through ingrained habits and assumptions. People who are less familiar with those constraints may be more willing to try unusual combinations or question why things are done a certain way.

He urges developers to:

• Ask why a practice exists rather than assuming it is necessary.
• Attempt unconventional projects, even when they may fail.
• Treat discarded work and rejected pull requests as normal rather than as personal losses.
• Choose tools and technical approaches based on the goal, not personal preference.
• Stay flexible about working at different levels of abstraction.

As an example of adapting tools for AI agents rather than himself, Theo describes a customized Ubuntu-based system he is building. He says it uses Bash instead of his preferred Zsh because his agents make fewer errors with Bash, and it is designed to provide graphical environments for agents.

Browserbase sponsor segment

Theo presents Browserbase as infrastructure for AI agents that need to interact with websites. He highlights:

• Signing into services through a linked 1Password account.
• Managing and persisting browser sessions so agents do not have to sign in repeatedly.
• A Search API for finding web results.
• A Fetch API that can return page content as HTML, JSON, or Markdown.

He contrasts this with approaches that require manually opening a cloud computer to authenticate or rely on curl to access websites.

Format

This is an opinion and analysis video, not a step-by-step guide or formal product review. Its central claim is that increasingly capable models will make software creation and customization accessible to more people, while challenging developers to rethink their assumptions.

Main speakers and sources

• Theo (T3.gg): Main narrator and source of the analysis, project examples, and performance claims.
• Audience chat and commenters: Quoted or referenced by Theo during the discussion.
• Browserbase: Sponsor featured in the ad segment.