▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=3OshsLm3_OQ
• Original title :: Opus 5.5 is proof that we do need smarter models

Theo reverses his earlier "we don't need smarter models" stance. The trigger is the wave of decompiled, reverse-engineered and mashed-up games (Halo x Call of Duty, Xbox 360 ports to Rust + Bevy) built by people who are not developers. He lays out tiers of what AI lets people do - from helping with what one already knows, to building things one does not understand, to reverse-engineering software at its core - and argues each model jump moves different groups (devs, "technical people", non-technical people) by different amounts. Opus 5.5 did little new for him personally, but as one capable, well-priced model on a subscription it pushed technical non-devs deep into building things they don't understand. His spicy take: devs are not very creative, cling to "how it's always been done" and feel too bad about thrown-away work, so unconstrained technical outsiders may run laps around them.

Why his take changed

• Two years ago he predicted a plateau in model intelligence; later he thought models were already smart enough that extra intelligence would go unused. He felt that with Fable 5.
• Opus 5.5 changed his mind not because it is faster or cheaper for his own work, but because it is smarter in a way that benefits far more people than bleeding-edge devs.
• The evidence: game mashups and decompiled ports built by people who never coded, never opened Blender, paying ~$200/month for a Claude subscription to port maps from old favourite games into reverse-engineered ones.

Smarter models help devs ship more

• Since roughly Sonnet 4, devs could get through their roadmaps faster. Theo claims he now ships more code in a week than in all of last year, and that half of his AI-written code beats most devs' hand-written code.
• Example: Lawn, an open-source Frame.io alternative built in a few days (with Gemini 3 Pro / Opus 4.6-era models). Frame.io sold to Adobe for $1.3B in 2021; Lawn covers ~70% of its functionality at that point plus extras, and a few hundred people use it for YouTube video pipelines.

Technical people: the group that moved most

• "Technical people": smart, comfortable with computers, maybe dabbled in code but not developers (Linus Tech Tips as an archetype). They know what software does, not the limits of frameworks.
• As models improve, they can ask "why is my app slow?" and get the analysis, fork tools they rely on, debug systems beyond their understanding, and inject scripts into apps like Discord or ChatGPT.
• Until recently that work stayed personal - models weren't good at packaging work so others could adopt it.

Tiers of what AI enables

1. Helps with what you are already good at (autocomplete, editing a function you understand).
2. Customizes the things you use - forks, personal Chrome extensions, Discord scripts - still inside your own bubble.
3. Builds things you don't fully understand - Theo ships Rust projects although he barely knows Rust.
4. Decompiles, reverse-engineers and rebuilds existing software at its core (e.g. the Super Mario 64 decompilation nobody fully understands).

• Each model jump moves groups differently: OG Copilot helped only devs at tier 1; tool-using models let technical people edit websites; Opus 4.5 let technical people build apps that "kind of worked" while devs went deeper (side-project games, ports, decompilation). Fable 5 and Astra mostly widened the gap in favour of devs.
• Opus 5.5 reversed that: one well-priced model on subscription tiers pushed technical people into tier 3 and 4.
• Non-technical people have barely moved yet - an opportunity for the "person who fixed mom's computer" to build tools or mods and share them with friends.

Devs are not very creative

• Devs are generally poor at anything except delivering code; that is why teams pay non-technical designers and PMs. Theo says his channel works because he is "6/10 at communication where the average is 3".
• Technical non-devs are the people who abused Zapier or built OBS setups with 40,000 scenes used as dynamic layers - weird, creative thinking that now has its shackles off.
• They don't "know" that an IW4 game shouldn't run inside a UE3 engine, so they try it.
• His "why can't we commit our environment files?" short works as a creative-thinking test for engineers; comments show many can't question assumptions like "every file in Git is shared with everyone".
• Like music scenes where outsiders remix a genre, people who don't know Java from JavaScript will build surprising things because they don't know not to ask.

Willingness to fail and throw work away

• Devs expect PRs to merge, and many have guilt-merged a PR. A dev who spends 2 hours generating a 2,000-line PR feels worse about it being closed than a technical person who spent 5 hours, because devs value their time and expect results; technical people treat it as a fun experiment.
• YOLO projects are now cheap: Theo's TypeScript-compiler-in-Rust project, started as "a thing AI would never do", worked in under 20 hours, is 2-3x faster than tsgo, runs in WASM, and matches TypeScript on real projects at 99.99%.
• "Try first, then understand why not."

Be flexible about depth and preferences

• Winners are people flexible to go higher or lower in abstraction and drop personal preferences for the goal.
• Theo keeps the React Native app instead of his preferred SwiftUI rewrite because agents work faster in it.
• T3OS: a heavily customized headless Ubuntu image built for agents, not for him - Bash instead of Zsh (over 30% of Claude's errors on his machines came from Zsh newline handling breaking ripgrep commands), Ubuntu instead of Arch/CachyOS, XFCE-fork GUIs spun up per thread for computer use.
• If devs don't think this way, they stop benefiting from smarter models while unconstrained technical people accelerate past them.

Closing

• Theo admits he was wrong: the knowledge that made devs better at wielding Opus 4.5 and Fable is starting to hold them back.
• A model that can write its own compiler turns out to be useful to people who want to rewrite their favourite games - and things are still accelerating.