Video summary
[한영자막] 이제 무엇을 만들어야 할까요? — Theo Browne, @t3dotgg
Main summary
Key takeaways
Main ideas / technological concepts
“AI psychosis” as a progression of model capabilities
The speaker frames recent AI model improvements as escalating usefulness—moving through:
- Sonnet 3.5
- Opus 4.5
- “Methus” (and mentions Fable)
Capability shifts described
- Sonnet 3.5: became a breakthrough because it could reliably complete multi-step, end-to-end tasks, enough for day-to-day coding (i.e., strong tool-use consistency across a codebase).
- Opus 4.5: improved longer-running tasks while maintaining context—less like “do step 1, then wait for the next prompt,” and more like continuing across steps with greater autonomy.
- Methus: described as an “orchestration” jump—able to
- understand a codebase and itself
- spawn additional models
- break work into sub-tasks
- verify results This reduces the need for elaborate external systems.
Push for “going bigger” is outdated—shift to “go wider”
The speaker argues that simply improving models isn’t enough; developers must adapt their thinking.
Instead of “bigger” (which is harder to define—training from scratch, building an OS, competing with package ecosystems), they propose “wider”:
- Software has breadth (range of things it covers) and depth (features within an area).
- A company doesn’t need to match the full breadth of a giant (e.g., AWS).
- A better approach is to offer deep features within a targeted spectrum, enough for users to start and then extend.
Skeuomorphism / identity-based tooling constraints in developer workflows
The talk critiques how developers overvalue familiar interfaces and traditions, such as:
- treating the terminal as the “ultimate interface” despite limitations
- preference for known tools (e.g., Vim)
- identity tied to:
- specific languages
- version-control habits
- “sunk cost” in code changes
It uses an iOS 7 analogy: even if a UI looks worse, it can be more useful—mirroring how dev tools may need to evolve beyond familiar aesthetics/workflows.
Product / workflow examples and “review-style” takeaways
Personal experiences using models for practical work
The speaker claims:
- Many real engineering tasks (e.g., Jira tickets) could be handled with Opus 4.5.
- Methus adds more unique value for workflows requiring orchestration and verification.
They also emphasize that prompt-only changes can outperform setups requiring heavy custom “software factories.”
“G-brain tier”: entire products as a markdown file + cron
Several projects illustrate how low-effort automation is becoming possible:
- Reddit scraper: a 2–3 day side project to scrape programming humor posts and format them for reuse (copy-paste memes).
- Zoom for streamers (Ping): a Y Combinator-backed startup enabling high-quality collaborations in tools like OBS.
- Full-stack cloud: conceptually compared to Vercel, but extended with built-in database features.
Then they describe a “tier shift”:
- What used to be a startup can now be a side project.
- What used to be complex infrastructure can now be “just a markdown file” that runs by piping into models.
AI agent workflow described as a markdown-driven system
A service is described that:
- triages PRs
- uses AI to review and help prioritize
- updates static HTML
- uploads/sends output to S3
- runs automatically via cron at 9:00 a.m. to generate a daily work plan
Uncertainty about “too big”
The speaker lists open questions about what counts as “too big” anymore, including:
- training models from scratch
- building a custom operating system
- competing directly with ecosystems like NPM/Node
Response: redefine ambition as spectrum width rather than brute-force parity in both depth/breadth with giants.
Main guidance / tutorial-like takeaway (implicit)
If you want more value from AI, don’t only wait for better models—change how your product and systems are structured:
- Use prompt-based orchestration (less custom tooling).
- Architect products so users can extend missing features (build enablers, not complete universes).
- Ship whole workflows that can be executed from simple artifacts (e.g., markdown piped into models) and scheduled automatically.
Main speakers / sources
- Theo Browne (@t3dotgg) — the primary speaker.