Skip to main content

Lab Notebook · First Contact — GPT-3

GPT-3, a blog-post generator, and the moment code generation became the idea

In December 2022 we built a small Python tool using the GPT-3 API to draft blog posts and SEO plans — a content tool, not an engineering one. Using an LLM API directly for the first time surfaced a bigger idea than the tool itself: if it can draft a paragraph, it can draft a function. That side discovery is the first link in the chain that leads to vocabotics.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 15 December 2022
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 15 Dec 2022
Living document. Reviewed 15 Dec 2022
Entry date
15 December 2022
Category
Method
Access
🔓 Public
ResearchPartly proven

We built a tool to draft blog posts. It quietly proved something bigger: if an LLM can draft a paragraph, it can draft a function.

Dec 2022
first hands-on LLM API contact
3
content formats attempted: posts, SEO plans, creative writing
1
side discovery that became the whole arc: code, not content

Honest evaluation

Partly proven

The content-generation tool worked as built; the bigger code-generation insight it's remembered for was a conceptual hunch here, tested only later.

What would prove or disprove it further

What would prove or disprove it further: the code-generation half of the claim is arguably already settled by history — the entire Codex/Copilot-era industry effectively reran this test at scale and confirmed it. Within this portfolio specifically, the falsifiable next step was to point the same GPT-3 API at a small, defined coding task and check whether the output ran — a test this project never ran itself, but one the field answered a few months later anyway.

The evidence — full reasoning behind the verdict

Verdict: partly-proven. The narrow claim — a GPT-3-backed tool can draft readable, on-topic blog posts and SEO plans from a plain-English brief — worked and is proven. The bigger claim this write-up is actually remembered for — that an LLM API is a general instruction-follower and the same mechanism should work for code, not just prose — was never tested inside this project; it stayed a hunch.

The mechanism behind the proven half is straightforward: watching an API turn a natural-language instruction into structured, usable output repeatably, for content, is what the working pipeline demonstrated. The mechanism behind the unproven half is exactly its absence — no code was generated or evaluated here, so "would this work for code?" is recorded honestly as a side observation, not a result.

Every long arc has a first small step. This is ours: a Python script that used the newly available GPT-3 API to write blog posts — and, almost by accident, showed what an LLM API was actually good for.

What it was

A short proof of concept, built across December 2022 and January 2023: a Python tool wired to the GPT-3 API and python-docx, aimed squarely at content generation — blog posts, SEO plans, a bit of creative writing. Nothing more ambitious than that was intended at the time.

What we built

Research

A small pipeline: prompt GPT-3 for a piece of content, format it with python-docx, hand back a document. It worked, in the narrow sense a content tool needs to work — readable, on-topic output from a plain-English brief.

What mattered wasn't the blog posts. It was the mechanics underneath them: watching an API turn a natural-language instruction into structured, usable output, repeatably. That's a generic capability, not a content-specific one — and once you've seen it work for prose, the question "would this work for code?" isn't far behind.

What we learned — including the honest negative

This was a content project, and the honest tally is short: it drafted blog posts and SEO plans reasonably well, and nothing here was engineered to write code. The code-generation insight was a side discovery, not the original goal — read that as a caveat as much as a credit. No production system, no shipped tool, came directly out of this experiment.

What it did produce was a reframing: an LLM API is a general instruction-follower, not a content-specific tool. Once that reframing landed, the natural next question was task decomposition and autonomous execution — attempted for the first time seven months later.

Where it went / status

Archived as a completed proof of concept — the tool itself was never developed further. Its importance is entirely upstream: this is the first hands-on LLM API experience in the record, and the pivot point the rest of the arc traces back to.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • It started as a content-generation tool, not a code tool — the code-generation insight was a side discovery, not the original goal.
  • No code was written by the tool itself; the insight here was conceptual, proven out only in later projects.
  • Nothing from this experiment shipped — it is kept in the record as the origin point, not as a result on its own merits.

Where this connects

Sources

  1. vocabotics project audit — OpenAI GPT-3 Experiments (first hands-on LLM API contact), Dec 2022vocabotics internal project history · as of December 2022

    We use cookies.