mlllmstartupsengineering

Building with the GPT-3 API Before Anyone Had Heard of ChatGPT

5 min readUtso Sarkar

In March 2022 I was wiring GPT-3 into side experiments from my room — Telegram ingress, rough memory sketches, prompt templates that broke every time OpenAI tweaked a model string. Nobody in my extended family had heard of ChatGPT. That product would not exist until November.

This post is a field note from the quiet months: before Stable Diffusion, before the hype tsunami, when using the API felt like a hobby you apologized for at dinner.

The Landscape in Early 2022

If you were paying attention to ML Twitter in Q1 2022, the text side looked like this:

What did not exist in my world: ChatGPT, cheap open-weight chat models, LangChain as default plumbing, or a single Indian enterprise buyer asking for “an LLM strategy.”

I was still running Strato Inc apps and tinkering at night. The LLM work was not a startup pitch. It was me trying to see if a Telegram bot could remember context better than stuffing the last twenty messages into a prompt.

Why Telegram

Founders live on their phones. Async messaging maps cleanly to agent loops: message in, classify intent, maybe retrieve something, reply.

I picked Telegram because the Bot API is boring in a good way. No app store review for v1. Easy to share a bot link with three friends who would tolerate broken behavior.

The architecture in March was embarrassingly simple:

flowchart LR
    TG[Telegram message] --> Router[Intent / keyword router]
    Router --> Prompt[Prompt template + few-shot]
    Prompt --> API[OpenAI Completions API]
    API --> Post[Light post-process]
    Post --> Reply[Telegram reply]

No vector DB yet. No tools. No swarm. Just prompt + API + hope, which is where everyone starts whether they admit it or not.

What GPT-3 Taught Me the Hard Way

Prompts are code. I versioned them in git after the third time a “small wording tweak” turned a helpful bot into a passive-aggressive poet. If you are not diffing prompts, you are not engineering.

Base models simulate text. Ask for a JSON object, sometimes get JSON, sometimes get a Medium article about JSON. Instruct-style behavior was coming; the public API surface in early 2022 still rewarded people who understood that GPT-3 was a distribution learner, not an assistant.

Latency is UX. Non-streaming completions felt broken on mobile. Waiting eight seconds for “okay” is not a product. Streaming responses were not vanity — they were survival.

Cost is real. I tracked tokens before I tracked uptime. A badly designed loop that re-sent the system prompt every turn would eat lunch money in an afternoon.

Evals or chaos. I had no formal harness yet. I had a spreadsheet of prompts and “good / bad / hilarious failure” notes. That primitive eval was enough to stop me shipping the worst regressions to friends.

What I Was Not Doing

I was not fine-tuning with LoRA. That post comes later in 2022, after Stable Diffusion made open-weight adaptation culture normal.

I was not pitching enterprises on “AI transformation.” Indian B2B buyers were still asking for basic dashboards and WhatsApp integrations, not retrieval-augmented generation.

March 2022 me was plumbing, not thought leadership.

The InstructGPT Paper Mattered More Than the Hype

The InstructGPT result landed in January 2022 and changed how I thought about the stack even before products caught up:

  1. Alignment is a training recipe, not a side effect of scale
  2. Human preference data is a moat if you can collect it responsibly
  3. Smaller aligned models beating larger base models would reshape API economics

I did not have RLHF infrastructure. Almost nobody did outside labs. But I stopped assuming the biggest model string on the pricing page was automatically the right tool for a constrained bot.

That mental shift paid off when ChatGPT arrived nine months later and everyone conflated “chat UI” with “new science.”

Failure Modes I Hit in March

Each failure pushed me toward the memory and eval work I would write about later in 2022. The through-line is boring: continuity and measurement, not bigger models.

What Changed by Year End (Foreshadowing)

November 2022 would compress a decade of narrative into a weekend when ChatGPT launched. March me would have laughed if you said relatives would use LLMs for wedding speeches.

But the foundations were already visible in March if you were building instead of tweeting:

Stable Diffusion in August would add image pipelines. ChromaDB experiments in September would add memory. None of that was visible yet. The GPT-3 API was enough to learn on.

Takeaway

If you are building in 2022 (or reading this retroactively): you do not need the trendiest acronym to start. You need a narrow loop, a logging habit, and friends who will tell you when the bot lies.

ChatGPT would later make everyone think they were late. They were not. They were early if they had already learned that prompts are code and demos without evals are liabilities.

I was just a kid with Strato Inc on the side and a Telegram bot that sometimes worked. That was the right level of ambition for March 2022.

The enterprise pivots could wait. The API was already enough to teach the real lessons.

--claps