2022: The Year Everything Changed in AI

I started 2022 building with LLMs when it was still a weird hobby you explained at parties while people backed away slowly. I ended 2022 watching relatives use ChatGPT to write wedding speeches. The compression of “fringe” to “default” in eleven months broke my sense of how fast industries are supposed to move.

This is not a timeline of every paper. It is a personal field log: what I built, what failed, what suddenly worked, and what I wish I had not optimistically shipped.

January to March: Plumbing Before the Hype

GPT-3 API access was the toy that ate my Q1. I was not training models. I was wiring retrieval, prompt templates, and evaluation harnesses around models I did not own. Side agent experiments with Telegram ingress and memory prototypes that would take years to become anything serious.

Lessons from the quiet months:

• Prompts are code. Version them or drown in regressions.
• Evals before features. If you cannot score outputs, you cannot ship responsibly.
• Latency kills UX. Streaming completions were not vanity. They were survival.

Nobody paid much attention. That was the advantage.

April to June: Images Enter the Chat

DALL-E 2 and Midjourney made generative media real for normies. I split time between text agents and image pipelines for marketing assets. The startup pitch deck problem (“we need visuals”) got cheaper overnight.

I also hit the first wall of policy whack-a-mole: NSFW filters, copyright anxiety, client requests that made me update terms of service twice.

Internally I drew a map of skills I was accumulating. By midyear it looked like a messy graph, not a ladder.

flowchart TB
    subgraph skills [Skills Layer]
        PROMPT[Prompt Engineering]
        EVAL[Eval Design]
        RET[Retrieval / RAG]
        EMB[Embeddings]
        FINETUNE[Fine-Tuning LoRA]
        DIFF[Latent Diffusion]
    end

    subgraph projects [Projects Layer]
        AGENT[Telegram Agent Experiments]
        MEM[Memory Prototypes]
        IMG[Image Gen Pipelines]
        B2B[B2B SaaS Experiments]
    end

    subgraph infra [Infra Layer]
        PY[Python Services]
        API[OpenAI / HF APIs]
        GPU[Local GPU Experiments]
        TG[Telegram Bot Layer]
    end

    PROMPT --> AGENT
    RET --> MEM
    EMB --> MEM
    EMB --> AGENT
    DIFF --> IMG
    FINETUNE --> IMG
    EVAL --> AGENT
    EVAL --> B2B

    PY --> AGENT
    API --> AGENT
    API --> IMG
    GPU --> DIFF
    GPU --> FINETUNE
    TG --> AGENT

    MEM --> AGENT
    IMG --> B2B
    AGENT --> B2B

The edges are dependencies. The graph got denser every month. That density was the year’s real story: tools stopped being siloed and started composing.

July to August: Stable Diffusion and the Open Source Shock

I wrote about this separately because the week mattered. Stable Diffusion’s release moved image generation from API rent-seeking to checkpoint files on disk. My GPU went from gaming hardware to production-ish infra.

Fine-tuning culture exploded. DreamBooth, LoRA, community WebUIs. I shipped internal tools faster and slept worse.

September to October: Memory, Markets, and India Reality

September was ChromaDB memory work: making agents remember without stuffing prompts. October was market reality: B2B SaaS in India does not care about your US pricing calculator. I watched founders burn runway on GTM that ignored procurement and relationship graphs.

I straddled builder and researcher modes. Neither side felt optional. If you ignore GTM, you build demos. If you ignore memory and evals, you build demos that lie confidently.

November: ChatGPT and the Overton Window

OpenAI shipped ChatGPT on November 30. The date is burned in because everything after felt like damage control for attention.

Effects I felt within two weeks:

• Inbound interest spiked from people who had ignored LLMs for months
• Expectations detached from capability overnight
• Hiring conversations changed from “what is GPT” to “why are you not ChatGPT yet”
• My old demos looked quaint even when they were more specialized

ChatGPT was not magically more capable than what tinkerers had been chaining together. It was accessible, conversational, and free to try. Distribution beat depth for public narrative.

I rewrote onboarding for everything I showed investors. Screenshots aged in days.

December: Consolidation and Honest Accounting

By year end I had:

• A Telegram agent stack with real memory, still imperfect but debuggable

• Image pipelines I trusted for internal use, not legal-approved for all clients

• Strong opinions on Indian B2B GTM I wished I had learned cheaper Failures:

• Shipped features without eval coverage and paid in support time

• Underestimated how fast open models commoditize UI wrappers

• Overestimated how fast enterprises adopt anything without security theater answers

Themes That Survived the Hype Cycles

Composition won. Retrieval plus tools plus memory plus UI beat raw model size for most founder use cases.

Open weights changed bargaining power. APIs are convenient until your unit economics depend on them.

India is not a discount US market. It is a parallel GTM universe.

Memory is product, not research. Users do not care about your vector DB. They care that the bot remembers their kid’s name.

Speed of narrative outran speed of reliability. 2022 rewarded demos. 2023 would punish them. I could feel that coming in December.

What I Would Tell January-2022 Me

1. Buy more GPU before the scalpers wake up.
2. Log prompts and outputs from day one.
3. Say no to custom demos without contracts.
4. Study sales cycles in your actual market, not on podcasts.
5. ChatGPT is coming. Build specialized value, not a chat box.

Looking at 2023 From the Edge

The year everything changed in AI was also the year I stopped asking whether LLMs were a fad. The fad was thinking you could ignore them and keep a normal software career.

I entered 2023 tired, more technical than I was in January, and more skeptical of people who confuse virality with product-market fit. The tools got better. The hard parts stayed hard: memory, evals, distribution, trust.

2022 did not give us AGI. It gave us permission to build like the future was already here, with all the mess that implies.

I would not want to relive the chaos. I would not want to have missed it.

Happy new year. Ship evals first.