The Rise of LLMs

Google publishes "Attention Is All You Need" — the architecture behind every modern LLM. Nobody outside AI research notices.

OpenAI shows that scaling up transformers with more data and parameters makes them surprisingly good at generating text. GPT-2 was considered "too dangerous to release."

175 billion parameters. For the first time, a language model could write essays, code, and translate — without being specifically trained for each task. The era of "general purpose" began.

OpenAI wraps GPT-3.5 in a simple chat interface. 100 million users in two months. Suddenly everyone — from students to CEOs — is talking to an AI. The "ChatGPT moment" from your podcast.

GPT-4 raises the bar. Google launches Gemini. Anthropic releases Claude. Meta open-sources Llama. China produces DeepSeek and Qwen. In one year, we go from one player to an entire industry.

Instead of one model fits all, providers release tiers: powerful models for complex work, fast models for quick tasks, tiny models you can run on your phone. Thinking vs Doing vs Checking.

OpenAI's o1 and o3, DeepSeek R1 — models that "think step by step" before answering. They take longer but get hard problems right. A new tier: models that reason, not just predict.

The landscape moves fast. At NYMG, Anthropic Claude is primary (with different sizes for different agents), OpenAI Codex/GPT-5.4 is the fallback, and gpt-5.4 is used in the LangGraph translation pipeline.