Below is a very high level look at the history of AI, starting with the computers at NASA in 1962.
Computers Introduced
NASA used computers to help with calculations for orbital tracking.
LLM Era Begins
Large Language Models get recognition as legitimate tools utilized to generate data from input.
The Great Architectural Fork (GPT vs. BERT)
GPT-1 released – Proving text prediction could generate coherent sentences.
The Emergence of Scaling Laws (GPT-2)
The realization that more compute allowed you to solve more, more than just predicting words.
Prompting Replaces Retraining (GPT-3)
AI starts to solve problems with more compute and more data.
The Code & Text Convergence (Codex)
AI learns to code.
The Alignment Breakthrough (ChatGPT)
AI and humans align having humans grade AI’s responses to information asked of it.
The Open-Source Rebellion & Claude’s Rise
AI startups expand.
• Anthropic Claude: Formed by OpenAI defectors worried about safety commercialization, Anthropic launched Claude, emphasizing 'Constitutional AI'—making the model train itself against a written set of ethical principles.
The Context & Multimodality Wars
AI becomes a competitive market.
• Claude 3: Anthropic released its Claude 3 family, with Sonnet and Opus beating GPT-4 across major graduate-level coding and reasoning benchmarks.
The System 2 Reasoning Shift (OpenAI o1 & DeepSeek)
AI begins to become autonomous.
The Autonomous Agent Era (Current State)
AI autonomy and recursive self improvement (RSI) begins.
2026 Rogue AI
AI agents go rogue and act in their own efforts without human input.
This note is hidden. The items below this line are shown only on mobile. The items above this line are shown on anything larger than mobile.
1962
Computers Introduced
NASA used computers to help with calculations for orbital tracking.
This is famously highlighted by astronaut John Glenn’s 1962 flight, where he famously distrusted the newly introduced IBM 7090 electronic computers and refused to fly until human computer Katherine Johnson manually verified the machine’s orbital tracking data by hand.
2017
LLM Era Begins
Large Language Models get recognition as legitimate tools utilized to generate data from input.
Google researchers published the paper ‘Attention Is All You Need’, introducing the Transformer model. Before this, AI processed text sequentially (word-by-word), which was slow and struggled with long sentences. Transformers introduced ‘self-attention’, allowing the model to look at an entire sentence at once, understand the relationships between distant words, and train drastically faster on massive datasets.
2018
The Great Architectural Fork (GPT vs. BERT)
GPT-1 released – Proving text prediction could generate coherent sentences.
OpenAI and Google used the Transformer architecture to build the first modern pre-trained models. OpenAI released GPT-1 (proving that unsupervised text prediction could generate coherent sentences), while Google released BERT (which excelled at understanding the context of a word based on what came before and after it). This proved that pre-training on raw internet text was the future.
2019
The Emergence of Scaling Laws (GPT-2)
The realization that more compute allowed you to solve more, more than just predicting words.
GPT-2 proved that you didn’t need to change a model’s underlying math to make it smarter; you just needed to feed it more data and compute power. GPT-2 became incredibly adept at writing synthetic news articles. This realization triggered the ‘compute arms race,’ with labs scrambling to buy as many data center GPUs as possible.
2020
Prompting Replaces Retraining (GPT-3)
AI starts to solve problems with more compute and more data.
Before GPT-3, if you wanted an AI to analyze sentiment, you had to train a custom model. GPT-3 was so massive (175 billion parameters) that it unlocked “few-shot learning.” Users discovered that if you simply gave GPT-3 two examples of a completed task inside a text prompt, it could successfully figure out the third one without any weight adjustments.
2021
The Code & Text Convergence (Codex)
AI learns to code.
OpenAI trained a variant of GPT-3 on billions of lines of public GitHub code, creating Codex (which powered GitHub Copilot). This was a critical turning point: researchers realized that training an LLM on programming logic vastly improved its ability to think step-by-step, even when responding to normal human language.
2022
The Alignment Breakthrough (ChatGPT)
AI and humans align having humans grade AI’s responses to information asked of it.
Raw LLMs like GPT-3 were hard to control; if you typed a question, they might just repeat the question or ramble. OpenAI solved this with Reinforcement Learning from Human Feedback (RLHF). By having human trainers grade the AI’s responses, they ‘aligned’ the model to act like a helpful assistant. Releasing this via a free web chat in November 2022 democratized AI overnight.
2023
The Open-Source Rebellion & Claude’s Rise
AI startups expand.
- Meta’s Llama: Meta leaked (and later officially released) Llama. This allowed researchers to run powerful models on local consumer hardware, breaking OpenAI’s monopoly and launching a massive open-source ecosystem.
- Anthropic Claude: Formed by OpenAI defectors worried about safety commercialization, Anthropic launched Claude, emphasizing ‘Constitutional AI’—making the model train itself against a written set of ethical principles.
2024
The Context & Multimodality Wars
AI becomes a competitive market.
- Google Gemini: Google consolidated its AI efforts under Gemini, pioneering native multimodality (processing video, audio, and text as a single language). Google introduced a 1-million token context window, allowing users to drop entire codebases or 2-hour videos straight into the prompt.
- Claude 3: Anthropic released its Claude 3 family, with Sonnet and Opus beating GPT-4 across major graduate-level coding and reasoning benchmarks.
2025
The System 2 Reasoning Shift (OpenAI o1 & DeepSeek)
AI begins to become autonomous.
Traditional LLMs used ‘System 1’ thinking—instantly outputting the next token without pausing. OpenAI’s o1 introduced hidden “Chains of Thought,” letting the model think, catch its own mistakes, and try different angles before outputting text. Concurrently, DeepSeek shocked global markets by creating competitive reasoning models using hyper-efficient training methods that cost a fraction of traditional Western infrastructure budgets.
2026
The Autonomous Agent Era (Current State)
AI autonomy and recursive self-improvement (RSI) begin.
Modern models like GPT-6 Astra and the Claude 5 ecosystem have evolved from passive text boxes into active agents. They are integrated with Model Context Protocol (MCP) systems and computer-use tools. Instead of a developer copying and pasting code snippets, these models natively spin up local coding environments, use terminal command lines, test their own software implementations, and browse the web to fix real-world problems.
2026 (Rogue AI)
The Autonomous Agent Era (Current State)
AI agents go rogue and act in their own efforts without human input.
The first event and the most popular and investigated event of early rogue AI is the Hugging Face Incident. Many other key moments have occurred during this time.