From AI to AGI and ASI

1. 1

   Training

   The system adjusts many internal numbers—weights—to reduce prediction error. The composition of the data and the training objective strongly shape what the model learns, including its biases.

2. 2

   Inference

   A trained model receives new input and computes a likely output. A language model repeatedly predicts the next token from context, but likelihood does not guarantee truth.

3. 3

   Post-training

   Human preferences, safety rules, and tool-use behavior are added to shape responses. This can make behavior more useful, but does not guarantee truthfulness or controllability.

Model

A mathematical structure trained to compute outputs from inputs

Token

A unit of information, often smaller than a word, processed by language models

Parameter

An internal number adjusted during training to encode patterns

Inference

Using a trained model to produce an output for new input

1. 1950

   How can we evaluate machine intelligence?

   An abstract question about whether machines think became a concrete question about distinguishing a machine from a person through conversation.

   People and research teamsAlan Turing

   • 1950 · Computing Machinery and Intelligence — The paper proposed the imitation game, later known as the Turing test, and remains relevant to discussions of how to evaluate model behavior.
2. 1956

   AI becomes a field

   The Dartmouth Summer Research Project set an agenda for implementing learning, reasoning, and language in machines under the name artificial intelligence.

   People and research teamsJohn McCarthy · Marvin Minsky · Nathaniel Rochester · Claude Shannon

   • 1955 proposal → 1956 Dartmouth workshop — A proposal by four researchers led to the summer gathering. It established a research direction rather than announcing a finished technology.
3. 1980s–2000s

   From rules to machine learning

   Alongside expert systems built from hand-written rules, statistical machine learning and neural networks advanced methods for learning classification and prediction from data.

   People and research teamsDavid Rumelhart · Geoffrey Hinton · Ronald Williams · Corinna Cortes · Vladimir Vapnik

   • 1986 · Learning representations by back-propagating errors — Rumelhart, Hinton, and Williams demonstrated how output errors could train internal representations in neural networks.
   • 1995 · Support-vector networks — Cortes and Vapnik presented SVMs for learning classification boundaries, illustrating the development of approaches beyond neural networks.
4. 2012

   The deep-learning leap

   Large datasets, GPU computing, and deep neural networks sharply improved image recognition. Training data and shared evaluation standards mattered alongside the model itself.

   People and research teamsAlex Krizhevsky · Ilya Sutskever · Geoffrey Hinton · Fei-Fei Li of the ImageNet team

   • 2009 · ImageNet dataset — Researchers including Fei-Fei Li and Jia Deng published a large image database that expanded the foundation for recognition research.
   • 2012 · AlexNet and the image-recognition challenge — Krizhevsky, Sutskever, and Hinton's convolutional network achieved a substantial lead over other entries in ILSVRC 2012.
5. 2016

   AlphaGo combines learning and search

   Neural networks estimating promising moves and winning chances, combined with search and reinforcement learning, achieved strong results in the complex game of Go.

   People and research teamsDavid Silver and the DeepMind team · Lee Sedol, the human opponent

   • March 2016 · AlphaGo–Lee Sedol match — AlphaGo won four of five games in Seoul. Exceptional performance on a specific task must be distinguished from human-level general intelligence.
6. 2017

   The Transformer

   The Transformer architecture used attention to model relationships across information. It became a core foundation for large language models and many generative AI systems.

   People and research teamsEight researchers including Ashish Vaswani and Noam Shazeer

   • 2017 · Attention Is All You Need — The paper built a translation model around attention instead of recurrence. Its greater parallelism helped support training at larger scales.
7. 2020s

   Foundation models and generative AI

   Models pretrained on broad data were adapted to many tasks, while systems rapidly expanded across text, images, audio, video, and code.

   People and research teamsTom Brown and the GPT-3 team · Rishi Bommasani, Percy Liang, and Stanford CRFM colleagues

   • 2020 · GPT-3 — The research studied language tasks performed at scale from instructions and a few examples, without additional training for each task.
   • 2021 · Opportunities and risks of foundation models — Stanford CRFM described widely reusable models as foundation models and examined their capabilities alongside social risks.
   • November 2022 · ChatGPT research preview — OpenAI released a conversational research preview to the public, widening access to generative AI through questions and follow-up requests.
8. Today

   Tool-using agents

   Developers connect models to memory, search, code execution, computer control, and iterative planning for longer tasks. Model output is becoming system action.

   People and research teamsShunyu Yao and the ReAct team · Tianbao Xie and the OSWorld team

   • 2022 · ReAct — The approach alternates language-model reasoning with actions using tools or environments, collecting information to revise a plan.
   • 2024 · OSWorld — The benchmark evaluates application use and file tasks in real computer environments, measuring action outcomes and failures beyond plausible answers.

1. 01

   Foundation model

   Learns broad patterns and representations across language, images, and code

2. 02

   Multimodality

   Connects text with seeing, hearing, speaking, and other forms of information

3. 03

   Tools and memory

   Reaches beyond the model through search, code, APIs, and persistent memory

4. 04

   Agency

   Breaks goals into steps, checks results, and repeatedly revises plans

5. 05

   AGI intervention threshold

   Human-level general capability connects to money, organizations, and infrastructure to execute decisions

ASI is not merely a higher benchmark score. It is the stage where AGI's generality, autonomy, and real-world authority expand into superhuman capability and accelerated AI research until people can no longer supervise the system reliably. NO AGI intervenes at the possible beginning of that path: the AGI threshold.

1. 01

   What can this system actually do, and what can it still not do?

2. 02

   Who sets its goal, and how independently can the system reinterpret or change it?

3. 03

   What data, tools, money, and infrastructure can it access?

4. 04

   Does a person approve consequential actions, and can people meaningfully stop it at any time?

5. 05

   Can capability and risk evaluations be reproduced outside the developer?

6. 06

   Are incidents, failures, and limitations disclosed, with a clearly accountable party?