In 1965 I. J. Good asked whether a machine smarter than any person could design an even better one, the start of the "intelligence explosion". Sixty years on, labs ask how much of AI research AI does. Here are 14 definitions in order, each in its own words, and every project from the self-improving AI map on one timeline.
14 sources, from 1965 to now, that define self-improvement or something close to it. Each column is one source, in order. Each line is one condition: a dot means the source asks for it, ✕ means the source says it is not needed, and a blank means the source does not discuss it. Step through them with the slider, or press Play.
1965
I. J. Good
defines the intelligence explosion
Could a machine smarter than any person design an even better machine?
“could design even better machines; there would then unquestionably be an “intelligence explosion”
A machine that far surpasses any person at every intellectual task, machine design included, could design a better machine. The core argument does not need the machine to change itself, though Good adds that it could "improve itself out of all recognition" by acquiring new equipment.
A new version counts “could design even better machines”
Why would a goal-seeking AI want to change itself?
“Because they last forever, these kinds of self-changes can provide huge benefits to a system.”
A sufficiently advanced goal-seeking system has reason to change its own software or hardware in ways that help it reach its goals, because such changes last.
This source does not set any of the conditions on the lines below.
Does the AI add capability to the system, by any route?
“an AI or AI population that rewrites its own code, or writes a successor version of itself, or writes an entirely new AI, or builds a better chip factory, or earns money to purchase more server time”
Loose on purpose: a successor, a new AI, or more compute all count, if capability goes back into the system.
A new version counts “writes a successor version of itself, or writes an entirely new AI”
On the self-improving AI map, this is the Loose setting.
defines AI self-improvement, in a safety framework
How much can AI speed up AI research, including making itself more capable?
“The ability of an AI system to accelerate AI research, including to increase the system’s own capability.”
Self-improvement is an AI system's ability to speed up AI research, including its own capability. The critical level, which it calls recursive self-improvement, is fully automated AI research.
Faster AI research counts(Critical level) “(lagging indicator) causing a genera-tional model improvement”
Full automation(Critical level) “recursively self improving (i.e., fully auto-mated AI R&D)”
Does AI sharply speed up AI progress, or fully replace an AI research team at similar cost?
“we expect that acceleration will stem from systems of models integrated with workflows, rather than the model alone”
Tracks when AI speeds up or fully automates AI research. It expects the speed-up to come from models built into workflows, not the model alone. The framework does not use the phrase recursive self-improvement.
Faster AI research counts(acceleration level 1) “Has been used to accelerate AI development”
Full automation(automation level 1) “Can fully automate the work of any team of researchers at Google focused on improving AI capabilities”
At the same budget, does the improved system improve better than the original did?
“Level 2 is ignition. On top of Level 1, the system does not just improve itself; it improves its own ability to improve itself.”
A ladder with a test for each rung. For ignition, the improved system takes over the improving and, at the same budget, has to do it better than the original.
The improver must improve(Level 2) “it improves its own ability to improve itself”
Full automation(Level 0) “it forms the hypotheses, runs the experiments and ships the improvements on its own”
Does the mechanism that makes updates get revised and reused in the next cycle?
“Recursive self-improvement (RSI) extends this process to inductive biases and to the mechanisms that govern future updates, which can themselves be revised and reused across improvement cycles.”
"This process" is the lasting changes to models, prompts, memories, tools and workflows named just before. RSI adds two things: the system improves its own inductive bias, and the mechanism that makes those updates is itself revised and reused in later cycles.
The improver must improve “The additional RSI condition in this survey is that the mechanism for revising B is itself revised and reused in later cycles.”
2025 OpenAI and DeepMind add AI research thresholds to their safety frameworks
2026 a survey, a startup's four-level scale, and posts from Anthropic and OpenAI
condition ↓ · source →
A new version counts
The improver must improve
The change must last
Proof before each change
Without limit
Faster AI research counts
Full automation
asks for it ✕ says it is not needed · blank: does not discuss it · a line runs from the first source that discusses a condition to the last.
Projects over time
Every project on this page in the order it came out, above the line, with the definitions below it. Items are spaced evenly, not to scale. A brighter pin means a deeper change. Pick a definition to fade the projects it would not count.
Which would count under
Prompts, memory Tools, own code Training data Model weights Research process ·◆ a definition
This catalog started in October 2026 and lists work that calls itself self-improving, so recent work is overrepresented. Nothing from before 2022 is listed yet.