A memory of how AI got here—and how working with it keeps changing. The moments that opened new possibilities, and the original sources that keep the story grounded.
A conversation. A creative tool. A collaborator in the work itself. These eight turning points make the progress easier to see, with ideas for what it could mean in your own work.
Eight turning points · 869 entries in the full archive · Source-reviewed through October 8, 2026
2022
November 30, 2022
AI becomes a conversation.
ChatGPT
ChatGPT gave people a familiar place to begin: a question in their own words, followed by another. An unfinished thought could become something to explore, explain, or develop.
The moment and its sources+
OpenAI introduced ChatGPT as a research preview that could respond in conversation and answer follow-up questions.
Public research preview.
The launch also documented plausible but incorrect answers. Conversational fluency did not establish factual reliability.
The question grew from what one person could do to how a whole organization could use AI. Shared access, information handling, and team practices became part of the picture.
The moment and its sources+
ChatGPT Enterprise launched with organizational administration, single sign-on, data analysis and a commitment not to train on business data or conversations.
Enterprise offering launched; onboarding was beginning.
These were product-specific terms and controls. Connecting company applications was described as future work in this launch post.
Longer documents and more kinds of material could be considered together. You could look for connections across a body of work instead of asking about one fragment at a time.
The moment and its sources+
Google introduced Gemini 1.5 Pro for early testing, including an experimental million-token context window in private preview. Its examples covered long documents, video, audio and code.
Early testing; million-token context in limited private preview.
The million-token feature was limited to selected developers and enterprise customers at this date. Input capacity alone does not establish accuracy.
Source material could become a conversation to listen to. It opened another way into a subject when reading a long document was not the easiest place to start.
The moment and its sources+
Google introduced NotebookLM Audio Overviews: generated discussions between two AI voices, based on the sources supplied to a notebook.
Experimental feature introduced in NotebookLM.
At launch, the audio was English-only and could introduce inaccuracies. A generated discussion reflected the supplied sources; it was not a complete or objective view of the topic.
AI began moving through software interfaces, clicking and typing to work through a task. The possibility expanded from advice about the work to help carrying out parts of it.
The moment and its sources+
Anthropic released computer use in public beta. Developers could direct Claude to view a screen, move a cursor, click buttons and type.
Public beta through developer APIs.
Anthropic described the capability as experimental, cumbersome and error-prone, and recommended beginning with low-risk tasks.
A defined question could lead to a report assembled from multiple sources. Finding material, comparing it, and tracing the answer back to evidence became part of the same task.
The moment and its sources+
OpenAI launched deep research in ChatGPT for multi-step web research. It could search, analyze online material and produce reports with citations.
Initially a limited Pro release; access expanded afterward.
The source warned of fabricated facts, incorrect inferences and weak uncertainty signals. Later connected-app features should not be attributed to this February launch.
You could describe an image, see an interpretation, and refine it through conversation. An idea no longer had to stay in your head while you worked out how to explain it.
The moment and its sources+
OpenAI introduced image generation in GPT-4o, including conversational revision and the use of uploaded images as context.
Rollout began in ChatGPT; the announcement described access by plan.
The launch documented limitations, including cropping, dense information, and precise editing. A generated image is not evidence that a real event occurred.
Documents, applications, and multi-step tasks became more connected. Getting an answer could become preparing a piece of work, with access and human review designed around the task.
The moment and its sources+
OpenAI described GPT-5.6 working with connected documents and workplace applications to produce editable work. In August, Okta made Agent SSO generally available for agents that support Cross App Access, with identity-based connection controls.
GPT-5.6 general-availability rollout began July 9; API multi-agent was initially beta. Okta Agent SSO reached GA August 24 for supported agents.
Connected-work features depend on product, plan and authorized access. Okta Agent SSO applies to supported Cross App Access agents; its broader governance product is a separate offering.
Follow the recent developments, see how they connect, and explore what they could mean for your work. Every entry links to its source.
The full memory.
All 869 entries are here. Search a subject, choose a year, or follow a thread through the original sources. This is a selected record of AI developments, with its source review date shown below.
869 selected source-linked entries.Source-reviewed through October 8, 2026.
