How far AI has taken over the world

AI is changing the world—and what you can do.

Will AI kill your idea? Take over your job? Reshape your industry?

Every AI update could change your answer.

PUBLIC WHITEPAPER
v0.2 · October 3, 2026
surviagi.com

1. Change begins with work you know

Building a product means designing it, writing code, answering customers and keeping it running. Providing a service means understanding a need, preparing the work, checking it and delivering a useful result. You put time and skill into each part.

When AI can take on some of that work, your choices change. Something you could only attempt with a team may become possible on your own. A routine task may take less of your day. A service you planned to sell may become something customers can do for themselves.

Whether you run a business, work independently, have a job or simply want to make something with AI, those changes matter. They affect what you can offer, what you need to learn and where your effort is still needed.

SurviAGI is an initiative to make AI’s progress toward independently completing major production and service work more transparent and credible. We actively collect and analyze AI developments, connecting them to the work they affect. Our website carries this record; community contributions can help fill gaps and correct our judgments.

We want users to see what AI has actually achieved, where people are still needed and which possibilities are worth exploring.

2. From a demonstration to everyday use

Imagine watching AI finish in minutes something that usually takes you a day. Before relying on it, you need to know what happened around the demonstration: who prepared the materials, whether anyone repaired the result, and whether it would work again with your own task.

A release announcement may describe a feature without showing all the work required to use it. A benchmark tests particular tasks under particular conditions. A business case can show that a system is being used while leaving its failure rate or human workload unclear. Each offers part of the picture.

Developers, businesses and users each hold part of the picture. Developers know how the system was built. Businesses know how it fits into their operations. Users know whether the result was useful and how much they had to fix.

We bring those pieces together where evidence is available. We want a useful account of what was attempted, under what conditions, with what outcome and with how much human help. A missing part stays an open question.

3. Connect each update to the work it changes

An AI update becomes useful when you can connect it to something you do or want to do: preparing a report, editing a video, repairing a program or handling a customer request.

We are building a record that follows that connection. Users should be able to find an update, see the work it affects, inspect the evidence and follow how the assessment changes over time. Markets and occupations help users find the work relevant to them.

See what AI can take on. Identify the parts it can complete, the conditions it needs and what a person still has to do. This can help a small team plan its work or someone working alone explore a service they could offer.

Understand why an update matters. A new version may make a task possible, improve the result, reduce effort or make repeated use more reliable. The record should explain which of these changed and what supports that conclusion.

Learn from real attempts. A case is useful when its task, conditions and outcome are clear. It gives you a starting point for trying a similar approach and shows what you will need to check yourself.

These records can help users judge both opportunities and pressure on their work. They do not tell us the probability that a business will fail or that a person will lose their job.

4. Completing the job includes the work people still do

By “AI taking over the world,” we mean AI becoming capable of independently completing the major production and service work society relies on. Our scope includes work on computers and work in the physical world through robots and other tools.

Consider a report for a customer. Writing ten fluent pages is one part of the job. Someone also needs to find the right information, check dates and figures, resolve contradictions and deliver a result the customer can use. If a person must redo those steps, their effort counts when we assess what AI completed.

What did AI finish, and what did a person still have to do? Preparation, checking, corrections, retries and handling unexpected situations all belong in that account when they are known. Repeated usable results offer a stronger basis than a single successful attempt.

We collect original announcements, evaluations, product information and accounts of real use, then connect them to specific work. A publisher’s claim and what we accept on the available evidence are recorded separately. Our assessments must show which conclusions come from the source and which come from our analysis.

We also distinguish three questions: what AI can do, whether people use it in practice, and what changes as a result. Production use does not by itself show that AI completes the work independently. A task result does not by itself establish an effect on employment or income.

The broader picture grows from these specific assessments. Work without enough evidence remains unresolved; it is not treated as proof that AI cannot do it.

5. Your experience can help improve AI

When you correct an AI’s answer, you contribute knowledge about the work. You might catch a wrong figure, explain a customer’s requirement or show why an apparently finished result cannot be used. Your correction helps define what a good result requires.

If you retain the task, result, corrections and reasons for accepting or rejecting it, you preserve more than the final answer. These records can help identify recurring mistakes, prepare tests and, when selected and organized appropriately, become material for post-training: further training that adjusts a model’s behavior after its initial training. 1

Users should be able to put their own experience to use. That includes retaining material they have the right to use, choosing a model that supports further training and testing whether it handles their work better. Keeping a conversation is one step in preparing material; improvement still requires suitable data, training and evaluation.

Post-training data, and the opportunity to use it to improve AI, should be accessible beyond model companies. A designer’s corrections, an engineer’s checked repairs or a small business’s failed automation attempts may also help other people prepare evaluations and training material.

This is the data sovereignty we advocate: you decide how experience you have the right to use is retained and used, and whether to contribute it. Private material stays outside the community. We want voluntarily contributed material to become public resources that others can reuse to improve and evaluate their own systems.

6. Our commitments to evidence and sharing

Explain the change in terms of work. Say what AI completed, what changed and what a person still had to do. Include failures and difficulties alongside successful results.

Make the basis of a judgment accessible. Show the original source, what has been checked and what is our interpretation. Keep the conditions and limits that affect the conclusion. When a judgment changes, explain why and preserve the earlier record.

Keep gaps visible. Missing evidence remains missing. We will distinguish a capability change from an expansion of coverage or a revision to our method.

Respect your choice to share. You should be able to keep and use experience you have the right to use without contributing it to the community. Participation is voluntary.

Keep contributed resources public and reusable. We do not collect sensitive material. Before uploading, users must review and select their own content, confirm that it contains no private personal information, client confidential information or other sensitive content, and ensure they have the right to make it public and allow reuse. Uploading means agreeing to public sharing and reuse, including evaluation and model post-training. If you are unsure or do not want the material to be public and reusable, do not upload it.

Support informed choices. Present what the evidence supports without turning uncertainty into fear or a possible improvement into a guaranteed result. Our analysis remains open to correction.

7. Current work and ways to contribute

Local collection, processing and early analysis are underway. We are organizing public AI updates and evidence, linking them to specific work and developing assessments of what AI can complete. The current analysis is machine-generated and has not been individually reviewed by a person. A calibrated public progress baseline and a community submission channel are not yet available.

As the record develops, users should be able to find work similar to their own, inspect original evidence and see why an assessment changed. We also want people familiar with that work to help identify missing conditions, weak evidence and mistaken conclusions.

You can begin with an attempt you know well.

Keep the whole attempt. Record what you needed, the model and tools used, the result, checks, corrections and retries, and whether the final outcome was usable. A failed attempt can reveal an important limit.

Explain what a good result requires. Note a missing check, an overlooked condition or work a person still had to do. This helps others understand what success means in your field.

Organize experience you may want to reuse. Keep examples and acceptance criteria that you have the right to use. They may help you prepare evaluations or training material for your own AI. You decide whether to share them; any future upload must meet the public-reuse conditions above.

For now, keep these records for yourself. We will publish contribution instructions on surviagi.com when the channel opens. The open-resource and post-training proposals in this whitepaper are directions for the initiative, not currently available submission or training services.

We want more people to understand AI’s progress—and what they can do with it.

Notes and version

v0.2 · October 3, 2026. This edition rewrites the opening, chapter titles and body around familiar work, useful evidence and participation. It preserves the initiative’s scope, data-sovereignty position and public-reuse commitments. The first bilingual edition was v0.1, dated September 13, 2026. This editorial revision does not establish a new measurement method or progress result.

A calibrated public progress baseline has not yet been established. This edition gives no global takeover percentage or completion date. Any future aggregate estimate will need to state which work it covers, how different kinds of work are weighted and how missing evidence affects the result. Capability, actual adoption and economic consequences remain distinct.

The following sources inform specific parts of the approach:

1. Human demonstrations and feedback: Training language models to follow instructions with human feedback — Ouyang et al., 2022.
2. Tasks grounded in paid work and checked outcomes: SWE-Lancer — OpenAI, 2025.
3. Task difficulty and reliability: Task-Completion Time Horizons — METR.

The English and Simplified Chinese editions express the same initiative and commitments. Future substantive revisions will identify changes to scope, method or conclusions.