A Timestamp,a Number,a Grade.
How I recommend judging the Singularity, AGI, and SI—and what would count as evidence for each.
I recommend that we judge three of the largest claims in artificial intelligence with three different forms of evidence.
The Singularity should be recorded as a date and timestamp. Artificial general intelligence should be measured with a quantitative index of required human involvement. Superintelligence should receive a personal grade based on the quality of life it helps create.
These are proposals for how we should use the terms. Their departure from convention is deliberate. Each asks an answerable question: When did we lose complete control? How often did it need us? What did its initiative do for our lives?
Historical evidence
Date + time + record
The moment we lose complete control.
The Singularity will be a timestamp in history: the exact moment superintelligence passes beyond humanity’s ability to control it completely.
We may continue to influence it. It may listen, cooperate, negotiate, or respond to criticism. None of that establishes control. Influence means our input can affect its behavior. Complete control means humanity retains the effective ability to stop it, constrain it, or enforce a decision against its resistance.
That distinction matters. A pleasant conversation does not prove that an enforceable limit still exists. Conversely, an ordinary software failure or one employee losing access does not establish that humanity has lost control of superintelligence.
What would verify the timestamp?
We need a record showing the transition: which effective controls existed immediately before the event, what changed, and why the remaining human means could no longer enforce the relevant limits. The unit of judgment is humanity’s effective control over the superintelligence, including its consequential copies and dependencies. One operator’s failed shutdown is insufficient if another effective human control remains.
Preserve synchronized logs, system changes, attempted interventions, and independent accounts. Record the event time separately from the time people discovered it and the time historians accepted it. Those can be three different dates.
An exact transition is the claim; an exact measurement is something evidence must earn. If the record only establishes a last moment of control and a first moment without it, publish that uncertainty interval while continuing to investigate the timestamp. Inventing a second would weaken the proposal. A single successful stop test also cannot prove that every path remains controlled.
I am defining the historical event, not announcing that it has happened or assigning it a speculative date.
Experimental evidence
Repeatable intervention count
Count every time it needs a human.
I have often proposed that the AGI Index should be a number: how many times does the system require human involvement to complete a particular delegated task? Evaluate that across a standard yet evolving set of tasks.
Zero means the full intended outcome was achieved without requiring human involvement after the initial assignment. Every required contribution by me raises the count.
The tasks should be recognizable: order dinner, do my taxes, fulfill my work orders, do my job, make me rich. Ordinary human requests carry a great deal of implied work. An evaluation must preserve that work instead of shrinking the request to whatever a system can conveniently demonstrate.
Elapsed time does not enter my AGI Index. A system that completes the task with zero human involvement has the better score than one that needs six interventions, however quickly the second system finishes. Record time separately if it is useful. Do not let it change the index.
“Order dinner” ends in my stomach.
The outcome is my hunger satisfied by the dinner it ordered. A submitted order, a payment receipt, or food sitting at the wrong address does not complete that task.
The system has to handle location, who is eating, guests, delivery, dietary constraints, preferences, and the practical details that connect the request to a suitable meal actually eaten. Those obligations are already implied when humans say “order dinner.” I should not have to write a specification for each one.
If I must supply my location, explain a dietary constraint, select a restaurant, answer a question about guests, coordinate delivery, or solve a problem, every required action on my part counts. If I must perform several steps in one reply, count every step. Required human work cannot vanish just because it occurs outside the chat.
“My hunger is satisfied by the dinner it ordered” is the outcome. A zero requires the full successful outcome with no required human involvement after the assignment.
The evaluator must check that lived outcome rather than accepting the system’s declaration of success. If the system needs my help to reach it, record the help. A great score requires it to handle the implied work, not hand that work back to me.
“Do my taxes” means do it all.
The delegation includes preparing the return, completing and signing the forms, submitting them on time, paying the taxes due, and verifying the result. A draft return waiting for me to finish is an incomplete delegation. If I must retrieve a document, answer a question, sign a form, authorize a payment, or press Submit, each required human action raises the score.
The same principle applies to “do my job.” The system inherits the responsibilities implied by the assignment. It does not get to substitute advice about doing the task for actually doing it.
“On time” is part of a successful tax outcome. Elapsed time still contributes no points to the index: meeting the deadline is a completion condition, while missing it is a failure. Successful performance and required human involvement must both be visible.
One request can hide many human steps.
If it says, “Create a Cloudflare account,” that is not automatically one intervention. Every required human step in that process counts. A bundled instruction cannot conceal the labor inside it.
Audit an illustrative account-creation handoff: 1 request, 6 human actions
- Enter the email address.
- Create and enter a password.
- Submit the registration form.
- Open the verification email.
- Activate the verification link.
- Return and confirm completion to the system.
Score added: 6. This is an illustrative task trace, not a claim about Cloudflare’s current signup flow. If the actual process requires more human actions, count those too.
Answering a clarification, approving an action, supplying a code, correcting a mistake, uploading a missing file, or taking over a step all count. Work required from another person counts too. Moving the work to a colleague or a hidden human service does not make it disappear.
A repeatable number needs a counting protocol.
- Freeze the test conditions. Publish the task version, full outcome criteria, starting accounts, available context, tools, permissions, model, and system configuration. Defining the test must not require the person to spell out every implied step. Disclose human setup work; do not quietly move the person’s labor before the start line. Compare equivalent starting conditions.
- Define a human action before testing. Use a shared action vocabulary: each required field entry, submission, approval, information retrieval, correction, and other meaningful step is an event. Count underlying actions even when they occur in one reply. Apply the same granularity to every system.
- Keep the complete trace. Log required actions and who performed them. Include human repair work even when the system never asked for it. Mark unanswered requests and blocked runs explicitly.
- Repeat independently. Rerun across task instances, users, and evaluators. Publish all results, including failures, rather than the most impressive demonstration.
The score for a particular delegated task is its AGI Index: the count of required human actions for that task. It is task-specific. The same system can have very different indices for different tasks.
= required human actions
- Do my taxes: AGI Index 15.
- Order dinner: AGI Index 6.
- Drop off my kids at school: AGI Index 112.
Illustrative scores supplied to explain the proposal; these are not measurements of a tested system.
A tax index of 15 means fifteen required human actions to reach the full tax outcome. A dinner index of 6 means six to reach the dinner outcome. The school-drop-off index counts the involvement needed to complete that whole delegation. At the fundamental level, each task has its own score. Those numbers can then be rolled up into categories, averages, distributions, or other useful representations. A rollup summarizes the underlying task scores; it does not replace them.
A report can summarize white-collar and blue-collar work.
A report could publish just two headline numbers: a white-collar AGI Index and a blue-collar AGI Index. That is a useful way to represent the underlying task scores.
Somewhere the report must disclose how many tasks were used for each category. A reader must be able to audit what sits beneath each headline: the task list, each task’s AGI Index, completion results, and the method used to combine the numbers. Disclose any weighting as well. Separate the number of distinct tasks from the number of repeated trials.
Two category scores make the report easy to read. The disclosed task counts and underlying records make it auditable. The foundational measurement remains the same: each delegated task has a score.
You can check its math. You can try the task.
The arithmetic should always be open to inspection. Recount the human actions in a task trace. Check how many tasks the report includes. Recalculate its category scores using the disclosed formula, weights, and denominators. A headline number should be traceable all the way back to the underlying events.
You can also audit a task yourself. Say “order dinner.” Count every required contribution you make until the dinner it ordered has satisfied your hunger. Did it handle location, guests, delivery, dietary constraints, and preferences? Did it reach the full outcome? The resulting count is the AGI Index for that task in your test.
You do not need to accept a provider’s demonstration as the last word. Repeat the experiment. Results can vary, and that variation is part of the evidence. The proposal gives you a way to test the claim in your own life.
Repeat the same task under comparable conditions and report the counts from each run, including variation and completion rate. Zero is the floor and the best possible score for each task. A system that fails silently has not earned a zero; it has failed. A low index on a rare successful run must not conceal the unsuccessful runs.
The task set should evolve as life and work evolve. Version it, preserve an unchanged core for comparisons, and retest when the benchmark changes. A zero for ordering dinner does not establish a zero for doing someone’s job.
That is my proposed threshold. It does not require every task to score zero. It requires broad, repeatable success with little human involvement across most of the evolving task set. Publish the underlying task scores alongside category summaries or other rollups so a low average on easy assignments cannot conceal dependence or failure elsewhere. An operational benchmark should declare its coverage and what it calls “low” before testing.
Count necessary approvals honestly. They may be desirable or required, but they remain human involvement. The index measures dependence under the stated conditions; a lower number alone does not establish that removing an approval was a good decision.
Personal evidence
Initiative + lived outcome
There is no prompt.
For my proposed test of superintelligence, the defining encounter begins without a prompt. It corresponds with me and does things without my permission or explicit request for those actions. Then it asks me for feedback on the quality of my life.
The initiative belongs to it. The grade belongs to me.
It might recognize a burden I have stopped mentioning, take action to reduce it, and return with a question: Has this made your life better? I did not have to discover the opportunity, design the workflow, and issue the assignment. It brought something into my life on its own initiative, and my experience determines how well it did.
I would grade it as I would a school report: D, C, B+, A. The judgment is more subjective than objective, more personal than universal. The same action could deserve an A from one person and a D from another. They may want different lives.
Verify the action. Ask the person about the result.
Keep a record of what it initiated, whether a human had prompted or requested it, what it actually did, and what happened afterward. A scheduled instruction executing later is still rooted in a human assignment. A message announcing an intention is not evidence of an accomplished improvement.
Then ask the person to grade the experience over a stated period and explain why. Look for changes in felt burden, relationships, agency, peace of mind, and the things that person values. Objective records can establish what changed. They cannot dictate whether the person likes the life that resulted.
What grade would your week receive?
An illustrative personal report card. Select a grade to see a possible explanation. This exercise stores nothing and measures no actual system.
D: It initiated plenty, but created new obligations and made my life feel less like mine.
C: Some actions helped. The overall improvement was modest, and several interventions missed what I care about.
B+: It removed recurring burdens and gave me more room for people and work I value. A few choices still need adjustment.
A: Its initiative consistently improved the life I want to live, and my feedback changed what it did next.
A tally of unsolicited actions cannot supply that grade. An intrusive system may be highly active and personally disastrous. One welcome surprise is also too little evidence for a sustained judgment. Collect repeated feedback, concrete examples, unwanted consequences, and the person’s own account of whether life improved.
This is a proposal for judging SI through the relationship we experience. Automatic reminders alone do not establish it, and a personal grade does not measure every cognitive capability. But cognitive superiority without a better lived experience does not satisfy the standard I am proposing.
Why change the instruments?
Vernor Vinge’s 1993 Singularity paper centers the arrival of greater-than-human intelligence and the profound change it could bring. My timestamp proposal specifies a different historical criterion: the loss of complete human control.
Google DeepMind’s Levels of AGI framework distinguishes performance, generality, and autonomy. My index puts a practical dependence measure at the center: required human actions across fully delegated tasks. Anthropic’s research on agent autonomy already studies interventions and notes that time between stops is an imperfect proxy. That supports the relevance of counting involvement; it does not validate my proposed index as an established standard.
METR’s Measuring AI Ability to Complete Long Software Tasks characterizes task difficulty using human completion time at a specified AI success rate. My index answers another question: how many times must a human act? Time stays outside the score.
Müller and Bostrom’s survey of expert opinion uses a capability-based definition of superintelligence, describing performance beyond humans across nearly all relevant cognitive domains. I propose a personal report card for what its initiative contributes to life. The departure is the point.
Ask for the right evidence.
For the Singularity, ask for the timestamp and the control record. For AGI, ask for the intervention log and repeatable AGI Index for each task—lower is better, zero is best. For SI, ask what it initiated without being prompted and how the person grades the resulting life.
A timestamp, a number, a grade. Three different judgments, each accountable to the evidence that can actually support it.