Nobody can honestly hand you that number, so this page doesn’t. It breaks the scenario into five things that must all be true, shows you what this site’s graded evidence actually says about each, and lets you set the probabilities yourself. The multiplication is done in front of you.
Every slider starts at 50% — that is a blank default, not this site’s estimate. This site publishes no probability for any of these.
Four of these five links are live graded predictions. When one flips, the Thesis Tracker score moves and your number here changes with it. That is the email — not a newsletter about AI in general.
Email me when the score moves →The five links are treated as a chain of conditional probabilities and multiplied: P(Matrix) = P(AGI) × P(explosion | AGI) × P(physical control | explosion) × P(we lose control | that) × P(it farms us | that). Multiplying conditionals is the standard way to price a scenario that needs every step; it is also why long chains produce small numbers even when each individual step feels likely.
The honest weakness, stated plainly: these five are not fully independent, and a chain treats them as if the conditioning is already baked into your inputs. If you think the same underlying cause drives several links, your true joint number is higher than the product. The tool cannot fix that for you — but pretending the issue doesn’t exist would be worse than naming it.
The verdicts shown against links 1–4 come from /data.json (CC BY 4.0), the same eight graded predictions behind the AGI-2027 Thesis Tracker, each with a pre-registered flip condition and a primary source. They are evidence about the preconditions. They are not a probability of the scenario, and this page never converts them into one.
Verbatim positions from the dataset — the spread here is the honest state of link 1.
| Elon Musk | AGI by end of 2026 |
| Dario Amodei | "Powerful AI" possibly 2026-27 |
| Leopold Aschenbrenner | AGI 2027 |
| Sam Altman | Near-term; superintelligence "a few thousand days" |
| Demis Hassabis | ~50% by 2030 |
| Samotsvety | ~28% by 2030 |
| Metaculus community | 25% by 2029; 50% by 2033 |
| Andrej Karpathy | About a decade out |
| AI researcher survey (n=2778) | 50% by 2040 |
Copy this. The embed stays current — the four graded links update whenever a verdict changes.
<iframe src="https://agiscorecard.com/matrix-odds?embed=1" width="100%" height="900" style="border:1px solid #33333a;border-radius:12px" loading="lazy" title="Matrix Odds Calculator — your own probability, not ours"></iframe>
This site publishes no probability for it, and this page deliberately does not either. Every number you see is one you entered. What the site does publish is graded evidence on four of the five preconditions, with pre-registered flip conditions and primary sources.
Because there is none, in either direction. “A superintelligence would keep humans alive in a simulation” is not a falsifiable claim about the present world — there is no observation today that would confirm or refute it. Indifference, extinction and cooperation are separate branches with the same evidentiary status. Anyone quoting you a percentage for this link made it up.
The physical one. AI capex has already exceeded what the source text predicted, and compute scaling is on track — both graded from primary sources. The buildout is the least speculative part of the whole scenario, which is not the part most people argue about.
Because the scenario needs every step, not an average step. If any one link fails, the scenario fails. The caveat is that the links are not independent — see the method note above, which states plainly that correlation would push the true number higher than the product.
Neither. It is a reasoning tool over a public dataset. It tells you what your own stated beliefs imply, and shows you where the evidence runs out.