A clearer conversation about AI futures

How the interview works

Code picks each follow-up⁠^[1] from a fixed list of authored questions,⁠^[2] and most people see results after 4–8 of them. TypeSafe’s Jev reads your answers through small, focused judgments against an authored rubric⁠^[3] and never writes a question or a verdict about you.⁠^[4] Code combines those judgments into your map, its ranges and P(doom).⁠^[5] Batching small judgments keeps this cheap enough to run for free.

What your result is, and isn’t

The map

Our reading of what you wrote, not a verdict, a test or a forecast. The dashed box shows other plausible readings, not the chance that you’re right.⁠^[6]

Behind the map

Eight worldview dimensions and seven qualities of reasoning, such as willingness to update. Missing evidence stays unplaced; it never counts as a low score.⁠^[5]

P(doom)

Either the number you typed or an estimate inferred from your answers, and it says which.⁠^[7]

Editorial choices

Answers aren’t fact-checked. The questions, the rubric and even the name involve editorial choices, and the name may prime you toward risk and upside. Mixed views, uncertainty and coherent extremes are all valid.⁠^[8]

Thought leaders

Simulations: another model (GPT-5.6 Sol) answers the same questions from each person’s linked public sources,⁠^[9] and our engine reads those answers without knowing whose they are.⁠^[10] They aren’t the people’s own answers or endorsements. Treat them as prototypes: in a 2026 Pew test, answers from AI stand-ins differed from real people’s by 12 percentage points on average.⁠^[11]

Evidence so far

Your own placement

Readings land closer to people’s own placements than one guess for everyone would, with no average lean toward hope or worry.⁠^[13] Readers agree with each other far more than with people, so most of the gap is between how people write and how they see themselves. Where the two differ by more than a quarter of either axis, your result asks about it.

Rereading and rewording

No reread or reworded interview moved the outlook more than our limit of 0.10 for answers that mean the same.⁠^[14] Map ranges are never drawn narrower than ±0.05, the variation seen when a simulated person retakes the interview.

Bigger models

Jev matches GPT-5.6 Sol within the noise, for about 1/21 of the cost.⁠^[12]

“Does this feel right?”

68 of 80 people said yes (85%).⁠^[13] Read that as a ceiling: people tend to agree with an AI’s reading of them. In a NORC study, independent coders backed an AI’s codings about 20% less often than the respondents did.⁠^[15]

Inferred P(doom)

Rough. Checked against 113 numbers people typed, with those numbers hidden, it read 96 of them and came within 2× in odds only 22% of the time, no better than guessing 12.5% for everyone. It does rank people sensibly, and a percentage you type, such as “20%”, always takes precedence.⁠^[12]

Not done yet

We don’t yet have a blinded test set labeled by independent reviewers, which our evaluation plan calls for.⁠^[14] Until then, every check here compares readings with self-placements, reruns or other models. Results across interview languages haven’t been compared yet.

Who has taken it

As of October 3, 2026, 963 people had reached a result.⁠^[13] They chose to take it, mostly through a Hacker News thread (September 25 – 26) and posts on X (September 27 – 28), and the two waves landed far apart: a median outlook of 0.32 against 0.68 on the 0–1 Doom–Bloom axis (see the comparison). So our numbers describe who showed up, not the public, and we publish figures only for groups of at least 10 people.

Why the questions are authored

An AI could improvise the interview. Fixed, authored questions keep the wording neutral and identical, so every participant and every simulated thought leader answers the same questions and comparing them means something. AAPOR’s 2026 guidance rates AI interviewers that generate questions, change their wording or influence the flow as high risk.⁠^[16] Here the wording never changes, and Jev shapes the flow only through judgments that code uses to pick among the questions.

Versions and changes

Algorithm

0.7.5

Questions

0.4.0-draft

Rubric

0.1.0-draft

Model

jev-1.13.0

Each saved result keeps the versions it was made with. Fixes that need no new judgments also reach older results when they’re shown.

1. October 4, 2026 · 0.7.5Places people between the map’s five levels instead of only on them, and re-reads saved results⁠^[17]
2. October 1, 2026 · 0.7.4Interviews in 10 languages; Jev reads answers as written and is told the interview language
3. September 29, 2026 · 0.7.2–0.7.3Keeps typed P(doom) answers as short as “30%”, which had been replaced by inferred values, and asks one more question when a reading is split or far from your own placement⁠^[12]
4. September 27, 2026 · 0.7.0–0.7.1Fixes how P(doom) is shown (a typed 20% had appeared as about 3%), asks directly about the scale of change and P(doom), stops holding results back for reasoning coverage, and re-reads saved results⁠^[18]

Research notes on GitHub record each change with its evidence.

Example using Elon Musk

Here are some example JSON results from Elon Musk’s simulated assessment.