Not a real ad. Not sponsored.

You saw the ad.
Here's the fine print, at a size you can read.

I wired an AI model up as a logic gate and ran a tiny computer on it. Every claim in the ad is real, and every caveat is below. Independent experiment; I have no affiliation with TypeSafe AI (makers of Jev), Vercel, or OpenAI, and I paid for every call.

The ad

The fine print

  1. Zero mistakes. live calls, each checked by a separate program, wrong. That rules out an error rate above about 1 in ; it doesn't prove it's zero. Replay every call.
  2. s per decision. Typical time for one call, 6 calls at a time.
  3. %+ sure, every time. The least confident answer in the whole run. These are 3-bit yes/no questions, the easiest kind there is.
  4. Relative mistakes. Compared with a simulated gate that is wrong 0.1% of the time (not an AI): wrong cells after 64 steps vs .

What this does not show · The exact question sent · ChatGPT scene: a dramatization; ChatGPT was not tested.

Read this before you repost

1 Every decision, replayed

Each pixel is one cell; each row is one tick of the computer. Every bright pixel after the first row was decided by one live call to the model.

Loading recorded run…

on off repeating background (one call per pattern cell) differs from the correct answer wrong answer

Click any cell to see exactly what was sent and what came back.

23 Speed and confidence

From the same calls. The model gives a probability, not just yes or no; the answer is "yes" when it's 50% or more.

How sure the model was on every call. Nothing landed anywhere near the 50% line.
Time per call. Calls ran 6 at a time, so the run took about minutes.

4 Why zero matters: mistakes spread

In a computer like this, one wrong cell changes its neighbours on the next tick, and the damage grows. Left: a simulated gate that's wrong 0.1% of the time (not an AI). Right: the live run. Same start, 64 ticks. You can also replay the simulated run above with the "Simulated gate" button.

Side by side, 64 steps of Rule 110: a simulated gate with 0.1% random errors ends with 34 wrong cells; the live Jev run ends with 0.

The smaller pieces (not in the ad)

Diagram: three neighbour cells and the exact question sent to Jev; Jev answers 97% sure the cell is on, and a separate checker grades the answer.
One cell = one question. The answer table is never sent.
NAND truth table: 1,000 of 1,000 correct for each of the four inputs, with Jev's confidence 0.98 to 0.99 for true and 0.01 to 0.02 for false.
One call = one NAND gate (the building block of all logic): 4,000 of 4,000 right.
A stored bit shown across set, hold, reset and hold steps, from two AI NAND gates wired in a loop, with the 1,000-trial result.
Two AI gates in a loop remember a bit: 1,000 of 1,000 trials held.
Python versus the AI on the same 200-step run: milliseconds and free versus 22 minutes and 43 cents; both 0 wrong.
The honest comparison.

All live tests together: about 51,500 graded calls, 0 wrong, about 79¢.

The exact question sent

Each call sends the three cells (for example L=true; C=false; R=true) and this question. The answer table is never sent; the model works it out from the sentence each time.

"true" means
"false" means

The ad's fine print, verbatim