fak10
Anna shows what she understands
Anna uses two statements about a fictional family to infer a third: Walter is Merle’s grandfather. The recording lets you follow the entered words through the grammar and knowledge graph. It also shows two bugs found while filming. In the test at the time, Anna met 13 of 16 expectations without a false entry.
Ask an AI something and it answers. Whether it understood, you can't tell. With Anna you should be able to see it, so this episode doesn't show a result, it shows the way there.
What does Anna infer about Merle's family?
From “Merle is my daughter” and “Walter is my father”, Anna infers that Walter is Merle's grandfather.
I tell Anna about a made-up family: my daughter Merle, my father Walter and a cat called Pixel. Up top, the graph grows, a web of everything Anna knows.
- With every new name she asks whether I mean someone new. She doesn't guess.
- I never told her who Merle's grandfather is. When I ask, she still answers "Walter": she works it out from "Merle is my daughter" and "Walter is my father".
- For the cat, she asks its name. For a mosquito, she wouldn't.
- Every underlined word is a node in her knowledge. Tapping "Merle" shows what Anna knows about her and jumps to her in the graph. Tapping "is" shows the grammar: verb, present tense, root "be".
That's what "Grammar, data and logic. Every step you can follow." means.
What did filming catch in the app?
Filming exposed two bugs: Merle's gender was shown incorrectly, and the camera jumped back after moving to her. Both were fixed afterwards.
Twice the video showed us bugs in the app that nobody would have noticed otherwise. On the first take, Merle's window said "gender: male", right above "Merle is a daughter", because a list of names has Merle as a male name. Now the role you state overrides the list. On the third take, the camera flew to Merle and jumped back a second later. That's fixed too.
How did Anna and the big AI score on the same test?
On 30 September, Anna met 13 of 16 expectations with no false entries; the big AI met 11 of 16 and made one false entry.
Anna and a big AI read the same three texts. The test bench has changed since episode 1: "Jonas no longer lives in Leipzig" now counts as the end of living there, not as a negation. That makes 16 expectations instead of 14, so the old numbers can't be compared directly.
| Hits | wrote something wrong | |
|---|---|---|
| Anna (30 Sep) | 13 of 16 | nothing |
| big AI (29 Sep) | 11 of 16 | 1 entry |
With the new protocol the AI has stopped inventing relations. In episode 1 it invented four. Its one false entry is on me: my protocol can't express "no longer" as an end yet, so it writes a negation. Both fail at "works at Kranich Logistik": Anna reads the company as two participants, the AI picks the wrong form.
Every line, every text and every change to the test bench is on the test bench page.
If you want to be there when fak10 comes out: Join the waitlist
What exactly does Anna infer, and what is it based on?
Anna is not learning about a real family here. In the made-up example, she knows two statements: Merle is the narrator's daughter, and Walter is his father. From those two links, she infers that Walter is Merle's grandfather. The conclusion comes from the statements in the conversation, not from a name search or an online database. That is the difference between a stored statement and a relationship that follows from several statements.
The recording makes those steps visible. Tap a name to open its knowledge panel. An underlined word leads to a relation or a concept; tap “is” and Anna also shows the grammar she recognised. You can check which person is meant and which form the sentence uses. The graph is not a pre-drawn family tree: it grows from the statements entered in this example.
Where does the demonstration stop?
The cat Pixel is a deliberate edge case. Anna asks for the cat's name because the sentence mentions it but does not name it. She would not ask the same question about a mosquito, which the conversation does not treat as an important character. The example shows that the app does not automatically fill every gap. When a statement leaves something out, Anna can ask instead of making up a name.
This video is a look at one concrete flow, not proof that Anna understands every sentence. The test bench measures separate, explicitly defined expectations; it lists the texts and the limits of that comparison. This episode is about traceability: you can move from the sentence to a node and back again, and inspect where a conclusion came from.
For technical background, the W3C OWL 2 Primer explains inference from relations, and W3C PROV-O describes provenance. These references explain the concepts; they do not claim the app implements either standard.