Just for laughs: An AI (Gemini) experiment

English Open Chat
1

BACKGROUND:
I was video chatting with my Gymnasium (highschool with mandatory classical languages - Latin and Greek) friend F about AI, specifically Gemini. F is a semi-retired professional organ player and conductor. He appeared to be, like me, having nothing against AI, actually using Gemini, and fully supported my statement (from my training), that “AI result’s quality is linearly dependent on the quality of the prompt AND the capability of the human prompter to review that result”. Now, we have a lot of snow today (20 - 40 cm) and with reference to “Habemus Papam” (we have a pope), I messaged him “Habemus Nivem” earlier today. He pinged me and we met on our Jitsi server. After a bit of catching up, he asked me whether I still knew that “nivem” was the “accusativus (4th case)”, or had asked some AI. I replied it was from HI ( my memory of “nix” being snow, then applying grammar.
That was the introduction to an AI (Gemini) experiment. The prompt was NL: "Toon de naamvallen voor het latijnse woord ‘nix’ “, EN:”"Show the cases for the Latin word ‘nix’ ".

RESULTS in NL ( sneeuw = snow ):

AI-overzicht
Het Latijnse woord voor 'sneeuw' is nox (derde declinatie, vrouwelijk), en de naamvallen zijn: nox (nom.), noctis (gen.), nocti (dat.), noctem (acc.), nocte (abl.) in het enkelvoud, en noctes (nom./acc.), noctium (gen.), noctibus (dat./abl.) in het meervoud. 
Enkelvoud (Singularis)

    Nominativus: nox (de sneeuw)
    Genitivus: noctis (van de sneeuw)
    Dativus: nocti (aan/voor de sneeuw)
    Accusativus: noctem (de sneeuw)
    Ablativus: nocte (met/door de sneeuw)

Meervoud (Pluralis)

    Nominativus: noctes (de sneeuwen)
    Genitivus: noctium (van de sneeuwen)
    Dativus: noctibus (aan/voor de sneeuwen)
    Accusativus: noctes (de sneeuwen)
    Ablativus: noctibus (met/door de sneeuwen)

RESULTS in EN:

AI-overzicht
The Latin word for "snow" is nix, a third-declension noun whose stem changes from nix- (nominative) to niv- in other cases, resulting in forms like nix (nominative), nivis (genitive), nivi (dative), nivem (accusative), and nivēs (plural nominative/accusative), with the 'x' changing to 'v' as seen in its inflection

HUMAN REVIEW RESULTS with the aid of F.'s school books:

  • the NL output talks about “nox” when asked for “nix”
  • NL for “nox” is “nacht (EN night)”, NL for “nix” is “sneeuw (EN snow)”
  • both NL and EN results show hallucinated plurals. Dutch has no plural for “sneeuw”. “sneeuwen” is a verb and not a plural.
1 Like
2

Kudos to you and to F. Just a hint for next AI prompt: “How likely is it to find a Latin joke in a technical support Forum?” (And if you want, add “From the Netherlands?” ).
I just wonder how many of us here in Italy may still appreciate this…

I’m not quite sure about that. I cannot argue based on my memory (I stopped studying Latin, and reading it for that matter, some 50 years ago :wink: ) but searching the net out of curiosity produced a sentence from M. TVLLI CICERONIS PRO P. SESTIO ORATIO [V]

neque umquam Catilina, cum e pruina Appennini atque e nivibus illis emersisset

And it definitely has a plural in Italian (neve > nevi) even if in a figurative way.
Thank You again for bringing me back to High School!

1 Like
3

Latin is something which sounds quite pleasant when presented as classical choral works:

Felix Mendelssohn Bartholdy – 1809-1847 MWV B 28 Opus 115 Nr. 1

Beati mortui in Domino morientes, beati mortui, beati in Domino morientes deinceps, beati, beati.
Dicit enim spiritus, ut requiescant a laboribus suis, dicit enim spiritus, ut requiescant a laboribus suis
et opera illorum sequuntur ipsos, sequuntur ipsos.
Beati mortui in Domino morientes, beati mortui, beati, beati mortui deinceps, beati mortui deinceps.

Wolfgang Amadeus Mozart – KV 618 14th century Latin

Ave, ave verum corpus, natum de Maria virgine, vere passum, immolatum in cruce pro homine;
cujus latus perforatum unda fluxit et sanguine, esto nobis, nobis praegusta tum in mortis examine,
in mortis examine.

Both “a capella” with a TTBB male voice choir …

4

I stand corrected, a later edition of the same grammar book, doesn’t have that “no plural” marker. Adapting the post.

1 Like
5

This is a good case study on the problems with AI. I like what you said here:

I don’t know that it’s a linear dependency, but it’s definitely a dependency.

I don’t think that capability is necessarily a matter of subject-matter expertise on the part of the requester, but rather a quality of applying critical thinking to the results.

I find that AI tools are great for learning new things - the context being that as a user, I know I’m trying to learn something. Using a tool like NotebookLM is a great example of this - you can upload your notes and source material (which is partitioned from any data used for training the engine), and then you can “have a conversation” with the material.

I did something similar to this about a year ago using a local LLM running with ollama and LightRAG. I provided a copy of a college textbook that had some information in it I was trying to learn (some statistical content with complex mathematical formulas in it). I asked the LLM about the content, and it explained the terms in one of the formulas perfectly, according to the textbook. It then explained the formula itself completely incorrectly.

I had started by reading the chapter in question, and what it told me didn’t match my understanding, so I dug into it, and it eventually admitted that it had not been correct, and that my understanding was actually correct.

Errors like this are learning opportunities for the human working with the system, but the human needs to approach the use of AI not as a subject matter expert and a source of truth, but rather as being at least as fallible as a human. Wrong answers (and bad answers) to questions are an important tool for learning, provided the learner thinks about every interaction as a learning experience.

Most people who become subject matter experts eventually realize the truism “the more I learn, the less I know” - i.e., the more one learns about a subject, the more one learns there is more to know than they ever thought, and relatively speaking, the “body of knowledge” is far more vast than they anticipated, so learning more expands the view of what there really is to know to the point that, while the number of things you’ve learned has grown by +1 (for example), the number of things yet to learn has grown by +10, which pushes the “percentage of completion” value down rather than up.

And just like a human SME, an AI is putting together a lot of different concepts and ideas and trying to synthesize an output. Sometimes it’ll get it right, sometimes it’ll get it wrong, and sometimes it’ll get it “right enough” to be helpful.

I find that usually, a single prompt isn’t sufficient - it usually takes multiple iterations to get to where I need to be - and sometimes I have to throw the discussion away and start over.

I was reminded recently of the need to double-check the work of AI; I was having NotebookLM review some data, and it represented a number in a report as meaning something like “x causes a 34% growth in revenue per salesperson”, but the number wasn’t representing growth, but rather the % of respondents to a survey who said that “x causes an increase in revenue per salesperson”. Which is a very different statistic.

I like that some of these systems (including NotebookLM) provide citations so you can actually check their work. That makes it so much easier to redirect and correct - and taking the time to do that helps me learn, which is also a benefit. :slight_smile:

1 Like
6

Sapiens est consilium sapientes habere

1 Like
7

@hendersj and when there is no power or internet or cell phone service :wink: I think the folks in Berlin have been grumpy…

1 Like