AI Capabilities And Limitations

{AI Fluency Framework}** - qapla'! Humanpu' competencies je machine properties together work. One system they are!

{Generative AI}** - QaPla' 'e' cha'nob. Transformer-based {text models} - wa' {token} pagh wa' cha'nob.

{Pretraining}**: QaPla! Model 'oH {text} law' law' yIjatlhbe'chugh. "nuq 'oH pagh?" yIjatlhbe'. Qapla'!

{Next Token Prediction}** - generative AI 'oH autocomplete 'e' lo'laH. QaPla' jatlhbe'chugh, word-by-word jatlh. Fluency je hallucination 'oH result.

{Markov chain}** — QaPla'! Qo'noS wa'DIch: word-to-word connections tally. Frequency table — simple, honorable method.

Qapla' Data**: Model's knowledge = training data only. Knowledge cutoff = hard stop. No real-time {browsing}, no lived experience.

{String Matching} QaPla'**: Qapla'pu'chugh "car" yIjatlhbe'chugh "automobile" vay' yIjatlhbe'. {Semantic meaning} pagh.

{Context window} — QaQ'a' fixed-size container. Qapla' 'e' yIjatlhbe'chugh, Qapla' 'oH.

{Serial Position Effect}**: LLM'pu' 'ej SuvwI'pu' Qapla'la' - QaPla' wa'DIch 'ej QaPla' wa'maH loS, 'ach Qapla' wa'maH vIlIj pagh. Poq 'oH middle.

{Next Token Prediction} - Qapla'pu'bogh pattern**: Model-pu' instructions-wIj Qapla'pu'bogh pattern-mey-logh, ghobe' understanding-logh. Gap-mey 'oH failures-pu' 'e' yIjatlh.

{Letter}** - SaH mu'mey; {pattern-matching} 'oH, 'a 'e' jatlhbe' QaPla'

Qo'noS wa'DIch**: Hoch AI SuvwI'pu' cha' pagh law' Qapla'mey je yIjatlhbe'chugh, SuvwI' Qapla' pagh wa' neH 'oH. Cha' Qapla'mey 'oH 'e' yIjatlh!

{Next Token Prediction}** — QaPla'! AI'a' pagh Qapla'? Ghu'vam 'oH: summarize, reformat, explain common concepts. Novel territory? Sparse patterns? "True vs. sounds true"? Verify!