Round 12: Tossup 18

The number of these things is the numerator of a ratio that is empirically optimal at about 20, the Chinchilla point. Computations involving the Q, K, and V vectors generated from these things are stored in a KV cache, which improves a metric called the “time to the first” one of these things. The process of converting an input into these things is commonly done (-5[1])using byte-pair encoding. During the prefill stage, IDs corresponding to these things are embedded into a vector that is then passed through multiple transformer layers. The number of these (10[1])things that can be processed at once is the context window. The cost of AI compute is often reported (-5[1])as the price per one million of these things. (-5[2])For 10 points, name these units of text (10[1])that large language models split queries into and try to predict the next one of. ■END■ (10[2]0[5])

Back to tossups

Buzzes

Summary