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 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. (-5[1])The number of these things that can be processed at once is the context window. The cost of (10[1])AI compute is often reported as the price per one million of these things. For 10 points, (10[1])name these units of text that large language models split queries into and try to predict the next one of. ■END■ (10[1])