Before calculators, an engineer was nothing without his slide rule. These analog devices help compute large multiplications and divisions faster than with pen and paper. They take advantage of the fact that the log of a number, when added to the log of another number, is the same as the log of the numbers' product. Arranging the figures on a slide rule in log scale makes multiplication and division as easy as addition or subtraction.

Today, large language models like ChatGPT multiply and divide billions of numbers in the blink of an eye on GPUs. GPUs parallelize those operations so that while a number is being multiplied in one part of the device, thousands of other operations are happening in other areas of the chip.

In the same way GPUs multiply numbers in parallel, our software spawns hundreds of copies of a language model at the same time, all working together for you. It can apply one question to hundreds of sources or ask a hundred questions of one source, and it answers in seconds. Parallelization is just one technique we use to make sure that you are receiving fast and accurate responses.

Most innovations - from the slide rule to the GPU to AI - increase complexity, detaching the user from the result. We hope that explanations like these as well as the software itself demystifies the process. A tool that exposes its functionality can be far more useful than a black box, especially when in the hands of an intelligent user. This belief is reflected in everything we do.