OpenAI’s ChatGPT was under no circumstances the primary giant language mannequin (LLM) to be developed, but it surely was the software that introduced LLMs into the mainstream. The discharge of this chatbot led to a modern-day gold rush during which many organizations raced to develop their very own LLMs, every making an attempt to outdo their competitor’s newest launch. Throughout this preliminary frenzy it typically appeared nearly as if this aim was being wanted with little thought of what sensible purposes these fashions would truly be used for.
A lot of that overzealousness continues to be obvious right now, but with time to step again and contemplate the chances, extra concrete purposes are additionally rising. One of many extra fashionable purposes of LLMs at current is taking the type of a software that’s embedded in e mail shoppers and phrase processors that assists customers with their writing. A typical structure of such an software includes API calls to a distant, cloud-based server that responds to consumer prompts earlier than the native software parses the response and inserts it into the textual content editor window.
Utilizing the LLM-font in inventory gedit (📷: S. Jørgensen)
A mathematician named Søren Fuglede Jørgensen had a really unusual, but additionally fairly fascinating (however largely unusual), thought for a brand new technique to combine LLMs into textual content editors. Jørgensen developed a software referred to as llama.ttf, and because the identify implies, it’s a font. However it’s also an LLM. And an inference engine. Sure, all in a font. I did point out that it was unusual, didn’t I?
This odd hack will work with any software that leverages the font shaping engine referred to as HarfBuzz, which is a reasonably lengthy checklist, and contains some notable purposes like Google Chrome. HarfBuzz features a internet meeting shaper that may execute arbitrary code, which is supposed to present flexibility in defining textual content shapes. However on this case, Jørgensen stuffed a whole LLM (TinyStories, 15 million parameters) and inference engine into the font file.
Is anybody else’s spider sense tingling proper about now? In case you can transparently run an LLM in a HarfBuzz font, what else may you run? Or what else would possibly already be operating? However setting that apart for a second, the outcome on this case is llama.ttf, an clever font. By merely selecting to make use of what seems to be an ordinary font, one can then sort sure key phrases, like “abracadabra,” and characters with particular that means, like an exclamation level, and textual content completion options simply begin to occur.
Weirdness apart, with some refinement, llama.ttf may truly be a really interesting technique to package deal up an LLM for a lot of use circumstances. It simply works in quite a lot of present purposes, and doesn’t require sending any non-public information to unknown cloud-based programs for functions which might be by no means fully clear. If you wish to dig deeper, take a look at Jørgensen’s challenge write-up and GitHub repository.