Productizing Large Language Models
Large Language Models (LLMs) are known for their near-magical ability to learn from very few examples -- as little as zero -- to create language wonders. LLMs can chat, write poetry, write code, and even do basic arithmetic. However, the same properties that make LLMs magical also make them challenging from an engineering perspective.
At Replit we have deployed transformer-based language models of all sizes: ~100m parameter models for search and spam, 1-10B models for a code autocomplete product we call GhostWriter, and 100B+ models for features that require a higher reasoning ability. In this post we'll talk about what we've learned about building and hosting large language models.
Any sufficiently advanced bullshit is indistinguishable from intelligence, or so the LLM thought. LLMs are super suggestible -- in fact, the primary way to interact with LLMs is via "prompting." Basically, you give the LLM a string of text and it generates a response, mostly in text form although some models can also generate audio or even images. The problem is, you can prompt the LLM with nonsense and it will generate nonsense. Garbage in, garbage out. Also, LLMs tend to get stuck in loops, repeating the same thing over and over again, since they have a limited attention span when dealing with some novel scenarios that were not present during training.
Luckily, there are some controls to make these two problems more manageable. The first is tweaking the so called "temperature" of the model. This is a parameter that affects the likelihood of the model picking a less likely word. The higher the temperature, the more random the output. The lower the temperature, the more likely the model is to generate words that are similar to the words it has seen before. Setting the temperature to zero makes the LLM more predictable.
