Does anyone else miss when LLMs gave you raw access to the “predict the rest of the text” button? We never exhausted the possibilities

Title: Revisiting the Simplicity of Raw Text Prediction in Language Models

In the early days of large language models (LLMs), many users appreciated the straightforward ability to generate uninterrupted text through the “predict the rest of the text” feature. This direct approach allowed for boundless experimentation and exploration, unlocking creative potential without constraints.

Today, this functionality has become somewhat of a niche, with specialized tools like NovelAI and other writing applications offering similar capabilities. However, it remains unclear why the raw, unfiltered text prediction experience has diminished in popularity so rapidly. While instruction tuning and guided prompts are valuable, they often introduce an additional layer of complexity. The true power of these models lies in their ability to continue any given piece of text in a natural and realistic manner—an experience that feels more authentic when you can simply feed in a prompt and observe the model’s unfiltered continuation.

I’ve been exploring whether this kind of raw text generation can be fully replicated through carefully crafted instructions. So far, though, it seems that instruct-based approaches introduce an abstraction barrier, making it challenging to achieve the same level of spontaneity and flexibility. The simplicity of having unrestricted access to the model’s predictive capabilities remains a compelling feature, and one that continues to inspire creative exploration in AI-driven writing.

As AI technology evolves, it’s worth reflecting on what features truly empower users. Sometimes, stripping back to the basics—like raw text prediction—can unlock the most genuine and innovative uses of these powerful models.