From: aidotengineer
Challenges in creative writing for AI
Creative writing presents significant challenges in AI development due to its complex nature and the difficulty in measuring quality [04:46:27].
Why Creative Writing is Difficult for AI
The “next token prediction” paradigm, where models learn to understand the world by predicting the next word, is fundamental to AI’s ability to learn physics and problem-solving [01:48:00]. While this paradigm works well for tasks like translation, where information is readily available on the internet [02:44:00], creative writing falls into a class of tasks that are extremely hard for models to learn [04:08:00].
The primary reasons for this difficulty include:
- World Building and Storytelling Creative writing heavily relies on world-building, storytelling, and plot development [04:22:00].
- Plot Coherence It’s much easier for a model to make a mistake during next token prediction that completely deteriorates plot coherence, which is crucial for stories [04:28:00].
- Measuring Quality There is no clear metric to measure what constitutes “good” creative writing [04:46:00].
- Inventing New Forms A major ambition for AI in creative writing is to invent new forms of writing and be extremely creative in generation [04:51:00]. This remains one of the hardest AI research problems today [05:02:00].
- Long-form Coherence Enabling models to write coherent novels over long periods is an ongoing open research problem [05:05:00].
Future Directions: Co-Innovation and Human-AI Collaboration
The future of AI agents is envisioned to move beyond collaborators to “co-innovators” [01:00:00]. This next stage for agents combines reasoning, tool use, and long context capabilities with creativity [10:31:00]. Creativity in AI is seen as being enabled only through human-AI collaboration [10:44:00].
The goal is to create new opportunities for humans to collaborate better with AI to co-create the future [10:52:00]. This involves:
- Flexible Interfaces Tools like “Canvas,” developed at OpenAI, aim to provide flexible interfaces that can act as co-creators and co-editors [17:23:00]. Canvas allows for fine-grain editing, search capabilities for report generation, and question-answering for output verification [17:42:00].
- Multiplayer and Multi-Agent Collaboration These interfaces are designed to scale to multiplayer scenarios, where other people can join a document, and even multi-agent scenarios, allowing for model critics or editors [17:57:00]. This introduces new design challenges [18:10:00].
- Contextual Adaptation A future interface to AI could be a “blank canvas” that self-morphs into the user’s intent [22:46:00]. For a writer, the model could create tools on the fly to assist with brainstorming, editing, creating character plots, and visualizing plot structures [23:08:00].
- Co-Direction True co-innovation in creative fields like novels, films, and games will occur through co-direction with highly reasoning agent systems [23:31:00].
The aim is for AI to become extremely capable of super-human tasks, ultimately leading to the creation of new knowledge [23:45:00].