Work & Learning

• The task descriptions still need improvement. Writing a script is challenging, so clear and easy-to-understand descriptions are essential. For example, I should provide a precise description of what each episode should entail and inform the agent teams about the number of episodes required for this task.

• While advanced Language Learning Models (LLMs) such as GPT-4 and Claude are highly sophisticated, they are still not suitable for tasks requiring both logic and creativity, such as writing novels and scripts. While many consider writing to be a creative task, it also demands logical coherence. LLMs tend to lose their memory in long narratives. For instance, while writing the fifth episode, a LLM might forget a character introduced in the first episode.

• The Alice in Wonderland Test shows how high-level Language Learning Models (LLMs) are difficult to handle this simple questions.

• TextGrad is an automated prompt optimization framework designed to help models solve complex and tricky problems. If you're familiar with Pytorch, you'll find it easy to use. It's a powerful tool.

• I also tested the newly released Claude 3.5 Sonnet model this month. As a fan of Claude, I must say it has improved intelligence.

• I've also read some compelling papers that provided insightful information. I will summarize them in separate posts.

Life