Usability Upgrades and Advanced LLM Integration

[stevenlu137]

Since the underlying framework has been completed, I have spent a lot of effort improving the usability of AIlice over the past period. Now, the new web interface has replaced the previous Gradio-based version. Gradio is great but brought too many indirect dependencies. Currently, AIlice's default installation has been significantly streamlined, making it more suitable for ordinary users. Along with this, there is a major feature update: AIlice can now automatically save chat logs and resume from history to continue tasks. As an agent coupled with the state of the outside world, achieving this is not as simple as it is for a chatbot, so we implemented this feature late. The new interface supports the display and playback of various formats of images, videos, and audio, as well as code highlighting, LaTeX formula display, and more. In the future, we will consider adding dynamic displays of agents' call trees, but we need to think carefully about how to implement this.

The second new feature to announce is that AIlice can now configure different LLMs for different types of agents! This means we can use GPT-4 as the main agent and various open-source LLMs to drive other agents with different responsibilities, reducing inference costs as much as possible without significantly lowering success rates!

Finally, the most important development is that we finally have an open-source LLM that gives AIlice practical value! This model is openrouter/qwen-2-72b-instruct. This model has very good reasoning capabilities and is the closest open-source model to GPT-4 I have found so far. Combined with AIlice's interrupt feature, we can perform quite complex tasks! Of course, more importantly, this is just a 72B model, which can run on a PC at relatively low cost (although for ordinary users I still recommend using openrouter, as 2x4090 can only run some highly quantized versions, and these versions of LLM do not perform as well as openrouter). I think it is possible to reduce the amount of knowledge in LLMs while improving their reasoning capabilities. It is likely that in the future we will have models with higher reasoning abilities and smaller parameter sizes, and that will be the time when standalone agents become widely popular.