novvoo
novvoo
> 另外我感觉这个所谓的AI动作都很随机，很多动作都是乱来。例如离敌人还很远就在乱出招。能赢最后的boss也有点盲拳打死老师傅的感觉，一对真人就没戏了。我感觉用这种方法训练的AI真的很不聪明。 别这样，作者虽然技术菜了点，但最少开源了LOL
such as https://www.huxiu.com/article/781201.html
> Sure. > > WebLLM is using TVM and compiling from C++ to WebAssembly, which is more performant but vastly more complicated. This is built to be both a solid...
> @jingslunt It's using onnxruntimeweb, mature but no webGPU backend for now. webgpu merged https://github.com/microsoft/onnxruntime/pull/14579
> @jingslunt You would actually need gpt-4 to better solve this problem, and it is quite slow it works for you? I am chatgptplus user,but can't access to use gpt4...
> GPT35 works just fine ``` % python3 example.py Using api_model gpt-3.5-turbo We receive a state of type For state: use 20,20,2,15 and basic arithmetic operations (+-*/) to obtain 24...
> We could add in an time delay between every generated thought and state! openai still works better https://openai.com/research/improving-mathematical-reasoning-with-process-supervision
https://github.com/microsoft/SynapseML
> 如果能识别reality就比gfw强了 它不是不能，是投入和影响
wireguard http3 shadowsocks