香港大学已经禁止学生使用，学生党还是要遵守相关规定。重要的事情，chatgpt就是一个聊天机器人，是聊天哈。
最开始了解ChatGPT居然是抖音上看到的，之前了解过GPT-3，最开始认为可能类似的语言模型，上手以后才发现，这玩意挺有意思，某些方面，比百度强，但是比人还差十万八千里，智力不好说，最起码现在能计算1+1=2，能解一元一次方程，可能更复杂的问题，比如leetcode的题，都是可以求解的（没试过，不清楚和搜索引擎上的答案差别多少）。
美国有一部电视剧，讲的就是一个强大的AI模型，可以预测未来发生的任何事情，阻止悲剧的发生，openai做的事情，感觉不仅仅是能成为一个强大的搜索引擎，他们的目标估计更远大。
预训练的大模型这些年特别火，从Bert开始，大模型能干的事情越来越多了，openai还是牛逼！
现在还有基于RL绘图的，有人演示一次性能出一批精美的文案(广告等等)。
国内也有，比如，百度，华为，智源的悟道等等。当然，我们都还在发展当中，有人不看好国内技术的发展，现在一提“自研”，免不了很多的酸酸的字眼。 现在研究上，这种级别的大模型，普通高校是做不了的，但是有人拿这个去解决自监督，无监督，多模态的问题，其实也是思路。
下面给大家演示一下怎么使用吧。
官方的路子。
去vscode弄一个ChatGPT的插件，有个中文的，现在很慢，你也可以用上面的账号，试一试理解代码和debug的功能，挺有意思的。这个目前是能用的。
必应。
具体能干嘛呢？
编辑
编辑
能做的事情很多。
我测试的有:
写一个用于ncnn的nms的代码
一堆废话了
下面是写的；虽然没写全，可能是网速的原因，但是写的还不错，当然，搜索引擎也是能够做到的，这个不算是很复杂的逻辑。
#include<algorithm>#include<vector>// Include the NCNN header#include"ncnn/net.h"// A struct representing a bounding boxstructBoundingBox{intx1, y1, x2, y2;floatscore;};// Compute the intersection-over-union (IOU) of two bounding boxesfloatiou(constBoundingBox& b1,constBoundingBox& b2){// Compute the coordinates of the intersection rectangleintx1 = std::max(b1.x1, b2.x1);inty1 = std::max(b1.y1, b2.y1);intx2 = std::min(b1.x2, b2.x2);inty2 = std::min(b1.y2, b2.y2);// Compute the area of the intersectionintw = std::max(0, x2 - x1 +1);inth = std::max(0, y2 - y1 +1);intintersection_area = w * h;// Compute the area of the unionintb1_area = (b1.x2 - b1.x1 +1) * (b1.y2 - b1.y1 +1);intb2_area = (b2.x2 - b2.x1 +1) * (b2.y2 - b2.y1 +1);intunion_area = b1_area + b2_area - intersection_area;// Return the IOUreturnintersection_area / union_area;}// Perform soft-NMS on a list of bounding boxes// with a given overlap threshold and a sigma valuevoidsoft_nms(std::vector& boxes,floatoverlap_threshold,floatsigma){// Sort the bounding boxes by their score in descending orderstd::sort(boxes.begin(), boxes.end(), [](constBoundingBox& b1,constBoundingBox& b2) {returnb1.score > b2.score; });// Iterate through the bounding boxes in order of scorefor(inti =0; i < boxes.size(); i++) {// Skip boxes that have already been removedif(boxes[i].x1 ==-1)continue;// Iterate through the remaining bounding boxesfor(intj = i +1; j < boxes.size(); j++) {// Skip boxes that have already been removedif(boxes[j].x1 ==-1)continue;// Compute the IOU of the two bounding boxesfloatiou_val =iou(boxes[i], boxes[j]);// If
一堆废话的过程
编辑
今天似乎不行了。😓
编辑
编辑
这种简单的代码写的很快，没仔细去和网上的对比，他这个注释，不知道是如何生成的，并且语言的表达如此的准确。
编辑
这玩意还能理解上下文信息。
编辑今天死活就不写代码了，估计英文会好使？
编辑
目前还不支持创造性方面，作为程序员也可以放心了，毕竟很多工作都是要理解业务场景才能写出来的。很少有人工作当中一直写冒泡吧。
编辑
不得不承认，是一个好工具。
当然，他也会格外的发挥，比如:
//
// Created by kanghaidong on 2022/12/10.
//
// Define a threshold for the minimum confidence score
// that a detection must have in order to be considered
const float threshold = 0.5;
// Loop over all the detections in the image
for (int i = 0; i < detections.size(); i++) {
// Get the confidence score for the current detection
float score = detections[i].score;
// If the score is less than the threshold, skip this detection
if (score < threshold) {
continue;
}
// Otherwise, keep this detection and suppress all other detections
// in its bounding box
for (int j = 0; j < detections.size(); j++) {
if (i == j) {
continue;
}
// Check if the bounding box of the current detection overlaps
// with the bounding box of the other detection
if (overlaps(detections[i].bbox, detections[j].bbox)) {
// If the other detection has a lower score, suppress it
if (detections[j].score < score) {
detections[j].score = 0;
}
}
}
}
// Define the overlap threshold for NMS
const float overlap_threshold = 0.5;
// Define the top-k parameter for NMS
const int top_k = 100;
// Perform NMS on the input data using the specified parameters
std::vector<int> indices;
ncnn::Mat::nms(bboxes, scores, indices, overlap_threshold, top_k);
// Create a neural network object
ncnn::Net ssdlite;
// Load the network parameters and weights from a trained SSD-Lite model
ssdlite.load_param("ssdlite.param");
ssdlite.load_model("ssdlite.bin");
// Define the input data for the network
ncnn::Mat in = ...;
// Run the network on the input data and get the output detections
ncnn::Mat out;
ssdlite.forward(in, out);
// Extract the bounding boxes and confidence scores from the output detections
std::vector<ncnn::Mat> bboxes;
std::vector<ncnn::Mat> scores;
extract_bboxes_and_scores(out, bboxes, scores);
// Perform NMS on the output detections using the specified parameters
std::vector<int> indices;
ncnn::Mat::nms(bboxes, scores, indices, overlap_threshold, top_k);
ncnn::Mat::nms is a method of the ncnn::Mat class in the ncnn library, which is a high-performance neural network inference framework for mobile platforms. The ncnn::Mat class represents a multi-dimensional matrix of data, and the ncnn::Mat::nms method performs non-maximum suppression (NMS) on a set of bounding boxes and confidence scores using the specified overlap threshold and top-k parameter.
The ncnn::Mat::nms method takes three arguments: the bounding boxes, the confidence scores, and a vector of indices representing the detections that were kept after NMS. The method applies NMS to the input data using the specified overlap threshold and top-k parameter, and returns the indices of the detections that were kept.
Here is an example of how you might use the ncnn::Mat::nms method:
// Define the overlap threshold for NMS
const float overlap_threshold = 0.5;
// Define the top-k parameter for NMS
const int top_k = 100;
// Perform NMS on the input data using the specified parameters
std::vector<int> indices;
ncnn::Mat::nms(bboxes, scores, indices, overlap_threshold, top_k);
实际上，Mat是不存在nms的。
你可以实现不同语言之间的转换，比如python2c++，但是正确性就不保证了。
