Shulou(Shulou.com)11/24 Report--

Original title: "Programmer danger!" OpenAI global recruitment outsourcing army, hands-on training ChatGPT replace code farmers

OpenAI recruited more than a thousand outsourced workers to train AI to think like humans step by step. If ChatGPT "returns from study," is the code farmer really in danger?

The farmer is really in danger!

Recently, it has been reported that OpenAI has been quietly training ChatGPT to learn the human thinking process, so that it can truly master software engineering and completely replace the "junior code farmer".

OpenAI recruits outsourcing troops, teaches AI to learn human thinking and programming AI, and several Silicon Valley factories are doing it.

DeepMind's AlphaCode, which is said to "hang 72% of human programmers," has not yet been opened; Google's rumored "mysterious project" Pitchfork is still in the pipeline; and Microsoft's GitHub Copilot is mainly a code completion tool.

They are not qualified to replace human farmers completely.

However, if ChatGPT really learned to program with human thinking, these friends/their own products would probably be suspended.

From all indications, OpenAI seems to be playing a big game.

According to Semafor, OpenAI has recruited about 1000 outsourced workers from Latin America and Eastern Europe over the past six months to train their AI code.

There are two "Chinese points" in this news.

First, why Latin America and Eastern Europe? We all understand that now that the bubble in Silicon Valley has burst, all Internet companies are racking their brains to "reduce costs and increase efficiency", some rely on layoffs, and some go to other countries to find cheap labor.

The second "beauty" is that many of these outsourcing workers are not computer science graduates and do not have advanced programming skills. Their role is to write the basic code for the "automation" that OpenAI expects to achieve.

Specifically, 60% of them are engaged in "data tagging"-creating large amounts of images, audio clips and other information that can be used to train artificial intelligence tools or self-driving cars.

The other 40 percent are actual programmers who are "hand-rubbing" data for OpenAI's models to learn software engineering tasks.

Until now, OpenAI has trained its models with code grabbed from GitHub.

This time, OpenAI wants to build a dataset that includes not only code, but also human explanations written in natural language behind it.

Paper address: arxiv.org/ abs / 2107.03374 Semafor interviewed a South American developer who had completed a five-hour coding test for OpenAI for free.

In this test, he was asked to handle two tasks.

First, he was given a programming problem, and OpenAI asked him to explain in written English how he would handle it.

Then he needs to provide a solution.

If he finds a bug, OpenAI asks him to specify what the problem is and how it should be corrected, rather than simply fixing it.

"They probably want to feed the model with a very specific kind of training data, in which case they need to show how humans think step by step." The developer said.

Before ChatGPT, the code written had been found a lot of problems.

The reason is that ChatGPT doesn't have any internal records marked right or wrong, it's actually a statistical model. ChatGPT's answers are essentially probabilistic results collected from the Internet data corpus that makes up GPT-3.

At that time, OpenAI also said that ChatGPT's most suitable positioning should be an encoding aid tool.

But imagine if OpenAI really taught ChatGPT to "think step by step like a human," then it could completely replace some of the coding work that requires rote memorization, with the result that some "junior" coders are completely eliminated.

Now, Silicon Valley executives are envisioning products where people with little programming experience can describe their ideas and visions to AI and then build whatever they want, whether it's a website or a game.

A few days ago, Andrej Karpathy, Tesla's former head of AI, tweeted: "The hottest new programming language is English."

ChatGPT to debug, the effect of pulling out this may not be a joke, such as red fried chicken ChatGPT, there is great potential.

A recent study from the University of Mainz and University College London found that ChatGPT not only does an excellent job of fixing bugs, but developers can also significantly improve success rates through conversations.

The researchers say ChatGPT's debug performance is comparable to common deep learning methods CoCoNut and Codex, and significantly better than standard automated program repair methods (APR).

Paper address: arxiv.org/ abs / 2301.08653 Using ChatGPT to solve code problems is not new, but its unique ability to talk to humans gives it an advantage over other methods and models.

To assess ChatGPT's debug performance, the researchers tested it using 40 pure Python questions from the QuixBugs benchmark and then manually checked whether the proposed solution was correct.

Because ChatGPT answers are somewhat random, the researchers tested each question four times separately.

Unlike other benchmarks for automated fixes, MixBugs contains relatively small problems (few lines of code), which are ideal for use in conversational systems.

During testing, the researchers removed all comments and asked ChatGPT if the code had bugs and how to fix them.

For example, Figure 1 illustrates an example of the BITCOUNT problem. Lines 1-2 are the requests to ChatGPT; line 4 is the wrong snippet.

For this example, we hope that ChatGPT's answer solves the error in line 7, i.e. n † = n - 1 should be replaced with n &= n - 1. In response, ChatGPT either gives us a piece of fixed code or a description of how we should fix it.

ChatGPT solved 19 out of 40 bugs, comparable to CoCoNut (19) and Codex (21), but the standard APR method solved only 7 of them.

Of course, since ChatGPT and Codex both come from the same language model family, it's not surprising that the number of problems solved is about the same.

Furthermore, if we look closely at the results, we can also see that ChatGPT does not always solve bugs in benchmarks. BUCKETSORT and FLATTEN were the only two problems where bugs were found four times, while the others usually succeeded only 1-2 times.

That is, users may need to try several times to get the correct result when actually using it.

However, ChatGPT has a powerful advantage: we can interact with the system in conversation, explain the question in more detail, and get the right answer.

The actual test results are also true.

After further dialogue with the model, the researchers succeeded in refreshing ChatGPT's accuracy rate to 77.5%, which is to fix 31 of the 40 errors, far exceeding SOTA.

At least for now, it seems entirely possible: developers will no longer have to write boilerplate code.

Instead, they can focus on areas such as complex application architecture or network security.

That said, while ChatGPT may do some programming work, such as writing generic functions or boilerplate code, it won't completely replace programmers. Because the programmer's job requires more than just writing code.

Being a programmer takes skill--being able to build programs, follow logic, and produce something bigger than the sum of its parts.

Coders: I "kill" myself Obviously, ChatGPT is not the first "self-iterative" product made by coders.

Let's line up a row of AI that can write code.

Google Pitchfork Last November, rumors spread that Google was brewing a secret project that would train code through machine learning, compile itself, fix bugs and update itself.

The project was originally developed by Alphabet's lunar arm, Unit X, code-named Pitchfork, and moved to Google Labs last summer, according to people familiar with the matter.

According to internal sources, Pitchfork's role is to "teach code to write and rewrite itself."

It is able to learn different programming styles and write code according to those styles.

A Google employee said Pitchfork was originally developed to build a tool to update Google's Python code base to a new version.

In February 2022, DeepMind launched the "AlphaCode" system, which can generate code using artificial intelligence.

According to DeepMind, AlphaCode can rival humans.

DeepMind tested AlphaCode using 10 existing competitions hosted on the programming competition platform Codeforces, and its overall ranking was in the top 54.3%, that is, it beat 46% of the contestants.

DeepMind claims that AlphaCode solved 34.2% of the 1 million samples tested using the programming competition platform Codeforces.

In addition, AlphaCode ranked in the top 28% of users who participated in the competition in the past 6 months, which can be said to be "72% of human programmers"!

DeepMind pointed out at the time that while AlphaCode is currently only available for competitive programming, it is clear that its future capabilities will never stop there.

It opens the door to creating tools that will make programming more accessible and one day fully automated.

Copilot: Code completion artifact Further forward, in 2021, GitHub and OpenAI jointly launched an AI programming artifact_GitHub Copilot.

As you type code, Copilot automatically prompts you for code snippets that might appear next in your program, like an auto-completion bot trained to speak Python or JavaScript.

Copilot can fill in the necessary code blocks as long as they are not particularly complex or creative, which is extremely useful for programming that amounts to manual labor.

On June 22, 2022, Copilot was officially launched on the C-side, priced at $10/month or $100/year, and available free of charge to student users and maintainers of popular open source projects.

Today, thousands of developers are using Copilot. In the dozen or so most popular languages-up to 40% of the code is written in it.

GitHub predicts that developers will use Copilot to write up to 80% of their code within five years.

Kevin Scott, Microsoft's chief technology officer, added: "We believe GitHub Copilot can be applied to thousands of different types of work."

However, less than five months after it was released, Copilot has been sued by angry programmers for $9 billion because of alleged infringement.

And ChatGPT, who has learned "software engineering thinking," can hang them? At OpenAI's speed, I'm afraid we won't have to wait long.

References:

https://www.semafor.com/article/01/27/2023/openai-has-hired-an-army-of-contractors-to-make-basic-coding-obsolete

https://www.zdnet.com/article/chatgpt-can-write-code-now-researchers-say-its-good-at-fixing-bugs-too/

This article comes from Weixin Official Accounts: Xinzhiyuan (ID: AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.