Programmers are in danger! It is said that OpenAI recruits outsourcing troops globally and trains ChatGPT code farmers step by step

Coders are really in danger!

Recently, there is news that OpenAI has quietly trained ChatGPT to let it learn the human thinking process, so as to truly master software engineering and completely replace the "junior coders".

OpenAI recruits outsourcing troops to teach AI to learn how humans think

Several major Silicon Valley companies are working on AI that can program.

DeepMind's AlphaCode is said to "beat 72% of human programmers", but it has not yet been opened; Google's rumored "mysterious project" Pitchfork is also still in the process of being brewed; and Microsoft's GitHub Copilot is mainly a code Completion tools.

To completely replace human coders, they are not qualified.

But if ChatGPT really learns to use human thinking to program, these friends/own products may be beaten down.

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

According to reports from Semafor, in the past six months, OpenAI has recruited approximately 1,000 outsourced personnel from regions such as Latin America and Eastern Europe to train their AI code.

There are two "wonderful points" in this news.

First of all, why were the locations chosen in Latin America and Eastern Europe? We all know this. Now that the Silicon Valley bubble has burst, major Internet companies are racking their brains to "reduce costs and increase efficiency." Some rely on layoffs, while others go to other countries to find cheap labor.

The second "hua point" is that many of these outsourcing personnel are not graduates of computer science, nor do they have advanced programming skills. Their role is to write the basic code for the "automation" that OpenAI hopes to achieve.

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

The other 40% are actual programmers who are "hand-scrubbing" data for OpenAI's models so that AI can learn software engineering tasks.

Previously, OpenAI had been training its models with code grabbed from GitHub.

This time, the data set OpenAI wants to build not only contains codes, but also includes human explanations written in natural language behind them.

Paper address: https://arxiv.org/abs/2107.03374

In this regard, Semafor specially interviewed a South American developer, and He once completed a 5-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 approach the problem.

Then, he needs to provide a solution.

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

"They probably want to use a very special kind of training data to feed this model. In this case, they need to show how humans think step by step." The developer said.

In the previous ChatGPT, many problems were found in the code written.

The reason is that ChatGPT does not have any internal records marking right or wrong. It is actually a statistical model. ChatGPT’s answers are essentially probabilistic results gleaned from the corpus of Internet data that makes up GPT-3.

OpenAI also said at that time that the most suitable positioning of ChatGPT should be a coding assistance tool.

But imagine if OpenAI really taught ChatGPT to "think step by step like humans", then it can completely replace some of the coding work that requires rote memorization. The consequence is that some "junior" coders are completely eliminated.

Now, Silicon Valley executives are envisioning products that would let people with little programming experience describe their ideas and visions to an AI, and then build whatever they want, whether it's a The website is still a game.

A few days ago, Andrej Karpathy, Tesla’s former head of artificial intelligence, just said on Twitter: “The hottest new programming language is English.”

Use ChatGPT to debug, the effect is outstanding

This may not be a joke, for example, the popular ChatGPT has great potential.

Recently, a study from the University of Mainz and University College London found that not only does ChatGPT do an excellent job of fixing bugs, but developers can also significantly improve their success rate through conversations.

Researchers said that the debugging performance of ChatGPT is almost the same as that of common deep learning methods CoCoNut and Codex, and is significantly better than the standard automatic program repair method (APR).

Paper address: https://arxiv.org/abs/2301.08653

Using ChatGPT to solve code problems is not new, but talking to humans Unique capabilities that give it an advantage over other methods and models.

To evaluate ChatGPT’s debugging performance, the researchers tested it using 40 pure Python problems of the QuixBugs benchmark and then manually checked whether the proposed solutions were correct.

Because there is a certain degree of randomness in the answers given by ChatGPT, the researchers tested each question 4 times individually.

Unlike other automated program repair benchmarks, QuixBugs contains relatively small problems (fewer lines of code), which is ideal for use in conversational systems.

During testing, the researchers removed all comments and asked ChatGPT if this code had a bug and how to fix it.

For example, Figure 1 is an example of the BITCOUNT problem. Among them, lines 1-2 are requests to ChatGPT; starting from line 4 are wrong code snippets.

For this example, we hope ChatGPT's answer will resolve the error on line 7, that is, nˆ= n - 1 should be replaced with n &= n - 1. In response, ChatGPT will either provide a piece of fixed code or a description of how we should modify it.

The results show that ChatGPT solved 19 of the 40 bugs, which is comparable to CoCoNut (19) and Codex (21), but the standard APR method only solved 19 of them. 7 questions.

Of course, since ChatGPT and Codex are both from the same language model series, it is not surprising that the number of problems solved is almost the same.

In addition, if we look closely at the results, we can also find that ChatGPT does not solve the bugs in the benchmark every time. Only on the two issues BUCKETSORT and FLATTEN, bugs were found four times, while others usually only succeeded 1-2 times.

In other words, users may need to try several times to obtain the correct result when actually using it.

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

The actual test results are indeed the same.

After further dialogue with the model, the researchers successfully refreshed the accuracy of ChatGPT to 77.5%, which means that 31 of the 40 errors were fixed, far exceeding SOTA.

#At least, it seems possible for now: developers will no longer need 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 common functions or boilerplate code, it won't completely replace programmers. Because a programmer's job requires more than just writing code.

Being a programmer requires skill—the ability to build programs, follow logic, and produce something greater than the sum of its parts.

Coder: I "kill" myself

Obviously, ChatGPT is not the first "self-iteration" product made by coders.

Let’s line up a row of AIs that can write code.

Google's Pitchfork

Last November, there were rumors that Google was brewing a secret project. This product would use machine learning to train code, program itself, and repair itself. bugs and can be updated by yourself.

According to people familiar with the matter, the project was originally developed by Alphabet’s lunar landing unit, the X division, code-named Pitchfork, and was transferred to Google Labs last summer.

According to internal information, Pitchfork’s role is to “teach code to write and rewrite itself.”

It can learn different programming styles and write code according to these styles.

A Google employee said that the original intention of developing Pitchfork was to build a tool to update Google's Python code base to a new version.

AlphaCode: Beat 72% of programmers

In February 2022, DeepMind launched the “AlphaCode” system, which can use artificial intelligence to generate code.

According to DeepMind, AlphaCode can rival humans.

DeepMind tested AlphaCode using 10 existing competitions hosted on the programming competition platform Codeforces, and it ranked in the top 54.3% overall, meaning it beat 46 % of the contestants.

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

In addition, among the users who participated in the competition in the past 6 months, AlphaCode’s data ranked in the top 28%, which can be said to “beat 72% of human programmers”!

At that time, DeepMind pointed out that although AlphaCode is currently only suitable for competitive programming fields, it is obvious that its future capabilities will not stop here.

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

Copilot: Code completion artifact

Moving forward, in 2021, GitHub and OpenAI jointly launched an AI programming artifact-GitHub Copilot.

When you enter code, Copilot will automatically prompt the code snippets that may appear next in the program, just like an auto-complete robot trained to speak in Python or JavaScript.

Copilot can fill in necessary code blocks, as long as they are not particularly complex or creative, which is very useful for the equivalent of manual labor programming.

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

Now, thousands of developers are using Copilot. Up to 40% of code written in a dozen of the most popular languages ​​relies on it to generate code.

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

Microsoft Chief Technology Officer Kevin Scott also said: "We are convinced that GitHub Copilot can be applied to thousands of different types of work."

However, due to alleged infringement, it could not be released before Five months later, Copilot was sued by angry programmers for $9 billion.

And can ChatGPT, which has learned "software engineering thinking", defeat them? At OpenAI’s pace, I’m afraid we won’t have to wait too long.

Reference:

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/

The above is the detailed content of Programmers are in danger! It is said that OpenAI recruits outsourcing troops globally and trains ChatGPT code farmers step by step. For more information, please follow other related articles on the PHP Chinese website!