"AI programmers on the job" may be a gimmick, but the anxiety of elimination is still close at hand and needs to be faced positively.
▲With the rapid development of related technologies, AI is also accelerating its intervention in human real life. Photo/Xinhua News Agency
Text | Malvern
Microsoft's AI programmer is still in the demo stage, and the first AI (artificial intelligence) programmer in China has been officially announced.
According to reports, a large model company in China is recently implementing AI programming internally, using large models to assist programmers in writing code, reading code, checking bugs (vulnerabilities), optimizing code, etc. This AI program has also been assigned a formal employee number, and according to a source related to the company, the company has decided that the future goal is that 20% of the code is written by large models.
As soon as the news of "AI programmers taking up jobs" came out, "code farmers unemployed" also became a hot topic of discussion. Will programmers who have created large AI models be the first to be impacted or even replaced?
Although the anxiety of elimination is close at hand, in fact, judging from the details of the report and the reality of the industry, the worry that AI programmers will replace "code farmers" is still a little early.
It's still just an accessory tool
It can be seen from the report that the role of the AI model that was "put on the job" this time is not a "programmer who writes code" in the usual sense, but as an auxiliary program for programmers. These functions undertaken by AI programmers are to use the data induction and analysis capabilities of large models to assist in the review and optimization of existing codes.
In terms of business substance, this programmer does not replace anyone's work, but only adds a "big data review" link to the existing R&D process.
This job function, like some media using AI to correct typos, and some hospitals using AI to identify test reports and provide basic analysis, only assumes the auxiliary function of human judgment and decision-making, which is far from the real "human programmer".
Therefore, if this report is to be accurately stated, it is not that "AI works like a human programmer", but that "a large model is provided to a human programmer as an auxiliary tool".
Moreover, judging from the current development status of the industry, large models cannot replace programmers' daily R&D work such as "writing code", and they have always participated in daily work as an "intelligent coding assistant".
In fact, AI large models, as an auxiliary tool for coding, have been widely used in the industry in the past two years, and they are all involved in the role of "assistants".
For example, another large model company in China also released tools and applications related to "coding assistants" last year. Overseas, there are also many large-scale model companies that have released tools and applications that are specialized in assisting programming.
Accuracy is a big issue
In the past year, large models have become more and more involved in programming applications, including participating in the auxiliary input and continuation of code, and natural language interaction with humans, which has made the trend of "low code" more and more obvious.
Perhaps it is based on this that many large-scale model entrepreneurs and experts have also proposed that programmers may lose their jobs within 5 years.
But until now, large models have been programmed as "assistants" and have become part of the daily workflow of human programmers. The long-awaited function of "writing code and developing" has never made a breakthrough, and it is naturally impossible to replace real programmers in the short term.
The accuracy of the code generated by the AI model is an important challenge in the first place.
In 2023, research data from multiple papers show that the probability of AI engaging in simple programming tasks to generate the right one time is about 50%, which is about the same as the probability of positive and negative results obtained by flipping a coin.
In May 2023, a group of foreign studies showed that the accuracy of code generation can be increased by about 5%-10% for every order of magnitude increase in the parameters of the model. Based on this estimate, the scale of model parameters will reach at least 10 trillion level, which is more than 10 times that of the current ChatGPT 4.0.
And even with a 90% accuracy rate, it is still difficult for large models to replace the programmers who generate code. In the actual development process, it is unlikely that the model will generate a large amount of code and then be screened by the programmer.
If this is the case, the cost of reading and testing these large models to generate code far exceeds the cost of actually hiring programmers to write code, and the benefits outweigh the losses.
AI still doesn't know how to reason logically
When writing code for large models, they are not really "writing", but they are searching in the past database according to human instructions to summarize and sort out the answers. AI can't verify the code based on the logic it runs.
In other words, from the perspective of thinking and production structure, the principle of AI large model is still induction, rather than logical reasoning.
Therefore, from a principle point of view, the production logic of a large model is completely different from that of a programmer "writing code". This also explains why AI can write very "standard" or even "simple and beautiful" code today, but the operation of the code itself is often out of order.
On this point, relevant foreign AI researchers also have a consensus.
For example, ABBY, a company that applies AI solutions in areas such as finance, healthcare, and data processing, acknowledged in a previous interview that even with the most advanced systems, AI artifacts or inaccurate outputs can occur, so human verification remains essential and crucial.
Some programmers on the Internet have shared their experience of using large models to produce code, and the results show that AI may be able to write a simple function module, but when multiple function modules are needed, the code generated by AI is prone to various bugs and cannot be used.
And even if it can barely run, it can't pass the company's internal code review. This is because, although AI can try to formally meet the needs of writing code through massive search and induction, it cannot understand the logical relationship between multiple modules, so it is inevitable that logical errors will occur.
It will accelerate the elimination of new and old talents
It can be said that as of now, artificial intelligence, as a coding assistant, is still an auxiliary tool for programmers, and a tool for amplifying the efficiency of human programmers, rather than a substitute for each other.
Giving "AI" a regular employee's job name and title cannot change the reality that it is not a real employee in the short term, let alone replace "code farmer". The scenario of full automation of AI, as envisioned by the public, is likely not to come in the near future.
In particular, it is difficult for AI to replace positions that are creative and original, and require a comprehensive mobilization of creativity and logic to complete the work.
However, it is undeniable that the AI model, as an efficiency tool, will still accelerate the elimination of the old and the new in the talent market for a foreseeable period of time. For example, the efficiency of some senior talents who are the first to apply AI has increased, while at the same time, the backwardness and elimination of others have also accelerated.
At the same time, a large number of basic mechanical positions that are easy to be replaced, such as copywriting, schedule assistants, content moderation and other positions, are also easy to be replaced.
How to further pay attention to and enhance the creativity of talents in the education system, talent training, and development is a new topic raised by AI for human production relations. It is not only the "code farmers" who need to face this issue urgently, but all of us.
Written by Malvern (Media Person)
Editor / He Rui
Proofreading / Zhang Yanjun
The Beijing News commented that readers and friends are welcome to contribute. Submission email: [email protected] comment topic must be the news reported by the institutional media on the day. Manuscripts will be published on a merit-based basis and will be paid. Please write the title of the article in the subject line of the email, and attach the author's signature, identity and occupation, ID number, phone number, and bank account (including account name, bank branch name) and other information at the end of the article in accordance with our publication specifications.