xutechs tells what specialists in the new profession do and what tasks IT companies hire them for
Artificial intelligence is becoming more and more powerful: neural networks can already write their own memoirs and describe in detail what photons are. AI trainers help models generate useful answers. The specialty appeared quite recently and is increasingly appearing on job search resources.
What do AI trainers do?
The training of any language model goes through two main stages. The first is pre-training: AI developers load general knowledge about the world into the neural network, for which they break huge amounts of data from different sources into fragments and give the model for training. As a result, it can generate coherent texts, but is not yet able to give meaningful answers to queries.
The next stage is refinement, or fine-tuning. Here, the model learns to fulfill requests and adhere to a given style, for example, to answer as if the question was asked by a child. To do this, the neural network needs to be fine-tuned: facts need to be checked, and it needs to be shown how to “speak” in a living language. This is where AI trainers come into play: they additionally train the neural network based on the skills it already has. To do this, specialists perform a number of operations:
- prepare the data on which the neural network will be trained: remove repetitions, make marks – for example, highlight slang and obscene language so that the model does not use them in its responses;
- compare responses from the neural network according to various parameters. Among them are literacy, correctness of facts, relevance to the request, simplicity of wording;
- decide which questions should not be answered. For example, if a user asks about health, the neural network can cause harm with its answer – even if the text is very detailed and truthful;
edit the generated answers to make them even more accurate and useful; - write standard answers – ones that will become a model example and help in its further training.
Where can I get a job?
American companies were the first to open positions for AI trainers. Thus, OpenAI was able to achieve high quality answers from ChatGPT largely due to the fact that it attracts special editors to work. According to specialists who train the model, they earn from $15 per hour.
In the Russian market, Yandex has become a pioneer in the search for specialists in the new profession, developing its own neural network YandexGPT: the company began publishing vacancies in April 2023. T-Bank also regularly publishes vacancies, and since the end of 2023, positions have begun to open at MTS and Sber.
What AI trainers need to know
The vacancies indicate that it is important to be able to:
- navigate large volumes of information and conduct fact-checking;
- quickly understand any complex topics, including unfamiliar ones;
- write in simple and understandable language;
- check and correct other people’s texts in different formats;
- argue your point of view.
As a result, the AI trainer will help the neural network answer questions correctly. For example, a person asks the model a question: “What foods should I eat to get more iron in my body?” The neural network has three possible answers. The AI trainer compares them, finds the most suitable one, and edits the text. To do this, you need to understand the topic, check all the facts, and ensure that there are no grammatical or speech errors. There is another scenario in which the AI trainer decides that there is no need to answer the question — it is better to suggest that the person consult a specialist. Here, it is important that the AI trainer can explain to colleagues why he made this decision.
There are also special requirements that depend on how the company uses artificial intelligence. For example, Sber was looking for AI trainers with experience in creating sales scripts and knowledge of the principles of designing dialog interfaces, and MTS — with an understanding of the specifics of communication with children.
In job postings, employers indicate that they give preference to candidates immersed in the topic of neural networks and technologies in general. At the same time, it is not always necessary to be an expert in the intricacies of machine learning and be able to write code. Often, companies do not expect a 100% match with the requirements, since the profession has appeared recently: the main thing is to successfully complete the test task.
Tests for AI trainer positions are very different from tasks for other content vacancies. In addition to literacy and knowledge of Russian language rules, they check fact-checking skills, compliance with the company’s ethical values, the ability to rank neural network responses – that is, to evaluate which of them are better and which are worse. The tests are extensive, and it may take several hours to solve them.
What does the future hold for the profession?
There are sometimes concerns that the profession may quickly become obsolete. However, experts agree that neural network training will continue for decades to come — and during this time, society will have new needs that will prolong the demand for specialists in the field of artificial intelligence. In essence, the task of AI trainers is to double-check and rewrite everything that can be found on the Internet. Even if thousands of people do this, it will take many years.
Artem Bondar:
“We are aiming to create an AI editorial team of hundreds of AI trainers. At the same time, we are trying to balance between quality and speed of hiring, given its complexity due to the specific competencies that our AI trainers must have. The trend for the profession will only grow, because there are more and more products based on large language models: assistants, internal tools for working with large data arrays, assistants in programming and analytics, etc.
We see from our own experience that AI trainers are a much more complex job than a classic copywriter. An AI trainer is a bit of an analyst, a bit of a journalist, a bit of a QA engineer. And most importantly, these are people who deeply understand the product they are working on. Perhaps the experience gained in the product and the insight into customer problems will transform the profession into some analogue of CX analysts (from English Customer Experience, “client experience”. — “RBK Trends” ).
There is a possibility that over time the profession will start to require deeper competencies: questions will become more complex, and it will not be fast enough to search for information — deep knowledge will be needed. Perhaps the demand for experts in different areas will grow, for example, trainers-scientists, lawyers or doctors. And the responsibility of AI trainers will become even higher: neural networks will cease to be a “toy” that is forgiven for mistakes, and will turn into a familiar everyday tool. And the increased trust will have to be justified.
