AI trainer Alexander Vasylenko.Alexander Vasylenko
I started training AI while between jobs, but now I do it after work and on the weekends.
I’m doing this work so that I’m prepared for when AI can do my day job as an investment analyst.
In my two years of AI training, I’ve already seen the models get much better.
This as-told-to essay is based on a conversation with Alexander Vasylenko, a financial analyst for a large steel producer who also does AI training work in his free time. He was born in Ukraine, worked for CIBC in Canada, and is now based in New York. The following has been edited for length and clarity.
From what I’ve seen, the workload in finance will be different in a couple of years, and it’s necessary to think about how you are going to fit into that.
Two years ago, I would have had to walk an AI model through every single step and input to have a large language model calculate free cash flow. Now, I can provide five PDF files, reference three other outside sources, and ask the model to use these assumptions to calculate free cash flow.
The technological progress over the past two years is just astonishing. It’s a bit scary when you wonder how useful you’ll be going forward.
I began AI training in 2023, while I was looking for a job that involved some form of equity valuation work. I had a lot of interviews, but no offers. A recruiter had reached out to me on LinkedIn to saying she was looking for experts in finance and economics to teach AI models. I was skeptical at first, but decided to give it a shot. Very quickly, I realized that I enjoyed the work and could share my background while getting paid to contribute to this technology.
Now, my work life is nine to nine. I work my day job until five and then do AI training until nine every night. I usually work on Saturdays and occasionally a couple of hours on Sundays. In that time, I’ve seen the models become drastically better at doing financial analysis, and it can take nearly a full day’s work to create one AI prompt that can actually stump the models.
The work pays well, and I find it interesting. I’m contributing my experience to a technology that will inevitably enter every industry. The financial industry is not an exception, and by understanding how this works and will impact my job, I’ll be better prepared when the changes come.
I first started in finance in my native Ukraine, where I worked as a stock trader at a proprietary trading firm before moving to one of the largest Ukrainian investment banks. I left Ukraine 9 years ago and moved to Canada, where I studied and then became an analyst at CIBC.
After the war started in Ukraine, my family moved to the US, and I followed them to New York. Here, I started an AI training job with Remotasks making $30 an hour, which I later learned was low.
A few months later, I was hired into my current full-time role working as a financial analyst at a large steel producer, where I value strategic projects and investments and analyze how the larger economy might impact our company’s performance.
I was then contacted by another company to do more AI training work. I have continued to pick up more projects since, working predominantly on the weekends and after my main job, around 15 to 20 hours a week.
I work a lot, but my family lost everything in Ukraine, so I have people to take care of here.
The work is project-based, but I don’t recall a single day when there wasn’t a project available. Once you’re onboarded with a company, team leads reach out to you with work you might be able to help with. These projects can have very tight deadlines.
As someone writing the prompt, the pay usually ranges from $50 to $70 an hour, while reviewers can make from $90 to $120. With my experience, I’ve seen the pay stretch up to $160. The growth in pay suggests that AI is still dependent on humans.
I’m now working for one company that makes challenging questions to sell to large language models for training, and also working directly for a tech company training its model.
The goal is to write a prompt that’s challenging enough to stump the model, alongside the criteria or steps needed to solve the problem. The AI engineers will judge the model responses against these criteria to understand why the model is failing.
As a writer, you have to put small traps within your tasks and ask it to grab information from as many different sources as possible. Your job is to make every task tricky without being ambiguous.
As a reviewer, I ensure that the prompts are actually failing to answer the question, and not because the task was poorly written.
Right now, it takes between three to eight hours to write a task that triggers model failure, and we’re heading to a world where it could take a day or two.
In the future, industry professionals with skills and knowledge in their field will have AI bots that perform the tasks they currently do. The professional’s job will be to check the bots’ work and take responsibility for the outputs.
The people who will evolve in this environment are those who know their subject matter well and those who know how to combine their knowledge with AI. The closer you are to these changes and the deeper your understanding of them, the better positioned you will be.
