Olga Megorskaya, Toloka AI CEO on Human Data, AI Training & Evaluation

Grace Gong (0:02)

We are live. Hi, Olga, welcome to Venture with Grace.

Olga Megorskaya (0:06)

Hello, Grace. Hi, nice to see you.

Grace Gong (0:09)

First of all, I feel like what you have accomplished is like super cool, and I'm super excited to dive in to our conversation. But before diving, I want to give a quick shout out to our amazing sponsor. This episode is brought to you by Nebius, the ultimate cloud for AI innovators. Nebius provides AI infrastructure. You can count down company reliability and speed with flexibility and engineering support outmatched by hyper-skillers.

AI leaders like Meta, Shopify and Higgsfield already partnered with Nebius to run their AI workloads. Plus, venture-backed startup can save up to 150K on compute costs when they apply for access. Visit nebius.com or nebius.com/startup to learn more.

Olga Megorskaya (0:46)

I can confirm that Nebius is a cool partner. We're working with them a lot.

Grace Gong (0:51)

Amazing. I want to talk about your background. You're the head of cross-source AI platform at Yandex, which is Nebius, where it originated from, and then now you're the founder of Toloka.

Basically, we did a web search about your company. One of some of the coolest things I want to shout out is, you guys have some of the biggest AI labs as your clients, like Anthropic and others.

I'm sure it's public information if I can search it. Then some of your investors, including Jeff Bezos Fund, as well as I believe Shopify Founders Fund as well. A lot of these really cool clients and investors. I want to start with, maybe we could start with what does the platform do since you guys are the leading data partner for all stages of AI development. Maybe let's start from your background to how you see the needs of this area and then how you created the company.

Olga Megorskaya (1:56)

Yeah, sure. In short, what we do, we provide data for training and evaluation and post-training AI models.

Specifically, we cook high complexity data based on human intelligence. Because as ironically as it sounds, artificial intelligence is based on human intelligence. This is something that for a long time, everybody preferred not to talk about. At the same time right now, data companies supporting all the major labs with data are all very famous. Everybody is talking about them. The industry is growing tremendously fast for the last couple of years. But at the same time, the demand for the human data has been there from literally the very beginning. Since the industry, it was not called AI back in the days, it was called machine learning. But it started from the need of humanly created ground truth labels. Because whenever machine learning algorithms are learning, to reproduce human logic, they need some ground truth to learn on. And this ground truth can be only obtained from the human labels.

That connects with our background. We started within Yandex, because Yandex used to be one of the leading search engines in the world.

Grace Gong (3:24)

It's kind of the people of Eastern Europe-ish.

Olga Megorskaya (3:27)

Yeah, yeah, yeah. Or even the whole Silicon Valley of Eastern Europe, because apart from search, it was developing all kinds of AI-powered products, like self-driving cars, ride sharing, and voice assistants, and so on. But it all started with the search, and the first, I would say, one of the first industrial applications of machine learning in general was related to the task of ranking search results, so that to ensure that the users see only the best of the best results, the most relevant results in their top three, because nobody would look beyond like top ten results anyway. And that task required human judgments of the relevancy of the tasks. And that was the first industrial application which required large scale production of human labels, and that has always been a bottleneck. Like scaling humans is hard. It is operationally hard. At the same time, we have always had this philosophy that this is purely technological task if you want to execute it at scale. And that's what we have been developing at first within Yandex.

Then we started at first within Search, like Search Quality Department was my first job within Yandex.

Then we expanded beyond Search to serve all the other applications of AI back in the days. Then we expanded beyond Yandex and then we expanded beyond Original Market to start working with the leading companies in the world. But the nature and the concept actually stays unchanged for the last 10 years. Although it might seem kind of unexpected, but still AI is grounded on human intelligence. And human intelligence and cooking it at scale continues to be the bottleneck. That's why our industry actually exists and is growing now.