On any given day, Ilya Poloskin has a dozen or so agents perform various “missions.”
One such mission, he said, could be “to be a better CEO.”
“So it effectively summarizes all your meeting notes, Google Drive docs, and Slack messages and provides coaching and an executive summary of what happened, what’s missing, and where decisions are stalled,” Polosukhin told Business Insider. “So it runs every week.”
Poloskin calls these agents “billionaire, chief of staff-level support.” The explanation was “literally” written in the prompt: “You’re a billionaire’s chief of staff,” he said.
It’s an early glimpse of the future Poloskin sees not just for individual workers and CEOs, but for the global economy as a whole: a world where agents can make deals, coordinate supply chains and broker deals on behalf of people and large corporations. And in his view, we are completely unprepared for it.
“I think the bigger problem is that systems are fundamentally not ready to take advantage of AGI (artificial general intelligence),” he said. These systems include “society, the Internet, government agencies, etc.”
Polosukhin is one of the key figures behind generative AI. In 2017, he co-authored a seminal research paper, Attending Is All You Need, introducing the Transformer architecture, a new approach to building AI models. This groundbreaking paper is why ChatGPT has a “T” at the end.
peel off the black box
The trajectory of AI hardly surprises the researcher-turned-founder.
In the same year that the Transformer architecture paper was published, Poloskin launched NEAR AI, based on the idea that machines could eventually generate software. His hypothesis was that humans would talk to computers in a natural language like English, and the machines would write the code.
“Back in 2017, that seemed pretty ridiculous,” he says. Today, it’s called vibe coding.
Polosukhin isn’t surprised by the features some models are currently displaying. human On Tuesday, the lab announced that the latest preview model, Mythos, is so capable of finding and exploiting vulnerabilities that it is restricting access.
Poloskin said he has been warning for years that “the model is going to start breaking everything.” He explained to Business Insider that each iteration of the model is a “cat and mouse game” in which anything fixed in the previous model may break.
In a world where people use AI agents to manage their health and businesses manage their logistics, Poloskin sees a need for back-end trust and security layers aimed at protecting against those risks.
At NEAR, Poloskin is building an infrastructure to reduce the reliance of AI agents on a single company, such as Frontier AI Labs, to control and oversee every step of a task.
In practice, this could mean that AI agents that process users’ login information, book travel, and move money to pay for airline tickets won’t require users to blindly trust a single gatekeeper.
“It’s going to contain all the information,” Poloskin said of the AI model processing the data. “It’s literally your life, so you don’t want any one company to control it or have access to it.”
Another risk Polosukhin wants to guard against is manipulation. From news summaries to investment recommendations, more and more people are turning to AI to stay informed. The AI lab, or malicious actors within it, could secretly shape these answers, Poloskin said.
One example of this occurred last year at xAI, which repeatedly brought up “white genocide” in unrelated responses after the company said Grok was an “unauthorized change” to its backend.
Polosukhin’s pitch to NEAR is to develop an open-source, auditable platform that gives users a deeper understanding of how AI systems work, rather than treating them as black boxes.
AI still requires oversight
His own agent is not entirely reliable at this time.
Poloskin shared with Business Insider how his agents can aggregate news about the US-Iran ceasefire and provide market perspective. There are also “development agents” who do the coding and “growth agents” who can suggest steps to improve specific metrics at your company.
They’re helpful, but Poroskin doesn’t let the AI off the lead. Researchers said AI systems still require extreme caution.
In his view, AI still struggles to make sound decisions, even though online conversations about AI may overestimate its current progress.
“If you let it run and do something, it reverts to meaningless things,” he said of AI models. “So you have to babysit it at your discretion.”