OpenAI races to build universal AI capable of handling any task

Shortly after Hunter Lightman became a researcher at OpenAI in 2022, he witnessed the explosive launch of ChatGPT—one of tech history’s fastest-growing products. Meanwhile, Lightman’s quieter work focused on training OpenAI’s models to tackle elite high school math competitions.

Today, his team—MathGen—plays a pivotal role in OpenAI’s quest to build industry-leading AI reasoning models. These systems form the backbone of "AI agents" capable of human-like computer tasks.

"Early on, we aimed to improve mathematical reasoning—an area where models struggled," Lightman told TechCrunch about MathGen’s origins.

Despite progress, OpenAI’s models remain imperfect. Even its latest systems hallucinate facts and falter with complex tasks.

Yet mathematical reasoning has dramatically improved. One model recently clinched gold at the International Math Olympiad—a contest for the world’s brightest math students. OpenAI believes these reasoning capabilities will extend to other domains, powering its long-envisioned general-purpose AI agents.

While ChatGPT’s success was serendipitous, OpenAI’s agent development represents years of deliberate effort. "Soon, you’ll simply ask computers to handle tasks," CEO Sam Altman predicted at OpenAI’s 2023 developer conference. "We call these ‘agents’—their potential is staggering."

Tech Titans Confirmed for Disrupt 2025

Netflix, ElevenLabs, Wayve, and Sequoia Capital lead an all-star lineup for TechCrunch Disrupt’s 20th anniversary. Gain actionable insights from tech’s top minds—secure tickets today and save up to $675 before price increases.

Tech Titans Confirmed for Disrupt 2025

Netflix, ElevenLabs, Wayve, and Sequoia Capital lead an all-star lineup for TechCrunch Disrupt’s 20th anniversary. Gain actionable insights from tech’s top minds—secure tickets today and save up to $675 before price increases.

San Francisco | October 27-29, 2025 | REGISTER NOW

Though Altman’s vision remains unproven, OpenAI stunned the tech world with its "o1" reasoning model in 2024. Within months, its 21 researchers became Silicon Valley’s most coveted talent—Meta poached five, offering nine-figure compensation packages.

Reinforcement Learning’s Revival

OpenAI’s agent breakthroughs tie to reinforcement learning (RL)—where AI models learn through simulated trial-and-error. Though RL dates back to 2016’s AlphaGo milestone, OpenAI spent years adapting it for computer-use agents.

Early GPT models excelled at text but faltered at basic math. The 2023 "Strawberry" project combined RL with test-time computation—letting models verify steps before answering—and pioneered "chain-of-thought" reasoning.

"Suddenly models could backtrack and self-correct—it felt human," recalled researcher El Kishky.

Expanding Reasoning Capabilities

OpenAI identified two scaling vectors: post-training computation and response-time allocation. "We don’t just build for today—we build for scalability," Lightman noted.

A dedicated "Agents" team emerged in 2023, laying groundwork for o1. Unlike rivals constrained by product demands, OpenAI prioritized AGI research—a strategic advantage.

Defining AI Reasoning

Researchers debate whether AI truly "reasons." Some emphasize computational efficiency; others focus on human-like outputs. Critics exist, but capabilities matter most—much like airplanes achieve flight differently than birds.

The Subjective Tasks Challenge

Current agents handle coding well but struggle with nuance—like online shopping. "It’s fundamentally a data challenge," Lightman explained. New techniques allow training on less-verifiable tasks.

OpenAI’s IMO solution used multi-agent exploration—now emulated by Google and xAI. Researcher Noam Brown sees rapid progress continuing: "There’s no slowdown in sight."

With GPT-5 looming, OpenAI aims to maintain its edge against Google, Anthropic, and Meta. The ultimate goal? An agent that intuitively executes any digital task—surpassing today’s ChatGPT entirely.