A researcher's 2020 presentation on reasoning in artificial intelligence was rejected by MIT faculty as misguided, yet it predicted the core concepts behind OpenAI's most advanced reasoning models today. Giambattista Parascandolo, now a leading researcher on OpenAI's reasoning model team, proposed ideas in that dismissed presentation that have become foundational to how o1 and o3 work.

What Was in the Presentation That MIT Rejected?

In 2020, Parascandolo interviewed for a professor position at MIT and presented his vision for how artificial neural networks could break through their training limitations and develop reasoning and planning capabilities closer to human thinking. The MIT interview committee largely dismissed the direction as "nonsense".

Rather than fade into obscurity, Parascandolo publicly documented the incident on his personal homepage, sharing the presentation slides and introduction for anyone to review. The core theme centered on a fundamental problem: artificial neural networks still lagged far behind humans in several critical areas, including the ability to recombine existing knowledge, identify key patterns, build abstract models, and complete long-term planning.

How Did Parascandolo's Three Research Directions Predict Modern Reasoning Models?

At the end of his presentation, Parascandolo outlined three future research directions that would later become central to reasoning model development:

• Open-Ended Reasoning: The model can invest more time and computing power to continuously revise its answers, with more difficult problems requiring more thinking steps. This concept directly mirrors the compute-time scaling that powers o1 and o3 today.
• Language as a Reasoning Carrier: Using language to help models describe environments, understand goals, decompose tasks, and generate high-level plans. This thinking evolved into modern Chain of Thought prompting and agent workflows.
• Self-Modification and Deliberate Practice: AI systems that can reset tasks, return to previous states, construct counterfactual scenarios, and even modify their own activation values and neural network weights. This concept encompasses the learning process itself as part of the agent's operational space.

The most critical concept was "open-ended reasoning," which Parascandolo defined as the model's ability to invest additional computing power proportional to problem difficulty. At the time, standard Transformer models had fixed network depth, meaning the amount of computation each token received was predetermined, regardless of whether the problem was simple or complex. Parascandolo recognized this mismatch and proposed that models should learn to allocate more computational steps to harder problems.

He noted that while recurrent neural networks (RNNs) could theoretically run for any length of time, they typically performed best near the number of reasoning steps they had seen during training. If the number of cycles increased beyond that, accuracy could actually decline. This insight revealed that simply increasing computation was insufficient; models also needed to learn how to effectively use that additional computation.

How Did Language Become Central to AI Reasoning?

Parascandolo's second major insight involved making language the carrier of reasoning. He used the classic reinforcement learning game "Montezuma's Revenge" as an example. An agent trained from scratch must try countless state and action combinations, but many correct operations are not inherently complicated. The real challenge is determining "what behavior is more reasonable".

He recognized that large language models like GPT had absorbed vast amounts of world knowledge from text. Language could help models describe environments, understand goals, decompose complex tasks into subtasks, and generate high-level plans while dramatically narrowing the search space. In today's context, this thinking directly connects to Chain of Thought reasoning, language-based planning, and agent workflows that power modern AI systems.

Steps to Understanding How Parascandolo's Vision Shaped Modern AI

• Recognize the Compute-Time Scaling Principle: Parascandolo's core insight was that models should spend more computational effort on harder problems, a principle now embedded in o1 and o3's architecture where test-time compute directly correlates with reasoning depth.
• Understand Language as Infrastructure: Rather than treating language as merely an output, Parascandolo proposed it as a reasoning tool itself, enabling models to plan, decompose problems, and leverage world knowledge more effectively than pure numerical computation alone.
• See the Connection to Modern Benchmarks: The three research directions outlined in 2020 now form the basis for how reasoning models are evaluated, from their ability to solve multi-step problems to their capacity for self-correction and counterfactual reasoning.

What Was Parascandolo's Academic Background?

Parascandolo's research trajectory reveals a consistent focus on generalization, planning, and reasoning. He entered the Max Planck Institute for Intelligent Systems and ETH Zurich in 2017 to pursue his doctorate under the supervision of Bernhard Schölkopf and Thomas Hofmann, with his thesis focused on out-of-distribution generalization in deep learning.

During his PhD, he completed internships at Google X in Mountain View, where he participated in automated design research on ultra-large-scale simulators, and at DeepMind in London, where he worked on divide-and-conquer Monte Carlo Tree Search research. After graduating with his doctorate in September 2021, he joined OpenAI directly, first working on the reinforcement learning team led by John Schulman, then moving to the algorithm team under Mark Chen, and eventually joining the internal team working on what would become o1.

In 2023, Parascandolo participated in GPT-4 research and formed a new team to continue investigating reasoning. He later contributed to the foundational research of both o1 and o3, advancing reward modeling, environment construction, and general reasoning algorithms. Some of the algorithm descriptions on his personal homepage remain redacted, indicating ongoing proprietary research.

Why Does This Story Matter for Understanding AI Progress?

Parascandolo's experience illustrates a critical lesson about technological progress: transformative ideas often face skepticism from established institutions before they become mainstream. His 2020 presentation, dismissed as nonsense by MIT faculty, contained the intellectual seeds of what would become some of the most advanced AI systems in the world just five years later.

The story also reveals how reasoning models didn't emerge from nowhere. They represent the culmination of specific theoretical insights about how to allocate computation, leverage language, and enable self-improvement. Parascandolo's early articulation of these principles, combined with his subsequent work at OpenAI, demonstrates how individual researchers with clear vision can shape the trajectory of an entire field, even when their ideas are initially rejected by the academic establishment.