The MIT‑Rejected 2020 PPT That Anticipated OpenAI’s o1 and o3 Reasoning Strategies

In 2020 Giambattista Parascandolo gave a MIT interview presentation on using GPT for reasoning; most professors labeled the direction "nonsense".

The report’s theme was how artificial neural networks could break out of their training distribution to achieve human‑like generalization and planning, proposing three research directions.

Parascandolo defined open‑ended reasoning as allowing a model to spend more time and compute to iteratively refine answers, especially for difficult problems, mirroring today’s compute‑extended reasoning.

Standard Transformers have fixed depth and compute per token, making computation independent of problem difficulty; a long question may have a simple answer, while a short one may need many reasoning steps. RNNs, which can run repeatedly, seemed more suitable, but experiments showed RNN accuracy peaks near the number of inference steps seen during training and can drop with extra cycles.

At the time, the strongest multi‑step planning relied on model‑predictive control and Monte‑Carlo tree search, where the neural network predicts the environment or evaluates value and an external search expands future trajectories. Parascandolo wanted to embed long‑range reasoning directly inside the network.

The second direction was to make language the reasoning carrier . Using the classic game Montezuma’s Revenge as an example, he showed that a reinforcement‑learning agent trained from scratch lacks the ability to judge “what action is more reasonable”. Language can describe the environment, understand goals, decompose tasks, generate high‑level plans, and dramatically shrink the search space.

The third direction envisioned AI systems that could reset to any memory state, construct counterfactual scenarios, and even read, copy, and modify their own activations and weights—effectively bringing the learning process itself into the agent’s action space for deliberate practice and transfer learning.

A 2021 blog by Parascandolo rebutted the claim that neural networks need massive data unlike the human brain, likening large‑scale pre‑training to biological evolution and fine‑tuning/contextual learning to an individual’s lifelong learning, arguing that scaling data and compute can still yield significant gains.

Parascandolo’s background includes a PhD on OOD generalization under Bernhard Schölkopf and Thomas Hofmann, internships at Google X and DeepMind, joining OpenAI in 2021, contributing to GPT‑4, and later leading research for o1 and o3, where many of the 2020 ideas reappear.

Thus, the once‑dismissed PPT essentially outlined the roadmap that modern OpenAI reasoning models have followed.