The future of AI interaction lies in shifting from single-turn, search-like queries to persistent, proactive AI assistants that collaborate with users on entire projects over long time horizons [26, 52, 39].
Achieving model truthfulness is a critical, unsolved problem. Supervised fine-tuning (behavior cloning) is fundamentally incapable of solving it, as it teaches models to hallucinate when they lack knowledge, necessitating methods like reinforcement learning to teach uncertainty [72, 80, 81].
The safe development of AGI requires a strategy of incremental deployment and careful coordination among the few major labs capable of training frontier models, with a willingness to slow down if capabilities outpace safety measures [34, 36, 50].
Post-training is a crucial and defensible competitive advantage in the AI industry, as it dramatically improves model performance and relies on complex, tacit organizational knowledge that is difficult for competitors to replicate [44, 46, 106].
Reinforcement learning algorithms he developed, like PPO, are central to current AI alignment techniques, but the field must address the challenge that RLHF optimizes for human approval, which may not align with objective truth [1, 9, 88].
2016
Graduates from the University of California, Berkeley with a PhD [11]. In its early days, OpenAI, where he is a co-founder, operates more like an academic group with small, exploratory research projects [105].
2016-2017
OpenAI attempts the 'Universe' project, aiming to create a general RL agent. Schulman later reflects that the core idea was correct but attempted about a decade too early, hampered by unwieldy systems and models that failed to generalize [116, 119, 120].
December 2017
Writes the blog post 'An Opinionated Guide to ML Research' for the OpenAI Fellows program, outlining his research philosophy [20].
Post-2017
OpenAI's culture shifts towards larger-scale engineering projects, influenced by DeepMind's success with AlphaGo [107]. The Dota project becomes one of the company's first major successes using this new approach [108].
Post-2020 (GPT-3 Release)
Schulman and his team at OpenAI fully pivot their research focus to language models, having not been fully convinced of their revolutionary potential by the earlier GPT-2 [7].
August 2022
OpenAI's GPT-4 model completes its pre-training [66], marking a major milestone in the lead-up to its public release and the further development of ChatGPT.
▶From Foundational RL Theory to Practical LLM AlignmentJul 2026
Schulman's career demonstrates a clear trajectory from developing core reinforcement learning algorithms like TRPO and PPO during his PhD for robotics [1, 3] to applying these principles at scale to align large language models with human feedback [37, 72]. His early decision to focus on policy gradient methods for continuous control problems, rather than Q-learning for discrete domains like Atari, proved foundational for the techniques now used in products like ChatGPT [1, 2].
This theme highlights the long-term, often unpredictable payoff of fundamental research; algorithms developed for robotic locomotion a decade ago are now critical infrastructure for aligning the world's most advanced AI systems.
▶A Pragmatic Approach to AI Safety and Deployment
Schulman approaches AI safety from a practical, engineering-focused perspective. He advocates for an incremental release of models where safety improvements match capability gains [36], and suggests slowing down development if AGI arrives unexpectedly fast [34]. He emphasizes the need for coordination among major labs to manage risks [50] and focuses on tangible problems like model truthfulness [73] and scalable oversight [92], while expressing skepticism about more abstract existential risk scenarios like instrumental convergence [29].
Investors and analysts should note that Schulman's safety framework is less about philosophical alignment and more about building robust, verifiable systems and establishing industry-wide engineering best practices.
▶The Future of AI as a Collaborative Partner
Schulman consistently articulates a vision for AI that moves beyond the current paradigm of one-off, search-like queries. He predicts future AI assistants will function as persistent collaborators on long-term projects, maintaining context, proactively making suggestions, and even performing work autonomously in the background [26, 39, 52]. This shift implies a deeper integration of AI into complex workflows, such as carrying out entire coding projects from high-level instructions [63].
This vision signals a strategic focus on products with high user retention and deep workflow integration, suggesting a future market based on project-based subscriptions or value-add services rather than per-query API calls.
▶The Centrality of Post-Training in AI Competitiveness
Schulman emphasizes that post-training techniques, particularly RLHF, are a primary driver of model performance and a key differentiator in the market. He notes that post-training accounts for most of the 100-point Elo improvement in GPT-4 since its initial release [44] and can substantially increase a model's 'effective compute' [106]. This process creates a competitive moat because it requires significant tacit and organizational knowledge to execute effectively [46].
The focus on post-training as a moat suggests that access to high-quality human feedback data and the complex infrastructure to implement it are as crucial, if not more so, than raw compute and pre-training data.