From Vibe Coding to Vibe Researching: OpenAI’s Mark Chen and Jakub Pachocki

Episode Summary

Executive Summary: OpenAI leaders Jakob Pohutsky and Mark Chen discuss GPT-5 as a step toward mainstream reasoning and agentic behavior, and outline a long-term goal of building an automated researcher that can discover new ideas. They argue evals are shifting from saturated benchmarks to economically meaningful, long-horizon, and discovery-oriented measures, with compute still the main bottleneck.

Main Topics: GPT-5 and mainstream reasoning (Priority: 5/5): The launch of GPT-5 is framed as an effort to unify instant-response and reasoning models, making deliberate reasoning and agentic behavior the default user experience rather than a special mode. Evals shifting from saturation to real-world discovery (Priority: 5/5): The speakers say traditional benchmarks are increasingly saturated, so progress should be measured by harder, economically relevant tasks, scientific discovery, and autonomous operation over longer horizons. Automated researcher as the core roadmap (Priority: 5/5): OpenAI’s research target is described as an automated researcher that can discover new ideas, extend reasoning time horizons, and eventually advance both ML and other sciences. Why RL continues to work (Priority: 4/5): Reinforcement learning is presented as a versatile, still-evolving method whose gains have been amplified by language modeling, better environments, and richer training objectives. Codex, real-world coding, and vibe coding (Priority: 4/5): GPT-5 Codex is positioned as adapting raw reasoning power to messy real-world software work, with dynamic latency and more useful coding behaviors; coding is already shifting toward AI-assisted defaults. Research culture, talent, and org design (Priority: 4/5): The conversation emphasizes protecting fundamental research, creating clear mandates between research and product, hiring for hard-problem solvers, and maintaining a culture that keeps people learning. Compute, prioritization, and constraints (Priority: 4/5): Compute is repeatedly described as destiny: even as algorithms improve, the org remains compute-constrained, so prioritization and portfolio management are central to the research strategy.

Key Arguments: GPT-5’s main purpose was to bring reasoning into the mainstream by removing user confusion about which model mode to choose. Current benchmark suites are close to saturated, so improving from 96% to 98% matters less than measuring discovery, autonomy, and economically meaningful performance. The most important frontier is not just better answers, but models that can independently discover new ideas and operate for hours or longer without drifting. Reasoning and long-horizon agency are tightly linked because both require consistent problem-solving, self-correction, and adapting after failure. RL keeps working because it can be applied to rich environments like language and reasoning, enabling targeted expertise rather than only broad pretraining gains. Real-world coding requires more than intelligence: models must handle style, soft preferences, workflow latency, and messy environment constraints. OpenAI protects fundamental research by giving researchers room to explore while maintaining clear company-level priorities and product coordination. Compute remains the most binding resource, more so than data, and the team does not expect that constraint to disappear soon. Great researchers are defined by persistence, honesty about failure, comfort with hard problems, and the ability to sustain motivation over years. The org’s strength comes from a deep bench, strong mission alignment, and a culture that avoids learning plateaus.

Data Points: High school coding norm: "vibe coding" - Anecdote about how younger coders now see AI-assisted coding as the default. Current reasoning horizon: 1 to 5 hours - Approximate horizon the models can now reason and make progress over, according to the speakers. Competitions mentioned as evaluation markers: IOI, AtCoder, IMO - Described as real-world success markers for future research capability. Previous model generation outcome: GPT-5 improved on O3 - The speakers say GPT-5 is a definite improvement over O3, especially for daily usefulness. Coding workflow example: 30-file refactor in 15 minutes - Used to illustrate how coding tools have already changed developer behavior. Release timing reference: today - GPT-5 Codex is described as having dropped on the day of the conversation. Model performance milestone: number two in AtCoder - Referenced as a notable coding competition result with only first place left.

Pivotal Quotes: "The big thing that we are targeting is producing an automated researcher." — Jakob Pohutsky: Stated as the central research goal and framing for future milestones. "The future hopefully will be vibe researching." — Jakob Pohutsky: A shorthand vision for AI-assisted research becoming as natural as vibe coding. "Compute is destiny." — Mark Chen: Used to emphasize that research progress is still fundamentally limited by compute resources.

Implications: The industry is moving from benchmark-chasing to autonomous, economically relevant AI systems. Winners will likely be labs that can scale compute, extend reasoning horizons, and turn research agents into practical discovery engines.