Mark Chen on GPT-5, Synthetic Data and the Future of AI Research

Matthew Berman asks OpenAI’s head of research Mark Chen about the atmosphere around the August 2025 GPT-5 launch. Chen says launches move from early excitement through uncertainty about whether a model will meet expectations, then back to excitement as the pieces come together.

Asked how a research lab balances research with a successful product, Chen says the two support each other: research breakthroughs create useful products, and the product enables further research. He wants the work to meet the real world so people can experience its capabilities.

Berman asks what OpenAI learned from GPT-4 and whether new public training data has become scarce. Chen says the assumption is partly right, but teams can still improve and expand data sources, including publicly available and licensed material.

Chen contrasts GPT-4’s scaling of pretraining with GPT-5’s combination of pretraining and the reasoning work developed through OpenAI’s o-series. The goal, he says, is for people to receive deeper reasoning when needed and a fast response when appropriate, without manually choosing each time.

Another part of the effort is synthetic data generated by models. Chen credits a program led by researcher Sébastien Bubeck and says it helps improve coverage in areas OpenAI cares about. He presents this as an active research program, not as a complete replacement for human-created data.

Berman asks whether training on previous models’ synthetic outputs can produce only marginal gains. Chen says OpenAI sees potential for higher-quality synthetic data to improve models meaningfully beyond surface-level knowledge, and has seen enough promise to use it in GPT-5.

Chen declines to disclose the exact synthetic-to-human data mix or compare it numerically with GPT-4, but says synthetic data has become more prominent over time. Berman asks which fields benefit most, suggesting math, science and code.

Chen says coding was emphasized in GPT-5, while he expects synthetic-data techniques to apply broadly. Asked where they fall short, he does not identify an excluded field; he says some areas are more amenable than others but views the approach as general in principle.

On early technical bets, Chen describes exploratory teams working on architecture, optimization, reasoning and infrastructure. They start with many ideas, narrow them as evidence arrives and integrate the most promising ones. He does not disclose a specific novel architecture.

Chen returns to the challenge of combining reasoning and pretraining. He credits post-training work led by Max Schwarzer for making reasoning faster, more robust and more reliable so the two approaches could be delivered through a shared user experience.

Berman asks how OpenAI divides compute and money between pretraining and reinforcement learning. Chen says the company invests heavily in both: RL still has many early research questions, while pretraining also offers room for work on synthetic data, architecture and optimization.

Choosing a model snapshot for release is partly an art, Chen says. Teams balance the desire to eliminate troubling behaviors and polish rough edges against waiting indefinitely for a model that is perfect on every dimension. He says hands-on checks complement the decision.

Berman asks what Chen’s personal “vibe check” involves. Chen uses a mathematical problem involving a uniform random generator modulo 42 and generators modulo each prime below 42, observing whether a model makes progress toward an optimal solution.

For more everyday checks, Chen asks models to generate user interfaces and physical simulations, such as bouncing objects or fluids. He looks at visual plausibility, code robustness and aesthetics. A good-looking simulation is useful feedback, though appearance alone cannot prove that its physics is correct.

Chen also uses models as writing partners, asking them to comment on and revise his documents. He looks for persuasive and compelling style, and says he notices an improvement in creative writing from GPT-4 to GPT-5.

Berman raises comedy as a difficult test. Chen recalls that older models could be funny because their text was oddly unrealistic; he thinks reasoning models are becoming better at humor and suggests understanding what makes a joke work could be a test of reasoning.

Both discuss using AI as a brainstorming partner for personal decisions. Berman values hearing angles he might have missed, while Chen says a model can offer perspectives on situations with competing pros and cons.

Berman asks whether Chinese open models and papers changed OpenAI’s approach. Chen says OpenAI’s research roadmap toward AGI has remained broadly stable for several years, including after DeepSeek releases. He praises DeepSeek’s architecture work and efficient kernels while saying OpenAI has largely continued its existing plan.

Asked how specific the roadmap is, Chen points to the development of models that reason for longer before answering. He regards that direction as central to OpenAI’s work toward general intelligence and says it was pursued independently of external releases.

Berman asks about emergent GPT-5 capabilities. Chen says it takes time and experimentation for users to discover them, but cites OpenAI’s internal preference results for coding, claiming a win rate above 70% against earlier models. He also claims improvements in reliability, hallucinations and tool calling; those statements reflect OpenAI’s assessment.

Berman says his own testing produced more visually appealing front ends and longer single-turn code, sometimes over 1,000 lines. Chen says GPT-5 was tailored for developers and that long context and handling large repositories were areas of focus. Neither example measures code correctness across projects.

Berman asks whether the future is one general “omni” model or many specialized models. Chen finds a large general system plausible but also describes OpenAI’s interest in organizations of agents collaborating on difficult objectives. Whether a single system or an organization works better remains, in his view, an active research question.

Berman asks how much opportunity remains for developers to build agent scaffolding around GPT-5. Chen expects application-specific scaffolding to remain useful, but hopes stronger, more reliable general models will reduce elaborate workarounds for model deficiencies.

Memory and context management are another part of that scaffolding. Chen calls the finite context window a major limitation and envisions systems that can retain useful context over much longer periods, including code and personal documents, so they need to interrupt users less often.

Berman asks whether a very large context window alone would solve memory. Chen says much depends on the underlying architecture: a system must integrate and retrieve relevant older information efficiently, not merely accept a long input. He does not claim a final solution.

On multimodality, Chen describes the existing ability to process images and audio and generate images. He says perception is central to intelligence and claims GPT-5 can extract relevant information from complex images with substantially less thinking time than prior reasoning models on those tasks.

Berman asks Chen about his role in the original Codex model and what led OpenAI to focus on code. Chen says coding models could accelerate research itself because researchers implement and test ideas in code.

Chen recalls that early Codex also required new ways to measure code-generation progress. He contrasts simple early programming tasks with much more complex competition problems and sees coding as a way both to help researchers and potentially advance science.

Asked whether coding and mathematics are the key to general reasoning, Chen says many fields can teach reasoning, including math, physics, writing and even humor. Coding is strategically important because it is central to OpenAI’s daily work and to technology more broadly.

Berman asks how reinforcement learning can verify subjective work such as creative writing or humor. Chen says making verifiers and RL more general is an important research goal, but when asked about model judges and other techniques, he says the system combines approaches and declines to give details.

Chen highlights both GPT-5 and the recently released open-weight models, as well as results in programming and mathematics competitions. He emphasizes an AtCoder top-three result as a milestone; this is his characterization of the competition system, not evidence that every ordinary user will get the same performance.

Berman asks how competition results translate to daily use. Chen says the teams sought relatively general reasoning systems rather than a model that could only solve one contest; he says they used general techniques and little, if any, specialized tuning. He argues that broad capability should carry beyond the contests, a claim that still requires task-specific evaluation.

The discussion turns to the 20B and 120B gpt-oss models. Chen says he is excited about models that can run on common consumer devices, making them more accessible to hobbyists, researchers and users with specialized or on-premises needs. Hardware requirements depend on the particular model and deployment.

Chen says OpenAI tested the open-weight models under its preparedness framework, including whether adversarial fine-tuning could raise capabilities in areas such as cybersecurity and biological or chemical risks. He says the models were assessed against the company’s thresholds before release; this is OpenAI’s account of its safety process, not a guarantee of safety.

Berman asks how the team chose total and active parameter counts. Chen says target devices and latency drove the design and that early conversations with developers helped shape it. The aim was to make the models practical to run.

As familiar benchmarks saturate, Chen sees a measurement problem: human-written tests that were once difficult may no longer distinguish frontier systems well. New benchmarks are being developed quickly, but there is not yet a universally agreed method and useful tests can have short shelf lives.

Asked about interactive environments such as ARC-AGI-3, Chen considers them useful but warns against optimizing too heavily for a single benchmark. He treats a broad collection of tests as signals for whether general reasoning is improving.

Berman asks what to tell students worried about coding careers. Chen advises learning to use AI tools to accelerate work while developing ideas and a deep understanding of how the technology operates. His suggested productivity multiples are aspirations, not measured outcomes promised to every worker.

For knowledge workers more broadly, Chen expects significant economic change but believes people can adapt as new kinds of work emerge. He hopes the technology will improve quality of life through science and economic activity. These are his forecasts, not guarantees about employment or distribution of benefits.

Looking six months ahead from the 2025 interview, Chen is excited about scaling reasoning at inference time and exploring RL objectives and optimization. Over roughly two years, he hopes AI systems could become as effective as he is at AI research and drive innovations in later systems; that is a prediction and research aspiration, not a demonstrated capability.