Open AI Researchers Breakdown GPT-5

Following the launch of ChatGPT-5, OpenAI researchers Christina Kim and Isa Fulford, along with a16z’s Sarah Wang, offered a deep dive into the model’s advancements, the nuances of its training, and its implications for the future of AI. The conversation reveals a development process centered on intentional design, data quality, and a significant shift towards more complex, agentic workflows.

A Leap in Model Capabilities

ChatGPT-5 represents a significant step-change across key areas, moving beyond incremental improvements to deliver a more useful and capable tool.

• Coding and Front-End Development: The model’s coding ability is described as a “huge step change,” particularly in front-end web development, which is now “totally next level” compared to previous versions. This leap wasn’t attributed to a single breakthrough but to a dedicated team effort focused on curating high-quality datasets and carefully designing reward models. The team “just really cared about like nailing front end,” which involved considering aesthetics and other details that contribute to overall usability.

• Reasoning and Behavior: The team took a “very, very intentional” approach to designing the model’s behavior, specifically to address issues like the sycophancy observed in earlier models. Post-training is viewed as an “art,” a balancing act of trade-offs. For instance, making an assistant “super helpful and engaging” can lead to it becoming “overly affusive.” With GPT-5, the team had a chance to “reset and rethink” these behaviors.

A key insight into reducing hallucinations and deception is the model’s ability to “think step by step.” This allows the model to “pause and before blurting out an answer,” preventing it from confidently stating incorrect information in its desire to be helpful.

• Creative Writing: One of the most celebrated improvements is in creative writing, which is now described as “very tender and touching.” The model can assist with emotionally nuanced tasks, like writing a eulogy, demonstrating a new level of sophistication that makes the output feel as though “someone should have written this.”

Training, Data, and Evaluation

The discussion underscored that progress is increasingly driven by data and sophisticated training methodologies rather than just architectural changes.

• The Primacy of Data: The researchers are “very data-pilled,” believing that high-quality data is paramount. As reinforcement learning (RL) algorithms become more data-efficient, the quality of the data becomes even more critical. The success of specialized models like Deep Research is attributed to “so much thought and careful attention to the data curation.”

• The Role of RL Environments and Tasks: The next frontier for improvement lies in creating “really good RL environments.” The more realistic and complex the simulated environments, the better the resulting models will be. The focus is shifting towards defining the right tasks to train on, as “tasks matter more at this point given the fact that we have such a strong algorithm.”

• Beyond Benchmarks: Standardized evals are becoming saturated. Greg Brockman noted a benchmark moving from 98 to 99, indicating diminishing returns. The team now works backward from desired capabilities (e.g., creating slide decks, editing spreadsheets) and builds internal evals if none exist. This practice “nerd-snipes” researchers into “hill-climbing” on metrics that translate to real-world utility. Ultimately, the true measure of progress is “usage”—the new use cases unlocked and the number of people integrating the tool into their daily lives.

The Shift Toward Agentic Workflows

A major theme was the move from single-shot answers to longer, more complex, and asynchronous tasks performed by AI agents.

• Defining Agents: An agent is defined as something that “does useful work for me on my behalf… asynchronously.” You can “leave it and then come back and get either get a result or a question about what it’s doing.”

• The Importance of Asynchronous Work: The launch of features like Deep Research proved that users are willing to wait several minutes for a high-quality, comprehensive result, shifting the paradigm away from pure speed. This was a bet that paid off, as users value the depth of work that would have taken a human hours to complete. However, user expectations adapt quickly, and the demand for faster, high-quality results continues to grow. An interesting psychological quirk noted is that users sometimes feel “a little bit insulted” when a hard question is answered too quickly.

• Bottlenecks and the Path Forward: A key bottleneck for agents is the availability of training data for a broad range of tasks, especially for general computer usage. The solution involves a bootstrapping process: “once you have good browsing models or good use models, you can like bootstrap them to help you make data.” The roadmap for agents includes improving synthesis from private data sources (services, documents), creating and editing artifacts (docs, slides), and handling consumer use cases like shopping and trip planning.

Ecosystem Impact and the Future

GPT-5 is positioned not just as a better tool but as an enabler for a new wave of innovation.

• The Golden Age of the “Ideas Guy”: With powerful, accessible coding capabilities, the barrier to creation is significantly lowered. Non-technical founders can now build “fully interactive” applications in minutes with a simple prompt. This is expected to lead to “a lot more indie type of businesses” where the core differentiator is the idea itself, not the technical execution.

• AGI and User Adaptation: Each major release demonstrates that the frontier of AI is still rapidly expanding, pushing back against claims that progress is hitting a wall. However, humans adapt to these powerful new tools with remarkable speed. Soon after a new capability is released, “you just kind of take it for granted that you literally have this like wizard in your pocket.” This constant adaptation continually raises the bar for what is considered impressive.

• Tags:
• Agentic AI
• Reinforcement Learning
• Model Evaluation
• ChatGPT-5
• Post-Training
• Data Quality
• AI Startups