OpenAI: Greg Brockman Claims GPT Will Lead to AGI

OpenAI and the Vision of AGI by Greg Brockman

Greg Brockman, president of OpenAI, recently stated that the debate over the limits of text-based models is over. According to him, the reasoning models of GPT represent a direct path to Artificial General Intelligence (AGI). He declared that the GPT architecture will lead to AGI, emphasizing that the company sees a "clear line of sight" toward this goal through text models.

Brockman clarified that this approach is preferred over multimodal models like Sora, which the company considers to be on "a different branch of the technological tree." This statement was made during the Big Technology Podcast, where Brockman insisted that one of the central unresolved questions in AI research is now settled: primarily text-trained models can develop a true understanding of the world.

Closure of Sora and Strategic Choice

OpenAI has recently decided to shut down the Sora application and model. While Brockman describes Sora as an "incredible model," he asserts that it is on "a different branch of the technological tree" than the GPT reasoning series. Research on world models will continue for robotics, but on a smaller scale and without consumer-targeted products.

Brockman explained that, with limited computing power, it is not feasible for OpenAI to pursue both approaches simultaneously. He emphasized that the decision to focus on the GPT architecture is a matter of "sequencing and timing." According to him, the applications "we have always dreamed of are starting to become accessible," and the way to achieve this is through the GPT architecture.

In a discussion with Alex Kantrowitz, Brockman acknowledged the risk of neglecting world models like Sora. Demis Hassabis from DeepMind stated that Google's "Nano Banana" image model seemed particularly close to AGI. Brockman admitted, "In this field, you have to make choices. Don't you? You have to place a bet."

Debate Within the AI Community

The question of whether purely text-based models can achieve general intelligence is far from settled in the AI research community. Yann LeCun, a renowned AI researcher, has argued for years that LLMs will not lead to human-like intelligence. He criticizes their limited understanding of logic, their inability to comprehend the physical world, their lack of permanent memory, and their incapacity to think rationally or plan hierarchically.

Demis Hassabis, founder of DeepMind, shares a similar viewpoint. He believes that the scale of LLMs alone is not sufficient and that further breakthroughs are necessary. Francois Chollet, another AI researcher, defines intelligence as the ability to learn new skills effectively. He emphasizes that what matters is a system's capacity to independently form abstractions.

In a recent paper, Richard Sutton from DeepMind and David Silver, a former DeepMind researcher, called for a paradigm shift. They suggest that systems should learn from their own experiences rather than being trained solely on human knowledge. Silver has since founded his own startup focused on learning through simulation.

Jerry Tworek, a former OpenAI researcher, also describes his field of deep learning research as "finished." He claims that the next step is to build simulations of human work where AI systems can learn skills. His new startup, Core Automation, is dedicated to this approach.

However, not everyone shares this skepticism. Adam Brown, a researcher at DeepMind, recently defended the potential of the current LLM architecture. He compares the token prediction mechanism to biological evolution: a simple rule that, through "massive scaling," creates emergent complexity that people perceive as understanding. Brown argues that this complexity could even lead to consciousness.