Last week, a significant upheaval occurred at OpenAI, the mysterious entity behind ChatGPT. The abrupt removal of CEO Sam Altman by the board triggered a revolt among numerous employees, resulting in Altman’s subsequent reinstatement. The aftermath of this event saw a flurry of media coverage examining the incident from various angles. Despite the intense scrutiny, it became apparent that our understanding of the inner workings of the company remains quite limited. Details regarding OpenAI’s technological development processes and Altman’s specific vision for future, more advanced iterations remain largely obscure.
The veil shrouding this vital aspect of the company was lifted last Wednesday when reports from Reuters and The Information came to light. Prior to Altman’s dismissal, concerns were raised by some staff researchers regarding a potentially groundbreaking advancement. The focal point of these discussions was an algorithm called Q* (pronounced “Q-star”), which purportedly showcased the ability to solve unfamiliar elementary math problems. While this may appear unremarkable initially, certain researchers at OpenAI speculated that this could signify a nascent leap in the algorithm’s capacity for reasoning—essentially, using logic to address new challenges.
The realm of mathematics often serves as a litmus test for this cognitive skill; formulating a new problem is relatively simple for researchers, but arriving at a solution necessitates a grasp of abstract concepts and systematic planning. The capability to reason in this manner is considered a fundamental element for the advancement of more intelligent, adaptable AI systems, often dubbed “artificial general intelligence” by OpenAI. According to the company’s narrative, such a theoretical system would surpass human capabilities in most tasks and could potentially pose existential risks if not carefully managed.
While an OpenAI spokesperson refrained from commenting on Q*, they did mention that the researchers’ concerns did not directly influence the board’s decisions. Two individuals familiar with the project, choosing to remain anonymous, confirmed that OpenAI had indeed been exploring the algorithm and its potential applications in mathematical problem-solving. However, in contrast to the apprehensions raised by some colleagues, they expressed doubts that this development warranted the existential concerns it seemed to provoke. Their skepticism underscores a well-known truth in AI research: the significance of advancements is often subjective at the outset. It typically takes time for a consensus to form regarding the true impact of a particular algorithm or research, as more experts delve into its reproducibility, effectiveness, and broader applicability.
For example, consider the transformer algorithm, which forms the foundation of large language models like ChatGPT. Initially developed by Google researchers in 2017, it was lauded as a significant breakthrough, but few foresaw its pivotal role in modern generative AI. It was only when OpenAI enhanced the algorithm with vast amounts of data and computational power that the rest of the industry followed suit, leveraging it to push boundaries in image, text, and now video generation.
In the realm of AI research, and indeed across scientific disciplines, the acceptance or rejection of ideas is not solely based on merit. Typically, entities with substantial resources and influence—whether scientists or companies—shape the prevailing consensus, directing the course of AI development. Currently, power in the AI sector is concentrated among a select few companies—Meta, Google, OpenAI, Microsoft, and Anthropic. This imperfect process of consensus-building is the prevailing mechanism, though it is increasingly constrained due to the shift from open research practices to more covert operations.
Over the past decade, as tech giants recognized the profitability of AI technologies, they lured academics with attractive compensation packages, enticing them away from academia to work in corporate research labs. Many AI Ph.D. candidates now join corporate entities before completing their degrees, while those in academia often receive funding or joint appointments from these companies. As a result, a significant portion of AI research now occurs within or in collaboration with tech firms driven to protect their cutting-edge advancements to outpace competitors.
OpenAI argues that its secrecy is partly motivated by the need to safeguard any developments that could accelerate the path to superintelligence, citing potential risks to humanity if left unchecked. Simultaneously, the company openly acknowledges that secrecy provides it with a competitive advantage. Ilya Sutskever, OpenAI’s chief scientist, highlighted the challenging process of developing GPT-4, emphasizing the collective effort and time invested by the entire organization. With numerous companies striving to achieve similar milestones, maintaining a veil of secrecy becomes crucial for OpenAI to preserve its competitive edge.
In the wake of the revelation about Q*, external researchers have speculated on its naming, drawing parallels to existing techniques like Q-learning and A*. The OpenAI spokesperson simply stated that ongoing research and innovation are central to the company’s pursuits. Without further insights and the opportunity for external validation of Q*’s robustness and relevance over time, speculation remains the primary avenue for all involved, including the project’s researchers. This underscores that the designation “breakthrough” was not bestowed based on scientific consensus but rather on the subjective evaluation of a select group within the organization.
