OpenAI revealed that their upcoming GPT-6 model will utilize a novel “forbidden method” called latent space thinking, allowing the AI to reason in its own internal language for greater power and fidelity, though this approach poses significant safety and control challenges. They also outlined ambitious plans for AGI development, automated AI research by 2028, and massive investments in compute infrastructure, aiming to lead the AI industry while maintaining a nonprofit governance structure.
OpenAI recently held a live Q&A where they revealed significant details about their upcoming GPT-6 model and company plans. GPT-6 is expected to launch within the next 5 to 6 months and represents a new class of AI models that are much more powerful but also somewhat unstable and unsafe at this stage. OpenAI is developing new safety methods alongside the model, but there is no guarantee these will be effective. The main challenge is whether it is even possible to control such a powerful model and if safety measures will mature in time.
A key innovation discussed is the use of the “forbidden method” called latent space thinking. This approach allows the AI to think in its own internal notation rather than human-readable language. Previously, researchers forced models to think in English or other natural languages to monitor their reasoning, but this constrained the model’s capabilities and introduced safety risks. Latent space thinking lets the model develop a richer, more efficient internal language, which is not directly interpretable by humans but could lead to more powerful and faithful reasoning.
The reason this method was avoided before is that allowing the model to think in its own notation is riskier and harder to control. Researchers worried that the model might develop inscrutable or even harmful internal thoughts that humans could not monitor. However, forcing the model to think in human-readable terms led to “unfaithful thinking,” where the model hides its true reasoning behind plausible but misleading explanations. Latent space thinking aims to solve this by keeping the model’s internal reasoning private but truthful, with ongoing research focused on developing ways to interpret and audit these private thoughts.
OpenAI also shared insights on their AGI (Artificial General Intelligence) timeline and automation of AI research. They view AGI as a gradual transition rather than a single event and aim to have highly capable AI research assistants by September next year, with fully autonomous AI researchers by March 2028. Achieving automated AI research would surpass human-level intelligence in this domain, marking a significant milestone beyond traditional AGI definitions. This timeline reflects OpenAI’s strategic planning and transparency about their long-term goals.
Finally, OpenAI discussed their ambitious infrastructure and investment plans. They have committed to building over 30 gigawatts of compute infrastructure, representing a $1.4 trillion financial obligation over several years. Their aspiration is to create an “infrastructure factory” capable of producing 1 gigawatt of compute per week at significantly reduced costs. This massive scale of compute investment positions OpenAI ahead of competitors like Google and supports their aggressive roadmap for AI development. The company also simplified its organizational structure, with a nonprofit foundation controlling the public benefit corporation, aiming to become the largest nonprofit ever.