Over the past two years of the Gemini program and GDM, we have achieved significant progress, thanks to our collective efforts. However, competition in the race toward AGI has intensified, prompting us to accelerate our initiatives. We possess the necessary elements to succeed, but we must enhance our endeavors. The quality of our code is paramount; AGI will advance rapidly when AI starts improving itself, likely with substantial initial human assistance. Thus, optimizing our own 1p code is essential, making us the foremost experts and most efficient coders in AI. Regarding productivity, about 60 hours a week is optimal. While some may work longer and risk burnout, others who contribute minimally not only compromise productivity but also affect team morale negatively. Collaboration is vital; working in an office enhances communication more than remote setups. We should minimize reporting structures across locations and prioritize in-person collaboration, ideally being present in the office every weekday. Organizational clarity is necessary, with defined responsibilities and high-functioning teams under shared management. We should embrace simplicity, opting for straightforward solutions when feasible. Instead of creating separate models, if prompting suffices, we should utilize that approach, avoiding unnecessary complexities. Our goal is to establish a single, adaptable model for various applications. Excellence in evaluations, data sources, dashboards, and internal communications is non-negotiable. Speed is essential; our products and tools must operate efficiently—waiting long periods for simple tasks is unacceptable. We should iterate through small-scale experiments, testing numerous ideas quickly to achieve scalable success.

Concentrating solely on large-scale projects often leads to ineffective adjustments and overfitting. We must aim for tangible wins that can be scaled. Additionally, we can’t continue to produce overly cautious products filtered by unnecessary constraints. We need to trust our users and develop robust, capable solutions. On the topic of information leaks, they can be classified into three categories: accidental leaks from careless sharing, misguided attempts to assist by trying to explain matters to the press, and malicious leaks that promote personal agendas over the company's interests. All three types are considered serious offenses that can lead to termination. We have zero tolerance for such leaks, and our dedicated team actively investigates them, resulting in the dismissal of over 20 individuals recently. Many more investigations are ongoing. Remember, do not screenshot or record sensitive information; we are improving our detection capabilities, and those involved are no longer with us. Regardless of intent or perceived inconsequence, breaches of this nature will lead to termination.