“Two years of Gemini and GDM have passed. We went a long way at that time with many efforts that we should feel very proud. At the same time, the competition accelerated enormously, and the last race to Aga is ongoing. I think we have all the ingredients to win this race, but we will have to pay our efforts.
The code has the most important – Agi will happen with the start when Al improves. It will probably be with a lot of human help initially, so our code performance is the most important. In addition, this must work on our 1P code. We must be the most effective scientists and scientists in the world, using our own Al.
Performance – in my experience, about 60 hours a week is a sweet performance place. Some people put much more, but they can burn or lose creativity. Many people work less than 60 hours, and a small number put into an absolute minimum. The latter group is not only unproductive, but also can be very demoralizing for everyone.
Location – it is important to work in the office, because physical being together is much more effective in communication than GVE, etc. And therefore you must be physically collaced with others working on the same. We need to minimize reporting lines in various countries, cities and buildings. I recommend being in the office at least every week.
Organization – we must have a clear responsibility and organization with groups operating on joint management and technological leadership.
Simplicity – let’s use simple solutions where we can. For example, if monitoring works, just do it, do not leave a separate model. Lack of unnecessary technical complexities (such as Lora). We will ideally have one recipe and one model that you can simply display a prompt for various applications.
Perfection – regardless of whether it is an evacuation, data source, navigation desktop or message in the internal hive, make sure everyone works and everyone is good.
Speed - we need our products, models, internal tools to make them fast. I can’t wait 20 minutes to run Python on Borg.
Iteration on a small scale – we need many ideas that we can quickly test. The best way to do this is small -scale experiments until you can increase and hope that by increasing the benefit on a large scale. This is a perfect validation. Too much work on a large scale has a habit of minor corrections and excessive adaptation to evolution, sniffing of the control point, etc. We need real wins on this scale.
Without a punt – we can’t continue to build nanny products. Our products are mastered with various types of filters and excavations. We need talented products and [to] Trust our users. “