Co-founder of Ukrainian EdTech startup Mathema Pavlo Pedenko shared key insights from two closed DeepMind events. The main theses concern the future of local AI models, development of applied services, and new trends in robotics that will define the industry in the coming years.
Co-founder of Ukrainian EdTech startup Mathema Pavlo Pedenko shared key insights from two closed DeepMind events. The main theses concern the future of local AI models, development of applied services, and new trends in robotics that will define the industry in the coming years.
The entrepreneur shared his observations and feelings of «maximum FOMO» on his Telegram page.
He noted that one of the most important conclusions of the meetings was to rethink the feasibility of creating an application layer in AI startups. According to Pedenko, building such products is only worth it if there is access to unique data that is closed to large models, or if the service significantly simplifies processes that the technology is not yet able to perform on its own.
Investors and developers have paid special attention to local models that run directly on users’ devices (on-device), such as Google’s Gemma. Currently, they lag behind the largest advanced models by about 18 months in terms of performance. However, the pace of development is impressive: everything that specialized developers create today will be able to run locally on smartphones or laptops in two years. Pavlo Pedenko notes that this is «brain-boggling,» and the on-device models themselves will find themselves under the control of large technology giants or device manufacturers.
There is also a noticeable focus among developers on so-called world models — simulations of reality that predict the consequences of decisions and help create synthetic data for training robots.
In terms of funding, startups that raise significant investment without building their own foundation layer typically focus on solving highly specialized business problems, such as logistics, where in-house teams simply lack deep technical expertise. Meanwhile, the human data market remains a promising niche, albeit one that is rapidly consolidating.
