The AIRI Institute has made 76 video recordings of lectures and seminars from the 'Summer with AIRI 2026' summer school publicly available. The school took place from July 21 to August 4, 2026, in the Krasnodar Krai. The material covers 11 areas of modern AI science — from VLA models and agentic systems to AI safety and inference optimization — and is available for free in playlists on VK Video and YouTube.

image

What happened

The AIRI Institute published the complete corpus of recordings from the 'Summer with AIRI 2026' summer school: 76 video lectures and seminars are collected in two playlists — VK Video (playlist -210514085_7) and YouTube (the 'Summer with AIRI 2026' playlist). The topics are divided into 11 tracks: VLA models (Vision-Language-Action — models that connect vision, language, and robot control), agentic AI, AI4Robotics, MedAI, SafeAI, genomics, weather models, autonomous driving, materials science, interpretability, and EfficientDL. The publication took place in two stages: on August 3, recordings of 20 streamed lectures were released, and now recordings of lectures and seminars from the school itself are available. Among the speakers are AIRI CEO and Dean of the AI Faculty at Moscow State University Ivan Oseledets, head of the 'Cognitive AI Systems' laboratory at AIRI Alexander Panov, as well as Vladislav Kurenkov (adaptive agents), Dmitry Senyushkin (autonomous vision), Oleg Rogov (safe AI), and other researchers from AIRI, MSU, MIPT, Skoltech, and HSE.

Context

AIRI is the largest autonomous non-profit organization in Russia for fundamental and applied AI research, bringing together more than 250 scientific staff. The summer school is a format in which active researchers combine theoretical lectures with practical seminars; in this edition, representatives from AIRI, MSU, MIPT, Skoltech, and HSE taught. The composition of the 11 tracks is a separate signal about the agenda: it includes the frontiers of embodiment and agency (VLA models, agentic AI, AI4Robotics), safety (SafeAI), inference efficiency (EfficientDL), and applied domains — medicine, genomics, weather models, autonomous driving, materials science, and interpretability. The phased publication — first the streams, then the complete corpus — indicates a systematic open-access policy at the institute, rather than a one-off action.

Why this matters for the industry

For the industry, this is not a product release, but a reduction in the cost of personnel training. Companies can build internal training in narrow applied tracks almost for free: choose two or three areas for the team's tasks, for example inference and SafeAI, assign the recordings as a curriculum, and use the seminars as material for reviews and internal discussions. The costs are reduced to the time of an employee who will select the relevant material. The longer-term effect is indirect, through the talent market: free Russian-language material of a research level from teachers at AIRI, MSU, MIPT, Skoltech, and HSE lowers the barrier to entry into applied AI areas for students and young scientists, and for AIRI itself, the school works as a funnel. If open transcripts appear, a layer of derivative tools will become possible on top of the corpus — navigation by tracks, search by lectures, educational assistants, — but this is already a task for third-party teams, not a condition of the release itself.

Why this matters for users

For the reader, access is open now: both playlists are free, and a self-study curriculum can be built across the 11 tracks — watch lectures sequentially in a chosen area or selectively for a specific task. The seminars with the analysis of real cases have the greatest practical value: model quantization, inference optimization, and AI system safety — these are engineering skills with direct transfer to work projects. The tracks also provide an overview of applied areas — from medical AI and genomics to autonomous driving — from practicing researchers, which helps those choosing a specialization.

What is still unknown / limitations

It is impossible to assess the content itself from the announcement: the characteristic of 'research level' is a formulation of the announcement, not the result of watching the recordings, and the depth and relevance of each track were not verified. It was not reported whether the seminars are accompanied by code, notebooks, configurations, or transcripts — this determines how applicable the analyses of quantization and inference are in practice. This is an educational release, not a scientific result: there are no metrics, benchmarks, or methodology in the sources, so it is incorrect to assess the event as a breakthrough. In addition, dynamic tracks like agentic AI and EfficientDL quickly become outdated, and in six months, not all topics will retain their value.

Sources

Author

Look at AI, editorial team