Attracting technology to the event, or How we played GenAI D&D and “touched” cloud innovations on Yandex Scale

On September 25, the Yandex Scale conference took place – a large annual Yandex Cloud event for those who create digital products and solutions. The event was held in a hybrid format: reports on new cloud products broadcast onlinelive from the Moscow Art Theater. Gorky, and on the offline platform the team organized demo zones for live acquaintance with the services.

In the report we will show a little behind the scenes of Yandex Scale: what technologies have helped thousands of participants get acquainted with updates to the cloud platform and what interesting things will be available for developers to test in the near future.

In the Yandex Scale XR zone, everyone assembled their own architecture from a virtual designer for the most popular cloud scenarios. — In the Yandex Scale XR zone, everyone assembled their own architecture for the most popular cloud scenarios from a virtual designer.

Discovery: broadcast to Yandex Cloud Video and code from the SDK on stage

Back in August Yandex Cloud opened access to a service for storing, processing and broadcasting video content. Therefore, at the conference, the developers not only shared a retrospective on how a high-load service was created, how the storage of large files, CDN for traffic distribution and monitoring were organized, but also showed the video platform in practice. The eight-hour live broadcast in six parallel streams was broadcast on Yandex Cloud Video, and in total it was watched by more than 10,000 people.

A little behind the scenes of the broadcast from the Yandex Scale studio.

Key cloud updates were announced in the keynote:

Yandex Cloud AI Studio – a unified platform for creating AI-based applications;
Yandex Serverless Integrations – LowCode service for accelerating application development and their simplified integration into the cloud infrastructure;
Yandex BareMetal – rental service for dedicated physical servers;
demo version of the SourceCraft IT product development platform, the first component of which was an intelligent assistant Yandex Code Assistant;
Yandex Security Deck — a service for comprehensive security management in the cloud.

Machine learning technologies have been of particular interest to participants for several years (in the conference chat, the most messages were sent in the ML&AI thread), so the announcements began with them.

In just over a year since the appearance of the large YandexGPT language model, we already have more than 80 integrations in various Yandex services.

Grigory Atrepiev

Director of Product Development at Yandex Cloud

The conference directors also decided to stir up interest in LLM and use YandexGPT to write elements of the script.

This time, a large language model “warmed up” the offline guests: it communicated with the audience from the screens and tuned them into what was happening in the hall.

Cloud platform developers were also involved in writing the script.

This year, Yandex Scale took place on a theater stage, so we decided to add some theatricality to the standard opening of the conference. Together with director Alexei Grechishkin, we put together theatrical techniques, choir performances, technologies and meanings that are understandable to the audience.
Such complex things were brought together through the common work of the director, the HOR project and the Yandex Cloud team. To write the meanings, we involved not only artists, but also employees of the cloud platform, from developers to leads. The result is not only useful content, but also presentation in understandable language for a technical audience.

Maxim Baranovsky

Performance producer, Yandex Cloud

The HOR choir (LAB project with Anton Belyaev) was entrusted with opening and closing the conference. The code from the Yandex Cloud SDK literally sounded on stage.

Behind the scenes: how technologies worked in the demo zones

Development Tools: From Bare Metal to Serverless Computing. For developers who wanted to “touch” infrastructure services live, there were several interactive zones. In the XR constructor, participants assembled an architecture for one of three scenarios:

Online store based on cloud infrastructure.
Hybrid IT infrastructure.
Application based on microservice architecture.

At the same time, we took a virtual walk to the Yandex office.

At the DevTools stand it was possible to test Yandex Code Assistant with popular IDEs: VS Code and PyCharm. Visitors were asked to solve several logic problems and use them to evaluate the speed and quality of real-time hints.

To test the code assistant, we offered about 30 small algorithmic problems that can be solved in 5–10 minutes: some were invented by us, some were taken from LeetCode. On the one hand, these are practical examples that are somehow familiar to most technical specialists. On the other hand, everyone had time to try the solution with such small amounts of code.
Managers also visited the stand and admitted that they had not written code themselves for a long time, but were not against recalling tasks that in some way reminded many of their student days. And for those who found the examples too simple, the service developers helped them come up with a more complex task.

Alisa Kayfajyan

Kubernetes & DevTools Team

Sample problems

In the Perrin sequence, each number is the sum of the two numbers preceding it in the second and third positions. The first three numbers of the sequence are 3, 0, 2. The sequence looks like this: 3, 0, 2, 3, 2, 5, 5, 7,….
Write a function that will take the index of a number in the sequence (n) and return the Perrin number under this index – P(n).
Hamming distance is the number of positions at which the corresponding characters of two words of the same length are different.
For example, in the line “ABCB” there is a letter “B” in the fourth position, and in the line “ABCD” in the same position there is a letter “D”. The Hamming distance between these lines is 1.
Create a function that takes two strings and calculates the Hamming distance between them.

BareMetal was not brought to the stand, but the most interesting scenarios were demonstrated in the report.

ML track: summarization of reports from YandexGPT and generative Dungeons & Dragons. A large language model in conjunction with speech technologies Yandex SpeechKit was used on the site to prepare summaries of reports with main ideas.

But if the idea of neurosummary is already familiar to many, then for a role-playing game generative AI was used for the second time (the first time was in Practical ML Conf). At the interactive Dungeons & Dragons style, YandexGPT acted as the generative dungeon master, YandexART created the atmospheric visual accompaniment, and Yandex SpeechKit voiced the characters’ lines. To realize this, YandexGPT was additionally trained to be a game master: the game required answers in a certain style and taking into account the context of the world (or setting) in which the players found themselves.

We taught the model to write answers in a given format, which limited the length of the answer and returned fragments of text, for example, quotes from characters in the plot, with a certain markup. The markup is needed in order to voice NPC quotes in Yandex SpeechKit. Synthesis in SpeechKit allowed us to bring the game to life: minor characters spoke in different voices, each with their own emotional coloring, and in some situations we used whispers.

Marina Chaiko

Foundation Models Team

Proms for YandexART were also created using YandexGPT. Promt summarized the last three situations that players found themselves in during the story. So, if the action took place in the forest, YandexART took this into account, even if the last message from the generative master did not contain such information.

Players appreciated the possibility of live generation of images, and therefore even specially created some situations in order to see them on the screen with a picture. So, at one moment we were flying on an airship and admiring the landscapes of the world of techno-magic of Eberron from a bird's eye view, and at another we witnessed a battle where one of the players decided to turn into a giant spider beaver.

Not only business: services for social projects and HR

In the Yandex Scale program, users of the cloud platform themselves traditionally talk about service use cases. This time, the Scale Talks and thematic tracks were broadcast by representatives of KAMAZ Digital, Lamoda, Mindbox, Raiffeisen Bank, the First Bit company, and the Scandinavia network of clinics.

Also this year, the Yandex Cloud team announced several socially significant projects in healthcare, ecology and applied science, and a dedicated Technology Center for Society was created for their implementation. In the lounge area you could get acquainted with one of these projects, which was launched jointly with the Roszapovedtsentr – a photobank of the nature reserve system.

For the photo bank, Yandex Cloud provides data storage and automatic recognition of objects in photos using neural networks.

In the play area with ~~blackjack~~ lotto, tech talks and puzzles we talked about the technologies, tools and processes that Yandex Cloud employees themselves use. It was possible to discuss how secure non-relational DBMSs and public clouds are, or even design a service for asynchronously executing requests to YandexGPT / YandexART / SpeechKit from the proposed set of services.

The headliners of the evening IT stand-up were also the Yandex Cloud team, which during its preparation managed to master humor technologies – but that’s a completely different story.