LAM (Large Action Models)

Main article: Artificial Intelligence

LLM (Large Language Models)

Main article: LLM (Large Language Models)

2025: How Russia is developing big action models - TA opinions

Artificial intelligence is rapidly evolving: if earlier language models (LLM) only gave advice and explained how to complete a particular task, now large action models (LAM) are coming to the stage - a fundamentally new type of AI that can not just reason, but independently act. In June 2025, TAdviser spoke with experts and market participants to find out how things are with LAM in Russia.

According to Sergey Nazarenko, head of the department for the development of data management and machine learning technologies at Reksoft, the development of LAM is at the stage of active experiments and local implementations, but there are already examples of successful projects in the Russian Federation:

𝓲

Фото: Freepik

• Sber develops AI agents for business processes and customer support;
• Yandex is testing Alice with auto-action functions in the browser;
• "MTS launched MWS Agent, a multi-agent platform for corporate tasks;
• pilot projects are being implemented in industry (Vintix model from AIRI for robot management) and the municipal industry (AI agents "Polyplastic "/MIPT).

Also, Solar is gradually starting to use the capabilities of LAM in information security products. According to TAdviser, the head of the R&D laboratory of the Center for Cybersecurity Technologies of the Solar Group Maxim Buzinov, by mid-2025, several promising prototypes have already been created. Thus, existing models can not only analyze information about threats and attacks, but also independently launch diagnostics of suspicious actions, for example, use existing algorithms for detecting anomalies and incidents as their tools. And then - such LAM systems can independently take measures to neutralize the threat or attack. For example, block dangerous user actions, Buzinov said. He added that LAM systems can relieve a significant load from information security employees for routine work. For example, preparing reports, configuring information security policies in information protection tools, etc.

ML-architect Softline Digital Dmitry Zboroshenko believes that by mid-2025 LAM is rather another clickbait designation. We are not talking about fundamentally new types of neural networks or architectures - these are essentially services and applications built on the basis of existing LLMs.

That is why it is premature to talk about LAM as a technologically independent breakthrough. This is an instrumental level, not a scientific revolution. If we talk about the development of such services, then the priority is the request from business, industry. Now such a request, in my opinion, is not observed, - said Zboroshenko in a conversation with TAdviser.

According to Sergei Nazarenko from Reksoft, the development of large action models is hampered by a shortage of modern GPUs (graphics accelerators), high implementation costs, a lack of qualified personnel and restrictions on access to foreign technologies due to sanctions. In addition, the business is still cautious about investing in LAM due to the complex payback model, the expert believes.

Alexander Borisov, head of Data Science at RAMAX Group, notes that alternative technologies that compete with LAM or complement it, but do not have such barriers when implemented. To such alternatives, the interlocutor of TAdviser attributed:

• RPA. The technology allows you to automate user actions in interfaces of various software, but strictly set the sequence of actions. LAMs will apply instead of RPAs if there is sufficient variability in these actions, and the number of instances of the business process is comparable to those where RPAs already apply
• Process Mining и Task Mining. They allow you to automate research of business processes by their digital traces, and find processes suitable for LAM.

Kirill Smelovets, chief architect of artificial intelligence and machine learning services Cloud X, sees one of the problems for the development of LAM in Russia in the need to adapt third-party systems: they must provide structured APIs so that agents invoke functions without human input. Another difficulty is that the so-called sandboxes are critical - specialized isolated environments with access to tools and APIs in which LAM can safely and efficiently interact with external services, the expert said.