Coding/theory: majority vote, ConversationHistory and LLM theory

Выводы и как готовиться

• The coding bar was practical Python rather than algorithmic tricks: grouping, counters, state cleanup and edge-case reasoning around tool calls.
• The second task checked production-readiness: clear types, data structure choice, invariants and paired deletion of tool calls/tool results.
• The theory block focused on current LLM systems: decoder-only Transformers, pretraining/SFT/alignment, MoE placement and long-context trade-offs.