Recorded in Las Vegas with shuki-cohen (VP Data at AI21, YouTube channel One Shot Learning), looking back at 2025 and ahead to 2026 across three layers. Models: omni-modal models (Amazon Nova), often at the cost of reasoning; AWS Nova Forge, letting companies continue training from early checkpoints on their own data mix (like Thinking Machines’ Tinker); niche expert models (latency, long context, task-specific); small encoder-only models for classification; and per-intent compute (“model choice”) as the market starts caring about ROI. Infra: speech-to-speech needs async reasoning delegation, async tool calling (Nova Sonic 2 keeps talking while tools run), and voice-specific evals; RL environments and reward endpoints; AWS AgentCore (memory, code runtime, browser, identity, policy); and MCP gateways. Apps: orchestrators return. Multi-agent systems bring context isolation but also overlapping agents and stale capability descriptions. Autonomous agents are narrowing to niches. Wishes for 2026: streaming omni-modal reasoning, and real memory management.
Show: langtalks
Concepts: multi-agent-orchestration · small-language-models · voice-agents · llm-evals · ai-gateway · model-selection