AI year in review: Trends shaping 2026

This special year-end episode of Mixture of Experts brings together a panel of veteran panelists to review the most significant AI developments of 2025 and offer predictions for 2026. The discussion covers four key areas: AI agents, open-source models, AI hardware, and multimodal AI.

Chris Hay: The Year of the Super Agent

Chris Hay defends his 2024 prediction that 2025 would be the “year of the super agent.” He argues that while the term isn’t widely used, the functionality is now deeply integrated into mainstream AI platforms.

• Super Agents Defined: Unlike specialized single-task agents, super agents combine reasoning and tool use to execute complex, multi-step tasks. Examples cited include Manis for presentation generation and advanced capabilities within ChatGPT and Claude that leverage external connectors and tools.
• Invisible Integration: Agentic features, such as ChatGPT’s Deep Research, have become so commonplace that users no longer perceive them as distinct “agents”; they are simply part of the AI’s core functionality.
• The Battle for the “Front Door”: A major competitive front has opened up over who will control the primary user interface for these powerful agents. This battle is being waged across chat windows, browsers (e.g., Perplexity Comet, Gemini in Chrome), and mobile operating systems (Siri, Android).
• 2026 Prediction: The focus will shift towards creating unified “agent control planes.” These dashboards will allow users to centrally manage, deploy, and monitor agents that can operate across various environments, from terminals to code editors, pointing to a future of more orchestrated and manageable agentic workflows.

Gabe Goodhart: Open Source’s Breakout Year

Gabe Goodhart posits that 2025 was truly the “year of open source,” marked by models like Kimi K2 Thinking achieving performance on par with, or even exceeding, proprietary counterparts.

• The “Packaging Problem”: The primary challenge for open source is not the quality of its individual components but the lack of a polished, standardized, and easy-to-deploy stack. Closed-source solutions excel by offering a seamless user experience that is difficult to replicate with disparate open-source parts.
• Orchestration is the Hurdle: The key difficulty lies in orchestrating various open-source models, guardrails, and tools into a cohesive system. This requires a level of standardization that can be at odds with the tinkering-focused open-source ethos.
• 2026 Prediction: The open-source community will make significant strides in solving the orchestration problem. We can expect the emergence of a high-quality, full-stack generalist open-source agent. The future will likely be a “two-sided world”: open source will provide the foundational components (akin to Linux), while polished, closed-source products will continue to thrive by offering superior user experiences built on top.

Kaoutar El Maghraoui: The AI Hardware Crisis and Future

Kaoutar El Maghraoui highlights that 2025 was defined by an AI hardware supply crisis, where demand for compute dramatically outpaced the supply chain’s capacity.

• Hardware Scarcity as a Strategy Driver: The lack of available compute became a structural constraint, forcing companies to design their business strategies around GPU availability rather than just model capabilities.
• A Diverging Hardware Landscape: The market split into two tracks: scale-up (massive superchips like NVIDIA’s B200 for frontier labs) and scale-out (efficient edge computing with small LLMs and NPUs).
• 2026 Predictions:
  • Efficiency Over Scale: The industry will pivot to “scale efficiency instead” of pure compute power. This will foster a divide between giant “frontier 2.0” models and highly optimized, hardware-aware efficient models.
  • Edge AI Becomes Reality: Running capable generative models (e.g., 10B parameters) locally on devices will become feasible, thanks to new NPUs and efficient architectures.
  • Specialized “Agent Processors”: A new class of chips may emerge, specifically designed for agentic workloads that require context persistence, low-latency reasoning, and multi-modal processing.

Aaron Baughman & Abraham Daniels: The Future is Modular and Multimodal

Aaron Baughman and Abraham Daniels discuss the evolution of multimodal AI, predicting a shift towards more capable and modular systems.

• Modular Multimodality: Abraham Daniels explains that the future lies not in single, monolithic “omni-models” but in orchestrating a network of specialized, best-in-class models and lightweight adapters (e.g., LoRA) for different modalities (image, audio, text). This approach offers greater flexibility and state-of-the-art performance for specific tasks.
• Rise of Digital Workers: Aaron Baughman envisions these multi-sensory models enabling autonomous “digital workers” capable of perceiving the world (seeing, hearing, reading) and performing complex tasks in fields like healthcare.
• Enterprise and Human-in-the-Loop: The focus for enterprise AI (like IBM’s Granite) will be on practical applications like document understanding (KVP extraction) and audio analysis, not text-to-image generation. Human-in-the-loop systems will be critical for fine-tuning models for specific, high-accuracy tasks, such as detailed analysis in sports.

• Tags:
• AI Agents
• Feature
• Multimodal AI
• Agentic Systems
• Modular AI
• Nvidia
• Open-Source AI
• AI Hardware
• AI Trends
• Super Agents
• 2026 Predictions