从愿景到实效：2026年人工智能演进路径研究报告

December 2025 1IntroductionCHAPTER 122025: The Technology MaturesCHAPTER 252025: The World Responds Introduction At the end of 2024, we reflected on it as the yeargenerative AI bridged imagination and implementation.In 2025, we crossed that bridge. It was the year of genAIimplementation, marked by a wave of model launches, Beyond the excitement of new features, generativeAI also moved from a collection of isolated tools to anintegrated layer within enterprise systems, consumerplatforms, and everyday workflows. Businesses began At the same time, the year revealed the limits of scale.The release of GPT-5 continued to show that ever-largermodels now yield diminishing returns. OpenAI’s trillion- As 2026 approaches, AI stands at a new inflection point,one defined by measurable impact. The focus will shift CHAPTER 12025: The TechnologyMatures 2025 was the year genAI became more than a collection of tools.Models stabilized, search became conversational, and earlyagentic systems took shape. Long-term memory and reasoning Then came November. For the first time in years, Google pulledahead in the AI race, with the launch of Gemini 3, a model that Across the ecosystem, the defining breakthroughs were notnecessarily the flashiest, but, instead, they were the quietadvances that made AI more usable and persistent. By the end of The Foundation is Set for AI Agents AI agents were widely expected to go mainstream in 2025. Instead,the year was defined by a surge of platform launches, expandinguse cases, and cautious enterprise adoption. One of the biggestsuccesses was the rise of deep research tools, introduced by While nearly every major AI company introduced agent-readyframeworks, most deployments remain in pilot mode. Expertshighlight hurdles around ROI, accuracy, trust, and governance, Long-Term Memory Takes Off In 2025, we also saw some interesting new enhancements,leading to newer, more creative ways to take advantage of AI. One 2025 was the year long-term memory was incorporated into AIsystems. OpenAI, xAI, Anthropic, Microsoft, and Google eachlaunched persistent memory capabilities, enabling their models to Until recently, each AI prompt was a fresh start. With long-termmemory in place, AI can move beyond one-off responses towardcontinuity and true personalization. Models can remember a user’sgoals and prior conversations, tailoring outputs to that specific Search Becomes Generative 2025 transformed search into a conversational experience.The major players moved quickly to make AI the primary In May 2025, Google introduced AI Mode, which providescomprehensive AI-generated answers to complex questions. In October, OpenAI launched ChatGPT Atlas, a browser withChatGPT built in, allowing users to navigate, summarize, andtake actions across websites in real time. Perplexity launched its GenAI Matures into Enterprise Systems After years of experimentation, companies began embeddinggenerative AI across departments. By the second half of 2025,the pace of enterprise innovation accelerated, with several major In October, building on previous efforts to bring AI capabilities tobusinesses, Google launched Gemini Enterprise, a comprehensiveAI platform for businesses. Gemini Enterprise is designed to letbusinesses securely create, share, and use AI agents for a variety That same month, OpenAI released ChatGPT company knowledge,giving organizations a way to safely connect internal data toChatGPT. It’s powered by a version of GPT-5 that directly connects Also in October, Anthropic announced Skills for Claude, whichis made up of “folders that include instructions, scripts, andresources that Claude can load when needed to make it smarter at In September, AlphaSense added Workflow Agents to its platform.Built on its Deep Research capabilities, Workflow Agents aredesigned around specific jobs, supporting tasks like M&A idea CHAPTER 2 2025: The World As AI’s capabilities stabilized, the ecosystem around it adjusted.In 2025, we saw massive infrastructure investments, shiftingregulations, and heightened awareness of AI’s risks and rewards. Big Tech companies doubled down on compute andpartnerships, with OpenAI betting trillions on the infrastructureneeded to sustain frontier models. Policymakers scrambledto keep pace, introducing or revising AI legislation across the The response wasn’t limited to governments or boardrooms.Within organizations, a new focus emerged around buildingAI-ready workforces: employees equipped to use AI tools This wasn’t a period of restraint or pragmatism. 2025 wasdefined by acceleration as the world tried to catch up with OpenAI’s Massive Infrastructure OpenAI has committed to spending more than$1.4 trillionwithpartners over the next five years. Some key deals are with: Oracle.In September, OpenAI signed a 5-year$300 billiondealwhere OpenAI will use Oracle’s infrastructure to power its AI models. Microsoft.OpenAI agreed to buy$250 billionof Microsoft’s AMD.In October, OpenAI