人工智能的未来：2025年上半年行业发展回顾（英）

alphasense.ai IThe Future of AIA Midyear Review of Leaders, Opportunities, and ThreatsSarah Hoffman, Director of Research, AIalphasense.aiWhat’s Inside1 Introduction CHAPTER 12 Model Advancements CHAPTER 2 5 New & Enhanced Capabilities CHAPTER 3 8 Strategic Infrastructure Investments CHAPTER 410 Enterprise Impact CHAPTER 512 Expanding Real-World Applications CHAPTER 613 Emerging Risks & Challenges CONCLUSION 15 Where We’re Headed alphasense.aiIntroductionThe generative AI landscape has undergone significant transformation in the first half of 2025, with established players solidifying their positions and new entrants disrupting the status quo. The period has been marked by rapid innovation in model architectures, particularly in the realm of reasoning capabilities, alongside infrastructure investments and shifting competitive dynamics. Strategic bets on agentic systems are beginning to reshape how genAI is applied across industries, and a wave of acquisitions is intensifying the race to build and own more of the genAI stack.alphasense.ai 1alphasense.ai 2CHAPTER 1Model AdvancementsSeveral model releases during the first half of 2025 have reshaped the competitive landscape by introducing more reasoning and multimodal capabilities.OpenAIIn March 2025, OpenAI released a new text-to-image generator that quickly impressed users with its ability to accurately follow complex prompts and generate high-quality visuals. Also in March, OpenAI closed a $40 billion funding round, the most ever raised by a private tech company.However, not everything went as planned for OpenAI in the first half of 2025. In late February, OpenAI launched its Orion model as GPT-4.5, instead of the expected GPT 5. In April, OpenAI rolled back a GPT-4o update due to its “sycophant-y and annoying” personality. GoogleIn March 2025, Google released Gemini 2.5 Pro, which stacked up favorably against other models on coding, math, and reasoning benchmarks. Like other models, Gemini 2.5 Pro can consult the web, but it also contains a recent snapshot of the world’s knowledge: Its training data cuts off at the end of January 2025.alphasense.ai 3In April, Google introduced an early preview of its Gemini 2.5 Flash hybrid reasoning model. Gemini 2.5 Flash is Google’s first fully hybrid reasoning model, giving developers the ability to toggle thinking on or off. In May, Google released its newest AI video generator, Veo 3, which has been amazing viewers with its realism. It can include dialogue, soundtracks, and sound effects and is available to subscribers of Google’s $249 per month AI Ultra plan. AnthropicAnthropic introduced Claude 3.7 Sonnet in February 2025, marking the industry’s first hybrid reasoning model. This innovation allows users to select between standard or reasoning responses for any query, potentially setting a new standard for model flexibility. In May, Anthropic debuted the first models of its Claude 4 series, including Claude 4 Opu