人工智能的未来：年中领导者、机遇与威胁回顾

A Midyear Review of Leaders,Opportunities, and Threats Sarah Hoffman, Director of Research, AI What’s Inside 1Introduction CHAPTER 1 2Model Advancements CHAPTER 2 5New & EnhancedCapabilities CHAPTER 3 8Strategic InfrastructureInvestments CHAPTER 4 10Enterprise Impact CHAPTER 5 12Expanding Real-WorldApplications CHAPTER 6 13Emerging Risks & Challenges CONCLUSION 15Where We’re Headed alphasense.ai Introduction The generative AI landscape has undergonesignificant transformation in the first half of 2025,with established players solidifying their positionsand new entrants disrupting the status quo. Theperiod has been marked by rapid innovation in modelarchitectures, particularly in the realm of reasoningcapabilities, alongside infrastructure investmentsand shifting competitive dynamics. Strategic bets onagentic systems are beginning to reshape how genAIis applied across industries, and a wave of acquisitionsis intensifying the race to build and own more of thegenAI stack. CHAPTER1 ModelAdvancements Several model releases during the first half of 2025 have reshapedthe competitive landscape by introducing more reasoning andmultimodal capabilities. OpenAI In March 2025, OpenAI released a new text-to-image generator thatquickly impressed users with its ability to accurately follow complexprompts and generate high-quality visuals. Also in March, OpenAIclosed a $40 billion funding round, the most ever raised by a privatetech company. However, not everything went as planned for OpenAI in the first halfof 2025. In late February, OpenAI launched its Orion model as GPT-4.5, instead of the expected GPT 5. In April, OpenAI rolled back aGPT-4o update due to its “sycophant-y and annoying” personality. Google In March 2025, Google released Gemini 2.5 Pro, which stackedup favorably against other models on coding, math, and reasoningbenchmarks. Like other models, Gemini 2.5 Pro can consult the web,but it also contains a recent snapshot of the world’s knowledge: Itstraining data cuts off at the end of January 2025. In April, Google introduced an early preview of its Gemini 2.5 Flashhybrid reasoning model. Gemini 2.5 Flash is Google’s first fullyhybrid reasoning model, giving developers the ability to togglethinking on or off. In May, Google released its newest AI video generator, Veo 3, whichhas been amazing viewers with its realism. It can include dialogue,soundtracks, and sound effects and is available to subscribers ofGoogle’s $249 per month AI Ultra plan. Anthropic Anthropic introduced Claude 3.7 Sonnet in February 2025, markingthe industry’s first hybrid reasoning model. This innovation allowsusers to select between standard or reasoning responses for anyquery, potentially setting a new standard for model flexibility. In May,Anthropic debuted the first models of its Claude 4 series, includingClaude 4 Opus, which Anthropic says is best at coding. Anthropic’s hybridreasoning modelallows users to selectbetween standard orreasoning responsesfor any query,potentially settinga new standard formodel flexibility. However, according to a May report, significant market share shiftsoccurred between January and May, including a 10% decline inAnthropic’s Claude models’ queries. DeepSeek In January 2025, Chinese AI startup DeepSeek made headlines withits open-source R1 large language model. The model performedcomparably to leading LLMs on industry benchmarks while beingtrained at under $6 million, a fraction of the cost of similar models. However, a report released in May showed DeepSeek’s R1 querieshad declined from 7% in mid-February to 3% by the end of April. Microsoft In the first half of 2025, Microsoft solidified its leadership ingenAI. The Copilot suite has evolved into a central component ofMicrosoft’s AI strategy. Recent updates introduced features such asmemory capabilities for personalized experiencesand the “Actions”function, enabling Copilot to autonomously complete multi-step tasks like booking travel or managing schedules. Microsoftalso introduced AI agents like “Researcher” and “Analyst” withinMicrosoft 365, designed to assist with tasks such as data analysisand report generation. In early May, Microsoft announced the release of Phi-4-reasoning-plus, an open-weight language model built for tasks requiring deep,structured reasoning. Despite its relatively modest size,Phi-4-reasoning-plus outperformed larger open-weight models, such asDeepSeek-R1-Distill-70B on a number of benchmarks. Later in May, Microsoft struck a deal with Anthropic to use Anthropic’snew Claude 4 models to power Microsoft’s own AI agent features,reflecting Microsoft’s willingness to branch out from OpenAI. xAI In February 2025,xAI’s Grok version 3 was unveiled. Similar to othernew models, Grok 3 contains advanced reasoning abilities. In April, xAI also expanded Grok’s capabilities by integratingmultimodal features, including real-time visual analysis throughsmartphone cameras. Grok Vision lets users point thei