人工智能的未来：初创企业视角

Table ofcontents 02Foreword02 10Advice for founders08 15What’s next in AI: Perspectives of industry leaders14 Amin VahdatVP/GM ML, Systems, and Cloud AI, Google Cloud James TromansManaging Director, Web3, Google CloudJennifer LiGeneral Partner, a16zJerry ChenPartner, GreylockJiaLiCo-Founder, President and Chief AI Officer, LiveX AIJill Greenberg ChaseInvestment Partner, CapitalGMatthieu RouifCo-Founder and CEO, PhotoroomMayada GonimahCTO and Co-Founder, Thread AIRaviraj JainPartner, LightspeedSalim TejaPartner, Radical Ventures, and Board Member,Aspect Biosystems, Promise Robotics, Intrepid LabsSarah GuoFounder and Partner, ConvictionMike VernalPartner, ConvictionYoav ShohamProfessor Emeritus of Computer Science,Stanford University, and Co-Founder, AI21 Labs Apoorv AgrawalPartner, Altimeter Capital Arvind JainFounder and CEO, Glean Chamath PalihapitiyaFounder and CEO, Social Capital,and Co-Founder and CEO, 8090 Crystal HuangGeneral Partner, GV David FriedbergCEO, Ohalo Genetics Douwe KielaCEO, Contextual AI Dylan FoxFounder and CEO, AssemblyAI Edo LibertyFounder and CEO, Pinecone Foreword AI is transforming every organization aroundthe world and represents an unprecedentedopportunity to solve complex problems, drivegrowth, create efficiencies, and open up newbusiness opportunities. This is particularly truefor startups, who are moving very quickly toaddress new market opportunities with AI.Google Cloud is at the center of AI innovation,and we’re proud of our technology leadershipthat continues to push the boundaries ofwhat’s possible for our customers, includingmore than 60 percent of all funded generativeAI startups globally. We are excited to partnerwith startups, the venture capital community,and industry leaders to ensure that foundersand their teams have access to the technologythat will help them redefine the future. Thomas KurianCEO, Google Cloud Google is building all the componentsof the AI technology stack, from custom chips,to data centers to frontier models. As a result,our new Gemini 2.0 models are more capable,faster and more efficient than previous versions.These models are natively multimodal—theyare able to process text, images, audio andvideo. They can also generate images andsteerable text-to-speech audio. With longcontext windows of up to 2 millions tokens,Gemini can power advanced applicationsthat require deep understanding and memory. Additionally, Thinking model is capable ofshowing reasoning skills for solving complexproblems, which is especially useful in mathand science. Gemini can also natively usetools like Google Search to access real-timeinformation, and DeepMind’s Project Marinerhas demonstrated that agents built with theGemini model can complete tasks using a webbrowser. Conversational experiences can nowbe built with the Gemini Multimodal Live API,which accepts audio and video streaming input.The combination of these capabilities enablesa new class of agentic experiences andwe’re excited to see what startups buildwith Gemini in 2025. David ThackerVP, Product, Google DeepMind AIpredictions Amin Vahdat VP/GM ML, Systems, and Cloud AI, Google Cloud Tight synchronization and massivecompute requirements will push infrastructureto never-seen-before levels of computedensity and capability. Arvind Jain Apoorv AgrawalPartner, Altimeter Capital Founder and CEO, Glean By combining voice, vision, and naturallanguage, multimodal AI will reducethe need for devices like computersand cell phones and make interactingwith the digital world more seamless. AI will continue to be a tool to augment humancapabilities, not replace them. The conceptof AI-based employees perpetuates a limitedperspective that hinders the true potentialof both AI and human intelligence. Chamath PalihapitiyaFounder and CEO, Social Capital, Crystal HuangGeneral Partner, GV and Co-Founder and CEO, 8090 Micro-booms and busts in AI willbe inevitable as the tooling forbuilding generative AI applicationsbecomes more readily availableand therefore commoditized. The future of software is about doing morewith less. With AI automation, the softwareindustry will get more efficient and the averageprofit margin of the S&P 500 will double ascompanies get more and pay less. Douwe KielaCEO, Contextual AI David FriedbergCEO, Ohalo Genetics Long context and RAG willconverge. Models may learnhow to decide when to uselong context and when touse RAG to optimize bothaccuracy and efficiency. The industries I think are most susceptibleto AI-driven transformation are media, SaaSand biology (therapeutic drugs and agriculture).For example, genome language models willbe able to predict the exact DNA sequenceneeded for any desired plant trait or biologicdrug, revolutionizing agriculture and human health. Dylan FoxFounder and CEO, AssemblyAI Edo LibertyFounder and CEO,Pinecone The timeline for widespread enterprise adoptionof AI will be slower than people think. A lot of last-mile issues need to be solved that