2025年的人工智能：25个模因中的25个主题

AI in 2025:25 themesin 25 memes January 30, 2025 Adrian CoxThematic Strategist(+44)-20-7541-7775adrian.cox@db.com Galina PozdnyakovaResearch Analyst(+44)-20-7547-4994galina.pozdnyakova@db.com A picture is worth 1,000 words If a picture is worth 1,000 words, this chartbook should save you fromreading 25,000 of them. That counts for something in a week when so many millions of words havebeen written about the surprise arrival of China’sDeepSeekAI model. AI has come of age in the era of the meme–and it turns out memes areone of the best ways of explaining where it is going. This chartbook has five sections with five key themes each:1.Technology: moving beyond large language models (LLMs) 2.The business of AI: the money magnet3.AI in enterprises: still multiple hurdles4.Jobs outlook: keeping up the good work5.The future: FOMO trumps safety fears AI is not just a popular subject for memes. It can also interpret and creatememes itself. We instantly generated the meme on this page usingimgflip’sChatGPT-based AI Meme Generator and the prompt “Create a meme about howgood AI is at creating memes”. Getting even more meta, we asked a competing AI chatbot to interpret it.Claude 3.5 Sonnet snapped back that the image was “playfully mockingboth groups by suggesting they’re equally amateur at making memes”. 1.A week is a long time in AI2.Bigger isn’t always better3.Agents are coming to the rescue… at a cost4.It’s back to the future on advertising5.Energy demand is surging 1.Technology: moving beyond LLMs 2.The business of AI: the money magnet 3.AI in enterprises: still multiple hurdles 4.Jobs outlook: keeping up the good work 5.The future: FOMO trumps safety fears •China’sDeepSeekopen-source small language model burst onto the scene this month,apparently trained at perhaps one tenth of the cost of comparable models. •It ignited concern about how much of the AI investment boom is justified, suggesting that:1.The latest hardware may no longer be essential for innovation.2.AI models for most uses will shrink and become commodities.3.Despite restricted access to chips, China is not lagging the US after all.For more, seeDeepSeekAI: Why drive a Tesla Model X when a BYD will do? •But a week is a long time in AI: Microsoft has now been reported to be investigating whethera group linked toDeepSeektook data from OpenAI without authorisation. And Alibabaunveiled its ownQwen2.5 Max model, saying it outperformsDeepSeekand Meta’s Llama. 1.2 Bigger isn’t always better •Critics say generative AI is “hitting a wall”, with diminishing returns for more compute.OpenAI and rivals have yet to significantly improve on GPT-4, launched in March 2023.•Instead the industry is focusing on squeezing more value out of models, often optimisingthe “inference” stage. For example, OpenAI’s o1 series takes time to reason step-by-step.•There is also a move to smaller models likeDeepSeekthat can run in phones, computersand cars without needing to connect to the cloud, making them faster, more robust, moresecure–and cheaper–for many applications than huge, generalist, multimodal models. 1.3 Agents are coming to the rescue–at a cost •AI agents are evolving from basic chatbots to autonomous systems that enablecomputers to perform tedious multi-step tasks 24/7, such as filing expense reports. Thatshould further increase the hyperscalers’ need to invest in computing power. •Major players have begun rolling out agents. OpenAI launched a preview ofOperatorlastweek and CEO Sam Altman said the first AI agents may “join the workforce” this year. •Microsoft’s Satya Nadella has said Software as a Service (SaaS) may be superseded asagents act directly with underlying data rather than through an interface like Excel. 1.4 It’s back to the future on advertising •Just as subscription models seemed to be the future, as popular for ChatGPT as fornews articles, advertising business models may be coming back in vogue. •AI-powered search engine Perplexity began experimenting with ads on its platformin November and even OpenAI has said it is now considering an ad model. •Meta is showing how spending on AI can benefit its advertising model. In the fourthquarter, sales rose by fifth, driven in part by an automated product promotion tool. •AI applications are leading to surging demand for energy. Whether theproliferation of smaller models likeDeepSeekaffects that remains to be seen. •Data centers consumed about 4.4% of total US electricity in 2023 and areexpected to consume as much as 12% by 2028, theDepartment of Energysaid. •Hyperscalers, led by Amazon’s AWS, Microsoft’s Azure, Alphabet’s GoogleCloud and Meta, are driving demand for renewable energy. They are also tryingto increase capacity by investing in small modular reactors, restarting oldreactors and betting on new technology such as fusion. AI in 2025: 25 themes in 25 memes 1.Technology: moving beyond LLMs 1.Nvidia has been the standout winner (for now) 2.The business of AI: the money mag