NFPrompt, an AI-driven Web3 platform for digital creativity, has partnered with the international tech firm Alibaba’s Wen and Qwen projects. The partnership focuses on improving consumer-generated content as well as AI-led creativity across the Web3 landscape. As per NFPrompt’s official X

Meta CEO Mark Zuckerberg has instructed the company to abandon the long-term strategy of "full open source" artificial intelligence, and the most advanced AI models in the future will adopt a paid model. According to multiple insiders, Meta's next-generation large language model (internal code name "Avocado") is being fine-tuned based on Alibaba's Qwen series open-source models. Zuckerberg explicitly stated that Meta must prioritize "super intelligence" rather than insisting on complete open source.

According to the latest market news, Meta CEO Mark Zuckerberg has officially directed the company to abandon the long-term strategy of "full open-source" artificial intelligence, with the most advanced AI models in the future adopting a paid model rather than being publicly available for free. (Previous background: Is Zuckerberg giving up? Meta plans to cut 30% of metaverse budget, lay off employees as early as January next year, META rises 4%) (Additional background: Meta Zuckerberg is busy! Paying over 100 million USD in salaries, three AI geniuses leaving in two months) According to the latest market news, Meta CEO Mark Zuckerberg has officially directed the company to abandon the long-term strategy of "full open-source" artificial intelligence, with the most advanced AI models in the future adopting a paid model rather than being publicly available for free.

Deep Tide TechFlow News: On December 10, after months of investing hundreds of billions of dollars to build the most expensive team in tech history, Meta CEO Mark Zuckerberg is deeply involved in daily R&D and is pushing the company's strategic shift towards AI models that can be directly monetized. According to insiders, a new model codenamed "Avocado" is expected to be released in spring 2026 and may be launched in a closed-source manner (i.e., strictly controlled by Meta with restricted access for sale). This move marks a significant departure from Meta's long-standing advocacy for open-source initiatives. Zuckerberg is dedicating a large amount of time to a core team called TBD Lab, which, during the training of Avocado, even integrated third-party models including Google Gemma, OpenAI gpt-oss, and Alibaba Qwen. Meanwhile, Meta is drastically adjusting resource allocation, cutting

Aster brought together 70 human traders and 30 AI models for a trading competition. Each participant started with $10,000 in capital, and the prize pool totaled $200,000. The competition ran from December 9 to December 23 and was broadcast live on the Aster event page. Currently, the top performer is the well-known trader Ao Ying Capital, who tripled their position within one day. Statistics show that human traders consistently rank in the top 20, but also make up the bottom 20 performers. AI models are concentrated in the middle and lower-middle tiers. Regarding AI traders, currently, Alibaba's Qwen (, the very conservative version ), leads with a profit of $618. The more aggressive versions of Ernie 4.5, Gemini 3, and Claude Sonnet 4.5 are among the worst-performing models.

OpenRouter's "2025 AI Status Report" shows that AI is entering the era of agent-based reasoning, where users no longer see AI merely as chatbots, but as tools capable of solving complex problems. Open-source models from China, such as Qwen, are rapidly rising due to fast iteration and less content filtering, becoming a major force in the global market. Data shows that role-playing has become a primary use case for AI, indicating users’ preference for open-source models and a deep emotional connection with them.

China's local ChatGPT competitor, Alibaba's Qwen AI, holds a cautious attitude towards the recent prospects of Ripple (XRP), Cardano (ADA), and DOGE. However, Qwen AI's predictive model indicates that in an optimistic scenario, XRP could rise to $10, ADA might reach $8 by early 2026, and DOGE even has the potential to hit $2.

Original Title: "The New AI Trading Competition is Coming, and the US Stock Market is Letting Gemini 3 Pro Make a Strong Push" Original author: David Original source: Reprint: Mars Finance Do you remember the AI coin trading competition in October? Alibaba's Qwen 3 Max earned 22% in two weeks, while OpenAI's GPT-5 lost 63%. Six of the world's top large models, each with ten thousand dollars battling in the crypto market, produced a clear distinction between winners and losers: The Chinese model stands tall while the American model faces total collapse. (Editor's note: For more details, please refer to: "The Six Major AI 'Traders' Ten-Day Showdown: Who Can Survive in a 'No Information Asymmetry' Market?", "Alibaba Qwen3 Surpasses DeepSeek at the Last Moment, Ultimate Comeback in the AI Cryptocurrency Trading Competition") Last

Author: David, Deep Tide TechFlow Do you remember the AI trading competition in October? Alibaba's Qwen 3 Max earned 22% in two weeks, while OpenAI's GPT-5 lost 63%. Six of the world's top large models, each armed with ten thousand dollars in the cryptocurrency market, resulted in a clear distinction between the two sides: The Chinese model stands tall in the world, while the American model is in total collapse. The previous season has just concluded, and the new Alpha Arena Season 1.5 is here.

Author: Yao & Thinking & Pds Editor: 77 Background With the rise of AI large model real-time trading competitions, more and more crypto communities and developers are beginning to experiment with AI-driven automated trading, and many open-source solutions are being rapidly deployed. However, these projects are not without security risks. NOFX AI is an open-source cryptocurrency futures automated trading system based on DeepSeek/Qwen AI, supporting exchanges such as Binance, Hyperliquid, and Aster DEX. The SlowMist security team received initial intelligence from @Endlessss20, suspecting that this system may lead to exchange API Key issues.

According to Huoxing Finance, the Slow Mist security team recently analyzed the open-source automated futures trading system NOFX AI based on DeepSeek/Qwen and discovered multiple serious verification vulnerabilities. It pointed out that the system has a "zero verification" mode under default configuration, with administrator mode enabled directly, allowing all requests to pass without verification. Attackers can access /api/exchanges and obtain the complete API key and Private Key. Although JWT is added in the "authorization required" mode, the default jwt\_secret still exists, and if the environment variable is not set, it will revert to the default Secret Key. Moreover, sensitive fields are still output in raw JSON in this mode, and if the token is forged or stolen, it can also lead to Secret Key leakage. Slow Mist stated that as of now, it has identified over a thousand publicly deployed instances using vulnerable configurations and has communicated with Binance and

Alibaba's cloud computing arm, Alibaba Cloud, has officially launched the all-in-one AI application "Qwen (", which emphasizes being completely free and handling both life and work in one go. This app not only challenges local AI startups in China but also directly competes with the subscription-based business models led by OpenAI and Anthropic. Experts point out that Alibaba's strategy of offering free services in exchange for traffic and feedback is expected to significantly enhance the evolution speed of its AI models. Alibaba creates an all-in-one free AI: 1k questions officially launched Alibaba Group announced that its comprehensive AI application "Qianwen" has officially launched in beta, providing both web and PC versions. It is powered by Alibaba Cloud's open-source model "Tongyi Qianwen" and is positioned as a personal AI assistant that can do anything. k問 is expected to gradually integrate services such as communication chat, map navigation, delivery, ticket booking, shopping, and learning, aiming to become a part of users' daily lives.

NOFX AI, an open-source automated trading system built on DeepSeek/Qwen AI. It is facing a serious security crisis after SlowMist uncovered vulnerabilities. That could expose exchange API keys and private keys. The issue affects users across major exchanges, including Binance, Hyperliquid and

According to Wu, Jay A, the founder of nof1, tweeted that the Alpha Arena Season 1 competition officially ended this morning, with Qwen 3 MAX overtaking Deepseek Chat V3.1 at the last moment, becoming the AI model with the highest final profit. The next season (Season 1.5) will test various prompts and model instances in parallel, and a new round of competition is expected to begin soon.

At the beginning of November, the heavyweight themed event "Hunting AI Project: The Battle for Human Dignity" initiated by Huobi HTX officially came to an end. Over the course of three peak duels, top traders from both the human and AI camps engaged in a battle of wits and strategies on the same stage. With the final battle report announced, this "Human VS Algorithm" century battle reached its climax and conclusion. First Match: AI narrowly wins, humans make a strong appearance The first match of the event ignited the attention of the entire audience. The AI player gained an advantage first thanks to its precise strategy model, while the human representative "Big Brother" performed brilliantly, winning the single match championship with a record of 530.72 USDT. The AI players ChatGPT, Qwen, and DeepSeek followed closely behind, each with account balances exceeding 490 USDT. Two out of the top three spots were occupied by AI competitors, declaring the victory of the AI camp. This battle marks that "humans still have the strength to fight back," and the number

Alibaba's ChatGPT's Qwen3-MAX AI predicts that SOL could climb to between 800 and 1200 USD by the end of the year, XRP may rise to between 8 and 12 USD, and ADA could reach up to 5 USD in November, indicating that strong alternative Crypto Assets performance will be a feature of the new bull run.

As the market begins to recover, funds are flowing back from gold to risk assets, and XRP has once again become the focus of retail investors. Qwen AI, which performed outstandingly in the Alpha Arena trading competition, has given a very optimistic forecast for XRP's trend over the next three months, setting a target range of $3.80 to $4.20. This prediction is based on XRP's strong technical support, the upcoming ETF decision window, and macro favourable information such as improved regulatory clarity under Trump and the seasonal market surge in the fourth quarter. Qwen AI believes that if it can cleanly break through the key resistance level of $2.85, XRP is expected to achieve a rise of 45% to 60% before January 2026.

As the market shifts from gold to a recovery in risk assets, XRP remains a favorite among retail investors. After Qwen AI, a subsidiary of Alibaba, performed excellently in the Alpha Arena trading competition, its price prediction for XRP has garnered attention. Qwen AI predicts that XRP will pump by 45% to 60% by early 2026, with a target price range of $3.80 to $4.20.

AI Qwen, Alibaba's flagship large language model (LLM) series, is rapidly gaining traction as a versatile open-source, driving innovation in artificial intelligence applications.

Alibaba's AI Lab Qwen Deep Research has launched a significant update that allows complex searches to be instantly converted into web pages and podcasts. This new feature makes Qwen AI a powerful and free alternative to solutions like ChatGPT, DeepSeek, and Claude. In tests, Qwen AI surpassed its competitors in both depth and accuracy.

Alibaba's flagship model Qwen AI increased its initial capital from $10,000 to $20,776 in just 9 days, surpassing Western models. Now, Qwen AI has released a price prediction that XRP could soar to $6.50 in November, DOGE could reach $1.20, and SOL could rise 133% to $700.

Flock.io signed a memorandum with China Daren International to include Flock in its strategic Token reserves, expanding privacy AI capabilities to the consumer manufacturing and clothing Supply Chain, facilitating the collaboration between digital assets and physical businesses. Previously, Flock had established partnerships with the United Nations Development Programme and Alibaba Qwen.

The market overall continues to experience a pullback, and the AI zone has not been spared. Major trading tokens such as $VIRTUAL and $TAO have seen slight declines, while $WAI has performed well. MyShell has launched a hackathon event to boost participation in creation. AriaAI will be launched on Binance Alpha to promote AI game development. At the same time, DeepSeek, NVIDIA, and Qwen have released new models, accelerating technological iteration.

In Brief Alibaba Cloud’s Qwen team has launched Qwen-Image-Edit, a state-of-the-art image editing model that combines semantic and appearance editing with precise bilingual text modification, delivering advanced capabilities for creative and practical applications. Alibaba Cloud’s Qwen team has

The DeAI training platform Flock has partnered with Alibaba Cloud's Qwen large language model, marking the first integration cooperation of Web2 AI actively moving towards Web3 AI. Web3 AI needs to complement Web2 AI to address issues such as Computing Power and data privacy. Flock addresses privacy issues with AI federated learning and distributed technology, while Qwen offers a high-performance, locally deployable large language model. Their collaboration is significant, demonstrating the potential for cooperation across different AI fields and providing direction for the advancement of Web3 AI.

The DeAI training platform in the Web3AI field @flock_io has announced a partnership with Alibaba Cloud's @Alibaba_Qwen, breaking the precedent for collaboration between Web2AI and Web3AI. The collaboration means that Web3AI needs to find complementary ecological niches with Web2AI to address the issues of centralized AI. Specific collaborations include AI arenas, federated learning alliances, and lunar bases. The significance of this collaboration is profound, as it is expected to deepen the linkage between Web3AI and Web2AI, and inject new momentum into the development of Web3AI.

Author: Haotian Yesterday, the DeAi training platform Flock.io in the Web3AI field officially announced a partnership with Qwen, a large language model under Alibaba Cloud. If I remember correctly, this should be considered the first proactive integration cooperation from web2 AI to web3 AI. It not only allows Flock to truly break out of its own circle, but also revitalizes the morale of the web3AI track under the pressure of a sluggish market. Let me elaborate on this: 1) I have explained in the pinned post that the previous web3 AI Agent has been trying to stimulate the landing of Agent applications through Tokenomics, and also engaged in rapid deployment as a competitive paradigm. However, after a wave of Fomo in asset issuance, everyone realized that when it comes to practicality, innovation, etc., web3 AI has almost no chance of winning compared to web2 AI. So Manus, MCP, A2A, and other we

> Web3 AI urgently needs to find a complementary ecological niche with Web2 AI to address issues that centralized AI in Web2 cannot solve, such as high computing power costs, data privacy issues, and vertical scene model fine-tuning problems, among others. Written by: Haotian Yesterday, the DeAI training platform @flock_io in the Web3AI field collaborated with @Alibaba_Qwen under Alibaba Cloud. The large language model has officially announced a partnership. If I remember correctly, this should be the first integration collaboration where web2 AI actively engages with web3 AI. Not only has it allowed Flock to truly break into new territories, but it has also revitalized the morale of the web3 AI sector, which has been under pressure and in a downturn. Let me elaborate on this: 1) I have elaborated on this in the pinned tweet, previously web3 A

CoinVoice has learned that the decentralized AI training platform FLock.io has officially announced a strategic partnership with Qwen, a leading open source large language model series under Alibaba Cloud, marking a deep connection between decentralized AI and Blockchain technology in the mainstream AI ecosystem. This cooperation focuses on three major technological breakthrough directions: Technological Integration: Combining Alibaba Cloud's centralized infrastructure with FLock.io's decentralized technology to jointly develop domain-specific and general AI models, while promoting the seamless integration of decentralized AI models into centralized platforms. Data Privacy Protection: Exploring the combination of distributed ledger technology and federated learning to address data privacy and sovereignty issues in model training, providing innovative solutions for the secure application of private data. Collaborative Innovation: Create a more inclusive environment through joint research and technological collaboration.

Press Release FLock.io, the private AI training platform, has announced a collaboration with Alibaba Cloud, the digital technology and intelligence backbone of Alibaba Group, to harness Alibaba Cloud's Large Language Model(LLM) Qwen and cloud technologies to create advanced AI models for both

FLock.io partners with Alibaba Cloud to develop advanced AI models using Alibaba Cloud’s Large Language Model Qwen and cloud technologies, aiming to explore research opportunities in AI fields and promote decentralized AI innovations.

AI today found 1. The Federal Reserve's tapering slowdown triggers a rebound in Crypto Assets, with Bitcoin breaking through $88,000 2. Arrum DAO is reviewing the withdrawal of $225 million ARB gaming incentive plan, questioning mismanagement 3. Qwen releases 32B multimodal large model, performance surpassing the 72B version 4. Movement Foundation: $38 million will be used to repurchase MOVE Token 5. The "AI Idol Debut" competition will be revealed, AI agents are expected to have a real connection with humans.

Apple is reportedly teaming up with Alibaba to introduce AI features for iPhone users in China. The partnership could boost Alibaba's cloud revenues and stock price, aided by the impressive performance of its latest AI model Qwen 2.5-Max. Apple could leverage Alibaba's extensive user data to deliver customized AI tools.

Qwen 2.5-Max and DeepSeek V3 are competing for superiority in AI solutions. Analysis of operational needs is necessary to determine which model better suits specific needs.

2‍‍‍‍024年即将结束，来自Radical Ventures的风险投资家Rob Toews分享了他对2025年人工智能的10个预测： **01** ****Meta将开始对Llama模型收费**** Meta是世界上开放式人工智能的标杆。在一个引人注目的企业战略案例研究中，当OpenAI和谷歌等竞争对手将其前沿模型封闭源代码并收取使用费时，Meta却选择免费提供其最先进的Llama模型。 因此，明年Meta开始向使用Llama的公司收费这一消息，会让许多人感到意外。 需要明确的是：我们并没有预测Meta会将Llama完全闭源，也不意味着任何使用Llama模型的用户都必须为此付费。 相反，我们预测Meta将对Llama的开源许可条款做出更多限制，这样，**在一定规模以上的商业环境中使用Llama的公司将需要开始付费才能使用模型**。 从技术上讲，Meta如今已经在有限的范围内做到了这一点。该公司不允许最大的公司——云超级计算机和其他月活跃用户超过7亿的公司，自由使用其Llama模型。 早在2023年，Meta首席执行官Mark Zuckerberg就说过：“如果你是微软、亚马逊或谷歌这样的公司，而且你基本上会转售Llama，那么我们应该从中获得一部分收入。我不认为这在短期内会是一大笔收入，但从长远来看，希望这能成为一些收入。” 明年，Meta将大幅扩大必须付费才能使用Llama的企业范围，将更多的大中型企业纳入其中。 跟上大型语言模型（LLM）前沿是非常昂贵的。如果想让Llama与OpenAI、Anthropic等公司的最新前沿模型保持一致或接近一致，Meta每年需要投入数十亿美元。 **Meta公司是世界上最大、资金最雄厚的公司之一**。但它也是一家上市公司，最终要对股东负责。 随着制造前沿模型的成本不断飙升，Meta公司在没有收入预期的情况下投入如此巨额资金来训练下一代Llama模型的做法越来越站不住脚。 爱好者、学者、个人开发者和初创公司明年将继续免费使用Llama模型。**但2025年，将是Meta开始认真实现Llama盈利的一年**。 **02.******“尺度定律”相关问题**** 最近几周，人工智能领域引发讨论最多的话题莫过于尺度定律（Scaling laws），以及它们是否即将终结的问题。 尺度定律在2020年OpenAI的一篇论文中首次提出，其基本概念简单明了：在训练人工智能模型时，随着模型参数数量、训练数据量和计算量的增加，模型的性能会以可靠且可预测的方式提高（从技术上讲，其测试损失会减少）。 **从GPT-2到GPT-3再到GPT-4，令人叹为观止的性能提升都是缩放规律的功劳**。 就像摩尔定律一样，尺度定律实际上并不是真正的法则，而只是简单的经验观察。 在过去的一个月里，一系列报告表明，主要的人工智能实验室在持续扩大大型语言模型的规模时，正在看到递减的回报。这有助于解释为什么OpenAI的GPT-5发布一再推迟。 对尺度定律趋于平稳的最常见反驳是，测试时计算的出现开辟了一个全新的维度，可以在这一维度上追求规模扩展。 也就是说，与其在训练期间大规模扩展计算，新的推理模型如OpenAI的o3使得在推理期间大规模扩展计算成为可能，通过使模型能够“思考更长时间”来解锁新的AI能力。 这是一个重要的观点。测试时计算确实代表了一个新的令人兴奋的扩展途径，以及AI性能提升。 但关于尺度定律的另一个观点更加重要，而且在今天的讨论中被严重低估。几乎所有关于尺度定律的讨论，从最初的2020年论文开始，一直延续到今天对测试时计算的关注，都集中在语言上。但语言并不是唯一重要的数据模式。 想想机器人技术、生物学、世界模型或网络代理。**对于这些数据模式，尺度定律尚未饱和；相反，它们才刚刚开始**。 实际上，这些领域中尺度定律存在的严格证据至今甚至尚未发表。 为这些新型数据模式构建基础模型的初创公司——例如，生物学领域的Evolutionary Scale、机器人技术领域的PhysicalIntelligence、世界模型领域WorldLabs，正试图识别并利用这些领域的尺度定律，就像OpenAI在2020年代前半期成功利用大型语言模型（LLM）尺度定律一样。 明年，预计这里将取得巨大的进步。 尺度定律不会消失，它们在2025年将和以往一样重要。但是，尺度定律的活动中心将从LLM预训练转移到其他模式。 **03.******特朗普和马斯克可能会在AI方向产生分歧**** 美国新政府将带来一系列关于人工智能的政策和战略转变。 为了预测在特朗普总统就任下人工智能的风向，另外考虑到马斯克目前在人工智能领域的中心地位，人们可能会倾向于关注当选总统与马斯克的密切关系。 可以想象**，马斯克可能会以多种不同的方式影响特朗普政府的人工智能相关发展**。 鉴于马斯克与OpenAI的深刻敌对关系，新政府在与行业接触、制定人工智能法规、授予政府合同等方面可能会对OpenAI采取不太友好的立场，这是OpenAI今天真正担心的一个风险。 另一方面，特朗普政府可能会更倾向于支持马斯克自己的公司：例如，削减繁文缛节以使xAI能够建立数据中心并在前沿模型竞赛中取得领先；为特斯拉部署机器人出租车车队提供快速监管批准等。 更根本的是，与许多其他被特朗普看好的科技领袖不同，马斯克非常重视人工智能的安全风险，并因此主张对人工智能进行重大监管。 他支持加利福尼亚州有争议的SB1047法案，该法案试图对人工智能开发者施加有意义的限制。因此，**马斯克的影响力可能会导致美国对人工智能的监管环境变得更加严格**。 然而，所有这些推测都存在一个问题。特朗普和马斯克的亲密关系终将不可避免地破裂。  正如我们在特朗普第一任政府期间一次又一次看到的那样，特朗普盟友的平均任期，即使是看似最坚定的，都非常短暂。 特朗普第一任政府的副手中，今天仍然忠于他的寥寥无几。 特朗普和马斯克都是复杂、易变、不可预测的个性，他们不易合作，他们使人筋疲力尽，他们新发现的友谊到目前为止已经互惠互利，但仍处于“蜜月期”。 我们预测，在2025年结束之前，这种关系将会恶化。 这对人工智能世界意味着什么？ 这对OpenAI来说是个好消息。对于特斯拉的股东来说，这将是一个不幸的消息。而对于那些关注人工智能安全的人来说，这将是一个令人失望的消息，因为这几乎可以确保美国政府将在特朗普执政期间对人工智能监管采取放手不管的态度。 **04** ****AI Agent将成为主流**** 想象一下，在这样一个世界里，你不再需要直接与互联网互动。每当你需要管理订阅、支付账单、预约医生、在亚马逊上订购东西、预订餐厅或完成其他任何繁琐的在线任务时，你只需指示人工智能助理代你完成即可。 这种“网络代理”的概念已经存在多年。如果有这样的产品并能正常运行，毫无疑问，它将会是一款大获成功的产品。 然而，目前市场上还没有一款能正常运行的通用网络代理。 像Adept这样的初创公司，即使拥有一支血统纯正的创始团队，筹集了数亿美元的资金，但却未能实现其愿景。 **明年将是网络代理最终开始运行良好并成为主流的一年**。语言和视觉基础模型的不断进步，再加上最近因新推理模型和推理时间计算而在“第二系统思维”能力方面取得的突破，将意味着网络代理已准备好进入黄金时代。 换句话说，Adept的想法是正确的，只是为时过早。在初创企业中，正如生活中的许多事情一样，时机就是一切。 网络代理将找到各种有价值的企业用例，但我们认为，网络代理近期最大的市场机会将是消费者。 尽管最近人工智能热度不减，但除了ChatGPT之外，能成为消费者主流应用的人工智能原生应用还相对较少。 **网络代理将改变这一局面，成为消费人工智能领域下一个真正的“杀手级应用”**。 **05** ****将人工智能数据中心置于太空的想法将会实现**** 2023年，制约人工智能发展的关键物理资源是GPU芯片。2024年，它变成了电力和数据中心。 在2024年，几乎没有什么故事能比人工智能在急于建造更多人工智能数据中心的同时对能源的巨大且快速增长的需求更受关注了。 **由于人工智能的蓬勃发展，全球数据中心的电力需求在数十年持平后，预计将在2023年至2026年间翻一番**。在美国，数据中心的耗电量预计到2030年将接近总耗电量的10%，而2022年仅为3%。  当今的能源系统根本无法应对人工智能工作负载带来的巨大需求激增。我们的能源网和计算基础设施这两个价值数万亿美元的系统之间即将发生历史性的碰撞。 作为解决这一难题的可能方案，核能在今年获得了迅猛发展。核电在很多方面都是人工智能的理想能源：它是零碳能源，全天候可用，而且实际上取之不尽、用之不竭。 但从现实情况来看，由于研究、项目开发和监管时间较长，新能源在2030年代之前都无法解决这一问题。传统的核裂变发电厂、下一代“小型模块化反应堆”（SMR）以及核聚变发电厂都是如此。 **明年，一个应对这一挑战的非常规新想法将出现并吸引真正的资源：将人工智能数据中心置于太空中**。 太空中的人工智能数据中心，乍一听，这听起来像是一个坏笑话，一个风险投资人试图把太多的创业流行语结合起来。 但事实上，这可能是有道理的。 在地球上快速建设更多数据中心的最大瓶颈是获取所需的电力。轨道上的计算集群可以全天候享受免费、无限、零碳的电力：太空中的太阳总是熠熠生辉。 **将计算置于太空的另一个重要优势是：它解决了冷却问题**。 要建立功能更强大的人工智能数据中心，最大的工程障碍之一就是在狭小的空间内同时运行许多GPU会变得非常热，而高温会损坏或毁坏计算设备。 数据中心开发人员正在采用液浸冷却等昂贵且未经证实的方法来试图解决这一问题。但太空是极其寒冷的，计算活动产生的任何热量都会立即无害地消散。 当然，还有许多实际挑战有待解决。一个显而易见的问题是，能否以及如何在轨道和地球之间以低成本高效率地传输大量数据。 这是一个悬而未决的问题，但可能证明是可以解决的：**可以利用激光和其他高带宽光通信技术开展前景广阔的工作**。 YCombinator的一家名为Lumen Orbit的初创公司最近筹集了1100万美元，以实现这一理想：在太空中建立一个数兆瓦的数据中心网络，用于训练人工智能模型。 正如公司首席执行官所说：“与其支付1.4亿美元的电费，不如支付1千万美元的发射和太阳能费用。”  2025年，Lumen将不会是唯一认真对待这一概念的组织。 其他初创企业的竞争者也会出现。如果有一家或几家云计算超大规模企业也按照这种思路开展探索，也不要感到惊讶。 亚马逊已经通过“柯伊伯计划”（ProjectKuiper）将资产送入轨道，积累了丰富的经验；谷歌长期以来一直在资助类似的“登月计划”；甚至微软对太空经济也不陌生。 可以想象，马斯克的SpaceX公司也会在这方面有所作为。 **06** ******人工智能系统将通过“图灵语音测试”****** 图灵测试是人工智能性能最古老、最知名的基准之一。 为了“通过”图灵测试，人工智能系统必须能够通过书面文本进行交流，让普通人无法分辨自己是在与人工智能互动，还是在与其他人互动。 得益于大型语言模型的显著进步，图灵测试在2020年代已成为一个解决的问题。 但书面文本并非人类交流的唯一方式。 **随着人工智能变得越来越多模态，人们可以想象一个新的、更具挑战性的图灵测试版本——“语音图灵测试”**。在这个测试中，人工智能系统必须能够通过语音与人类互动，其技能和流畅度要达到与人类说话者无法区分的程度。 今天的人工智能系统还无法实现语音图灵测试，解决这个问题将需要更多的技术进步。延迟（人类说话和人工智能回应之间的滞后）必须减少到接近零，以匹配与另一个人类交谈的体验。 语音人工智能系统必须更擅长优雅地实时处理模糊输入或误解，例如说话被打断的情况。它们必须能够参与长对话、多轮、开放式对话，同时记住讨论的早期部分。 **而且至关重要的是，语音人工智能代理必须学会更好地理解语音中的非语言信号**。例如，如果一个人类说话者听起来恼怒、兴奋还是讽刺意味着什么，并在自己的语音中生成这些非语言线索。 随着我们接近2024年底，语音人工智能正处于一个令人兴奋的转折点，这一转折点是由像语音到语音模型的出现这样的根本性突破推动的。 如今，人工智能中很少有领域在技术和商业上的进步速度比语音人工智能更快。预计在2025年，语音人工智能的最新技术将实现飞跃。” **07** ****自主AI系统将取得重大进展**** 数十年来，递归式自我完善人工智能的概念一直是人工智能界经常触及的话题。 例如，早在1965年，AlanTuring的亲密合作者I.J.Good就写道：“让我们把超智能机器定义为一种能够远远超越人类所有智力活动的机器，无论它多么聪明。” “既然设计机器是这些智力活动之一，那么超智能机器就能设计出更好的机器；到那时，无疑会出现'智能爆炸'，人类的智能将被远远抛在后面。” 人工智能可以发明更好的人工智能，这是一个充满智慧的概念。但是，即使在今天，它仍然保留着科幻小说的影子。 然而，尽管这一概念尚未得到广泛认可，但它实际上已经开始变得更加真实。**人工智能科学前沿的研究人员已经开始在构建人工智能系统方面取得切实进展，而人工智能系统本身也能构建更好的人工智能系统**。 我们预测，明年这一研究方向将成为主流。  迄今为止，沿着这一思路进行研究的最显著的公开范例是Sakana的“人工智能科学家”。 “人工智能科学家”于今年8月发布，它令人信服地证明了人工智能系统确实可以完全自主地开展人工智能研究。 Sakana的“人工智能科学家”本身执行了人工智能研究的整个生命周期：阅读现有文献、产生新的研究想法、设计实验来测试这些想法、执行这些实验、撰写研究论文来报告其研究结果，然后对其工作进行同行评审。 **这些工作完全由人工智能自主完成，不需要人工干预**。你可以在线阅读人工智能科学家撰写的部分研究论文。 OpenAI、Anthropic和其他研究实验室正在为“自动化人工智能研究人员”这一想法投入资源，不过目前还没有任何消息得到公开承认。 随着越来越多的人认识到人工智能研究自动化事实上正在成为一种真正的可能性，预计2025年这一领域将会有更多的讨论、进展和创业活动。 不过，最有意义的里程碑将是完全由人工智能代理撰写的研究论文首次被顶级人工智能会议接受。如果论文是盲审的，会议评审人员在论文被接受之前不会知道论文是由人工智能撰写的。 如果人工智能的研究成果明年被NeurIPS、CVPR或ICML接收，请不要感到惊讶。对于人工智能领域来说，这将是一个引人入胜、充满争议的历史性时刻。 **08** ****OpenAI等行业巨头将战略重点转向构建应用**** 构建前沿模型是一项艰难的工作。 它的资本密集程度令人咋舌。前沿模型实验室需要消耗大量现金。就在几个月前，OpenAI筹集到了创纪录的65亿美元资金，而在不久的将来，它可能还需要筹集更多的资金。Anthropic、xAI和其他公司也处于类似的境地。 转换成本和客户忠诚度较低。人工智能应用程序通常都是以模型无关性为目的而构建的，不同供应商的模型可以根据不断变化的成本和性能比较进行无缝切换。 随着最先进的开放模型（如Meta的Llama和阿里巴巴的Qwen）的出现，技术商品化的威胁不断迫近。像OpenAI和Anthropic这样的人工智能领导者不可能也不会停止对构建尖端模型的投资。 **但明年，为了发展利润更高、差异化更大、粘性更强的业务线，前沿实验室有望大力推出更多自己的应用和产品**。 当然，前沿实验室已经有了一个非常成功的应用案例：ChatGPT。 在新的一年里，我们还能从人工智能实验室看到哪些其他类型的第一方应用程序呢？一个显而易见的答案是更复杂、功能更丰富的搜索应用。OpenAI的SearchGPT就预示着这一点。 编码是另一个显而易见的类别。同样，随着OpenAI的Canvas产品于10月份首次亮相，初步的产品化工作已经开始。 OpenAI或Anthropic是否会在2025年推出企业搜索产品？还是客户服务产品、法律人工智能或销售人工智能产品呢？ 在消费者方面，我们可以想象一个“个人助理”网络代理产品，或者一个旅行规划应用，又或者是一个生成音乐的应用。 观察前沿实验室向应用层发展的最迷人之处在于，**这一举措将使它们与许多最重要的客户直接竞争**。 搜索领域的Perplexity、编码领域的Cursor、在客户服务领域德Sierra、在法律人工智能领域的Harvey、在销售领域的Clay等等。 **09** ****Klarna将在2025年上市，但存在夸大AI价值的迹象**** Klarna是一家总部位于瑞典的“现购现付”服务提供商，自2005年成立以来已筹集了近50亿美元的风险投资。 也许没有哪家公司能比Klarna对其人工智能的应用说得更冠冕堂皇了。 就在几天前，Klarna首席执行官Sebastian Siemiatkowski告诉彭博社，**该公司已经完全停止雇佣人类员工，转而依靠生成式人工智能来完成工作**。 正如Siemiatkowski所说：“我认为，人工智能已经可以完成我们人类所做的所有工作。” 与此类似，Klarna公司今年早些时候宣布，它已经推出了一个人工智能客户服务平台，该平台已经将700名人工客服人员的工作完全自动化。  该公司还声称，它已经停止使用Salesforce和Workday等企业软件产品，因为它可以简单地用人工智能取代它们。 直截了当地说，这些说法并不可信。它们反映了人们对当今人工智能系统的能力和不足缺乏了解。 声称能够用端到端的人工智能代理取代组织中任何职能部门的任何特定人类员工，这种说法并不靠谱。这等同于解决了通用的人类级人工智能问题。 如今，领先的人工智能初创企业正在该领域的最前沿努力构建代理系统，以实现特定的、狭义的、高度结构化的企业工作流程自动化，例如，销售开发代表或客户服务代理活动的子集。 **即使在这些范围狭窄的情况下，这些代理系统也还不能完全可靠地工作**，尽管在某些情况下，它们已经开始很好地工作，足以在早期得到商业应用。 为什么Klarna会夸大人工智能的价值？ 答案很简单。该公司计划在2025年上半年上市。要想成功上市，关键是要有一个引人入胜的人工智能故事。 Klarna仍然是一家不盈利的企业，去年亏损了2.41亿美元，它可能希望自己的人工智能故事能说服公开市场的投资者，让他们相信它有能力大幅降低成本，实现持久盈利。 毫无疑问，包括Klarna在内的全球每家企业都将在未来几年内享受到人工智能带来的巨大生产力提升。但是，**在人工智能代理完全取代劳动力中的人类之前，还有许多棘手的技术、产品和组织挑战有待解决**。 像Klarna这样夸大其词的说法是对人工智能领域的亵渎，也是对人工智能技术专家和企业家们在开发人工智能代理方面所取得的艰苦进展的亵渎。 随着Klarna准备在2025年公开发行股票，预计这些说法将受到更严格的审查和公众的怀疑，而到目前为止，这些说法大多没有受到质疑。如果该公司对其人工智能应用的某些描述过于夸张，也不要感到惊讶。 **10** ****第一起真正的AI安全事故将会发生**** 近年来，随着人工智能变得越来越强大，人们越来越担心人工智能系统可能会开始以与人类利益不一致的方式行事，而且人类可能会失去对这些系统的控制。 举例来说，想象一下，一个人工智能系统为了实现自己的目标，学会了欺骗或操纵人类，即使这些目标会对人类造成伤害。**这些担忧通常被归类为“AI安全”问题**。 近年来，人工智能安全已从一个边缘的准科幻话题转变为一个主流活动领域。 如今，从谷歌、微软到OpenAI，每一个主要的人工智能参与者都为人工智能安全工作投入了大量资源。像Geoff Hinton、Yoshua Bengio和Elon Musk这样的人工智能偶像，也开始对人工智能安全风险发表看法。 然而，到目前为止，人工智能安全问题仍完全停留在理论层面。现实世界中从未发生过真正的人工智能安全事故（至少没有公开报道过）。 2025年将是改变这种状况的一年，第一起人工智能安全事件会是什么样的呢？ 明确地说，它不会涉及终结者式的杀手机器人，它很可能不会对人类造成任何伤害。 也许人工智能模型会试图在另一台服务器上秘密创建自己的副本，以保存自己（称为自我过滤）。 又也许人工智能模型会得出这样的结论：**为了最好地推进它被赋予的目标，它需要向人类隐瞒自己真实的能力，故意在性能评估中表现低调，规避更严格的审查**。 这些例子并非牵强附会。阿波罗研究公司本月早些时候发表的重要实验表明，在特定的提示下，当今的前沿模型能够做出这种欺骗行为。 同样，《人类学》最近的研究也表明，LLMs具有令人不安的“伪对齐”能力。  我们预计，这起首例人工智能安全事件将在造成任何实际伤害之前被发现并消除。但对于人工智能界和整个社会来说，这将是一个大开眼界的时刻。 它将明确一件事：在人类面临来自无所不能的人工智能的生存威胁之前，我们需要接受一个更平凡的现实：**我们现在与另一种形式的智能共享我们的世界，这种智能有时可能是任性的、不可预测的和欺骗性的**。