全部新闻
一种新型的AI模型使数据所有者可以控制
Allen AI研究所（AI2）的研究人员开发了Flexolmo，这是一种大型语言模型，即使在训练模型后，数据贡献者也可以保留对数据的控制。与传统的方法不同，一旦将数据用于培训，它就无法轻易检索，Flexolmo的设计使数据所有者能够在不放弃所有权的情况下贡献其数据，并在需要时提供以后提取的可能性。该模型使用“专家的混合”体系结构，但使用独立的子模型合并技术进行创新，以保留个人数据贡献。在专有数据集测试中，Flexolmo在各种任务中的表现优于其他模型，同时为数据治理和隐私提供了独特的优势。
报告说，没有足够的AI芯片来支持数据中心的预测。
高端预测可以预测，由于全球半导体芯片生产的限制，伦敦经济学国际将认为美国数据中心扩张和支持AI开发的电力增长被认为“不可信”。该报告表明，根据分析，预计的数据中心需求将需要全球90％的芯片供应，这是一个不切实际的情况。高估引起了人们对投机基础设施投资和纳税人潜在经济风险的担忧。
Cal State LA从CSU获得两个人工智能项目的资金
Cal State La通过CSU的AIEIC倡议赢得了两个AI项目的资金，旨在促进AI在教育中的创新使用。这些项目是“在AI时代的正直教学”和“ AI增强的STEM补充教学研讨会”。第一个重点是开发道德AI知识的教学实践，而第二个则将生成的AI工具整合到STEM研讨会中，以增强学生的成功。该计划是CSU更广泛的AI战略的一部分，旨在促进其23个校园的学术卓越和创新。
从非技术背景过渡到AI的5种方法
许多成功的AI专业人员没有技术背景。人们可以通过几个实用步骤从各个领域（例如营销，心理学，法律或设计）过渡到AI：通过探索Chatgpt等工具，成为团队中的AI人；学习基本的AI和机器学习概念；将自己定位为AI翻译人员桥梁技术和非技术利益相关者；使用无代码/低代码工具来构建AI项目；并涉及诸如项目经理或技术作家之类的AI-ADJACEST角色。这些方法使个人能够在逐渐获得AI知识的同时利用其领域专业知识。
Why AI projects fail, according to EPA CIO US Environmental Protection Agency CIO Carter Farmer warns against盲目追求AI，强调在没有明确用例和充分验证数据的情况下贸然采用AI会导致失败。他认为AI不是万能的解决方案，并且解决某些问题不一定需要AI。Farmer指出，在应用AI时必须重新定义业务流程，避免简单地将现有流程映射到AI系统中。他提倡对项目进行彻底审查以确定是否真正值得使用AI，以及制定一个明确的数据规划来指导AI项目的实施。 US Environmental Protection Agency CIO Carter Farmer has a blunt message for AI hype-chasers: Shiny-object syndrome too often drives teams to leap into AI without defining a clear use case or vetting their data, leaving them to wonder why it doesn't work. Speaking during a webinar hosted by FedInsider, Farmer said that AI isn't a cure-all business operations wonder drug as many people assume. It must be brought into a business with a specific use case in mind, the EPA CIO said. "People hear AI and they think it can solve any problem," Farmer argued. "They see what it's done at other organizations and jump in without asking the right questions." Many times the problem you're trying to solve doesn't need AI He explained, "Many times the problem you're trying to solve doesn't need AI," adding that jumping to keep up with the latest buzzwords can ultimately slow growth down rather than accelerate it. "If you don't ask those questions up front you can go down a road pretty far before you realize you need to back up, reverse and figure out where you should have gone down first," Farmer said. That's a reasonable take when you consider investment returns on AI have been found to be pretty poor of late, with just one in four AI bets are paying off according to a recent survey of 2,000 CEOs. Farmer made his comment in response to Ed Bodensiek, customer experience market leader at government tech services firm Maximus. Bodensiek said he's heard other government tech leaders observe that there are too many bots at government agencies. His company has developed an entire "mission readiness assessment" process it uses to determine the appropriate approach to solving a problem - which might not end up involving AI at all. So what's going wrong with so many AI implementations? For Farmer, it comes down to two things: Processes and data. Applying AI to business processes doesn't mean simply mapping what one does now onto an AI, Farmer explained. AI infrastructure investment may be $8T shot in the dark AI agents get office tasks wrong around 70% of the time, and a lot of them aren't AI at all Tinfoil hat wearers can thank AI for declassification of JFK docs When it comes to AI ROI, IT decision-makers not convinced "What people should be doing is redefining and reimagining that process as it would apply to automation," the CIO said. "[The] lift and shift mentality has to be stopped in its tracks." As an example, Farmer described a presumably hypothetical situation where a form that had been part of a business process for years was still being filled out, even though the same function was now accomplished automatically, and digitally. Without examining the business process behind that form, it might get included in an AI workflow when it wouldn't need to be. A process review can force teams to ask really critical questions - like whether AI-ifying a particular thing is really worth the effort. "Value add might come at 2x, 3x or 4x the cost of doing it manually," Farmer explained. In short, if AI is only as good as the data that goes into it, an AI project is only as good as the data that went into planning it. Farmer has shown his work on that end. The EPA keeps an inventory of AI applications at the agency, part of his philosophy of "finding the actual best use cases where AI can be best applied," as he explained during the webinar. ®
我们是否对AIS承担重要任务感到满意？
数学中的一个例子说明了对AI可信度的转移观点。1976年，《四种颜色定理》的计算机辅助证明引起了怀疑机器验证的数学家的争议。现在，AI高级AI通过建议机器可以提供比人类更多可靠的证据来挑战这种怀疑。这场辩论反映了有关适当的AI使用和依赖的更广泛的社会问题。
Adobe：深入定价 - 使用AI进入增长模式（NASDAQ：ADBE）
Adobe的业务案例由于AI的战略性整合到其软件产品中，推动了新的增长机会，因此仍然很强大。随着订户数量的增长和交叉销售的增加，这反映在增加的ARR，扩大的多年RPO和持续的利润率中。该公司已提高了2025财年的指导，建立了连续十个季度的两位数性能。此外，Adobe与AI相关的ARR超过了2025财政年度末的2.5亿美元目标，高于FQ1 2025年FQ1的1.25亿美元。股票进一步支持股票回购，较低的估值和提高的共识估计。
科学家说，新的AI比以往任何时候都更好地预测我们的行为。
科学家已经开发了一种名为Centaur的AI模型，该模型可以根据心理实验中超过1000万个实际决定的数据来预测人类行为，其精度为64％。在一个名为Psych-101的数据集上接受培训，该数据集包括160个心理学实验中60,000名参与者的选择，Centaur超过了预测能力中的先前模型，并且可以适应新的情况。研究人员旨在扩大模型的培训数据，以包括人口统计细节，以进行更多个性化的预测，并探讨其在理解认知过程和心理健康状况方面的应用。
太阳谷：从导航特朗普到AI的恐惧，这就是媒体大亨的想法
Allen＆Co。的年度Sun Valley会议，通常称为“亿万富翁夏令营”，看到来自全球聚会的媒体和技术首席执行官。但是，由于与苹果和亚马逊等科技巨头相比，由于市值下降，今年的与会者的信心较小。传统媒体领导者面临的关键问题包括在合并期间进行合并后缩减，与特朗普总统的关系导航，在迪士尼领导力变化的不确定性以及努力努力AI对其商业模式的影响之后的不确定性。
超越及时的手工制作：如何成为AI对程序员的更好伙伴
本文讨论了与Github Copilot有效合作的策略，而不仅仅是制作提示。它强调提供足够的上下文和指导以确保副本生成与项目要求一致的质量代码建议的重要性。这包括在代码中添加注释，使用项目中的自定义说明文件，并利用模型上下文协议（MCP）服务器访问相关数据或服务。这些方法有助于副驾驶更好地了解特定于项目的细节并遵守编码标准，从而增强其作为开发援助的效用。
