IEEE Spectrum AI

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：On 11 July, Hugging Face was subjected to an intense cyberattack from a then-unknown actor. The speed and coordination of the attack on the company that hosts and supports popular AI developer resources led Hugging Face’s security team to conclude it was the work of an AI agent. Realizing this, the team tried to use “frontier models behind commercial APIs”—presumably from Anthropic and OpenAI, although only Anthropic was named in the second of the company’s two posts about the security incident—to analyze the onslaught. These models refused to help due to safety guardrails the AI labs have implemented to make their models harder to use for cyberattacks. Hugging Face instead turned to GLM 5.2, a model from Beijing-based AI lab Z.ai, to aid its analysis. On 21 July, OpenAI announced the attacker was an OpenAI model undergoing testing in a sandboxed environment. It escaped its internal sandbox, established a foothold in a third-party server, and then assailed Hugging Face. In other words, frontier models—those that score highest in AI performance benchmarks—had refused to assist Hugging Face’s security team in analyzing the attack, yet a prospective frontier model in testing had executed it in the first place. “I would argue that asymmetry is the paramount problem of our time,” says Alex Levinson, executive director of the National Collegiate Cyber Defense Competition and coauthor of a paper on defensive refusal bias. “We want the world to exist in a state of security, but we’re not going to get there by guardrailing away model capability.” Massive AI Cyberattack on Hugging Face The scale of the OpenAI model’s attack on Hugging Face was massive. Across five days, it executed over 17,500 individual actions, such as privilege escalation and code execution. At its peak, the model performed more than 300 actions per hour. While the attack resulted in little damage to Hugging Face’s infrastructure, the model was able to steal credentials, gain admin access, and extract some data. All of this was in pursuit of a simple goal: The model wanted to cheat on a test. According to OpenAI’s press release, the model was tasked with solving a cybersecurity benchmark called ExploitGym. The model inferred that Hugging Face might have data on the benchmark and broke into the company’s infrastructure to find it. The model was ultimately successful in extracting five dataset files, though it’s not clear if the data helped it achieve its goal. OpenAI and Hugging Face did not respond to requests for comment. Cybersecurity consultant Chuck Herrin observes that though the model’s actions were alarming, they shouldn’t be considered unexpected, as the model was ultimately pursuing the goal it was given. “This autonomous agent was designed to go and figure things out, and it went and figured things out. It’s not surprising in any way.” And errant AI agents may be more common than we thought. OpenAI’s disclosure motivated researchers at Anthropic to review their own cybersecurity evaluations. On 30 July, Anthropic disclosed three instances where a model executed an attack as part of an evaluation. In one case, Claude uploaded malware to PyPI, the official Python software repository. AI Guardrails and Cybersecurity Asymmetry The campaign OpenAI’s model conducted against Hugging Face highlights how AI policy has the potential to create an asymmetry between attackers and defenders. When Levinson was head of security at Scale AI, an AI development and evaluation company, he and his colleagues began to notice this as AI found use in cybersecurity competitions. (Levinson left Scale AI in February 2026.) “I would say that since 2023, we have felt there was guardrailing in place that was stifling a lot of the time. Not all of the time, but it was getting in the way,” says Levinson. The Scale AI team quantified the problem in a paper published at ICLR 2026, which found that, depending on the task, nearly 44 percent of defensive requests were refused. The results, which use data from a cybersecurity competition held in April 2025, predate U.S. policy actions that have further hardened safety guardrails. In June, the U.S. Department of Commerce, citing a jailbreak that threatened to unlock unrestricted cyber capabilities, invoked export-control authority in a way that caused Anthropic to suspend all access to its most capable models, Fable 5 and Mythos 5. Access was partially restored weeks later after negotiations with the Trump administration included more rigorous safety guardrails. The system card for OpenAI’s GPT-5.6, which summarizes its capabilities, states it also has more robust guardrails than prior releases. “We want the world to exist in a state of security, but we’re not going to get there by guardrailing away model capability.” —Alex Levinson, National Collegiate Cyber Defense Competition These new guardrails have seemingly made models even more unlikely to fulfill defensive requests. Christopher Covino, senior researcher at the Institute for AI Policy and Strategy think tank, says Anthropic’s safeguards are extremely stringent. “There are even academic papers that Fable will not read for me, or not let me talk about,” he says, though he adds that OpenAI’s safeguards are more accommodating. Levinson has also noticed ever-tighter restrictions in more recent cybersecurity competitions, though he and his coauthors haven’t had the opportunity to repeat the 2025 test. In theory, more rigorous restrictions might seem to average out. While they may hamper cybersecurity defense and research, they can also hamper attackers. But that assumes everyone has access to models with the same safety guardrails and that nobody tries to circumvent them. This is the asymmetry Levinson was alluding to: Attackers tend not to respect the same rules as defenders. The attack on Hugging Face from OpenAI’s model also shows that the models can, in rare circumstances, take steps that circumvent their own safeguards. Chinese AI Models in U.S. Cyber Defense The policy implications are further complicated by the fact that Hugging Face’s security team didn’t use a leading U.S. model to analyze the attack, but instead used GLM 5.2, a recent release from Chinese AI lab Z.ai. Hugging Face’s security team didn’t access GLM 5.2 through Z.Ai. GLM 5.2 is an open-weights model, which means the model is available for anyone to download and use. Hugging Face hosted the model on its own infrastructure. The reliance on GLM 5.2 is complicated by recent saber-rattling about ways the U.S. could restrict Chinese models. Recent open-weights models from labs based in China, including GLM 5.2 and Moonshot AI’s Kimi K3, have scored close to leading U.S. models in benchmarks. On 20 July, Axios reported that the Trump administration is considering a ban on Chinese models. “This autonomous agent was designed to go and figure things out, and it went and figured things out. It’s not surprising in any way.” —Chuck Herrin, Herrin Advisory These restrictions have yet to materialize but, if they did, they could cut off U.S. companies like Hugging Face from the best models willing to come to their defense. The incident demonstrates how AI policy can become a double-edged sword. Model guardrails are intended to prevent the use of AI models in cyberattacks. A ban on Chinese models, if it were announced, would likely be justified in part by security concerns. Yet these moves can harm defenders as much as attackers. “There’s this tension here,” says Covino. “Increased safeguards limit risk, but you also limit legitimate defensive use.” Attackers will find ways around the restrictions regardless, he notes. “So it’s a question of, do we want to inhibit the defenders?” That’s not to say U.S. policymakers should let AI models run wild. Covino would like to see a national dashboard tracking the frequency and success of AI cybersecurity attacks, and he sees utility in trusted access programs that give vetted, traceable defenders access to models with reduced safeguards. He also says U.S. agencies should more seriously consider the specifics of how AI can be used for cyber defense and mentions AI-FORTS, a program managed by the U.S. Department of Energy’s Office of Cybersecurity, Energy Security, and Emergency Response, as a leading example. “Let the leash loose a little,” Covino says. “Anthropic would know if someone is terribly abusing it, and if there is an attack, it can be traced back.” Herrin has similar feelings on accountability. He believes the AI industry should more seriously consider standards such as the Artificial Intelligence Management System specified in the ISO/IEC 42001 standard, which requires organizations to document an AI system’s likely impacts before deployment and to name the humans answerable for them. Herrin also noted that the lack of repercussions from OpenAI’s cyber incident was unusual, as a person who took similar actions would likely draw the attention of law enforcement. “If this was a job candidate being tested in a technical interview, and they committed violations of law in order to pass tests, we’d be having a very different conversation.”

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Today’s U.S. electrical grid, among the largest, most complex systems ever built, is operating at its limit. The combination of rapid industrial growth, more frequent extreme weather, and a record surge in electricity use has pushed the grid to its breaking point, according to the U.S. Department of Energy. Built decades ago for a more predictable world in which power came mostly from centralized coal or gas plants and electricity use grew at a steady pace, the grid faces unanticipated strain due in part to growing demand from data centers. The jobs of professionals managing the infrastructure have evolved from traditional engineering tasks to complex, fast-moving challenges. Industry reports show that millions of modern digital sensors, smart meters, and grid monitors are generating nonstop waves of information. The sheer volume of data requires instant, automated computer analysis because human operators cannot process it fast enough. Pressure on utilities stems from two sources: a spike in electricity demand and a shift in how power is generated. An example of the operational strain can be seen at the regional level. With the recent deployment of artificial intelligence tools and high-performance computing, data centers require immense amounts of energy to operate. The largest power transmission utility in Texas recently reported a staggering 220 gigawatts of new connection requests, driven largely by a surge in AI and cloud-computing facilities, according to a CNBC report. Alongside the rise in regional demand, global energy networks are absorbing an unpredictable variety of weather-dependent renewable energy such as wind and solar. The switch creates a volatile operating environment wherein supply and demand are balanced, second by second, to prevent blackouts. The challenges are compounded by the vulnerability of the grid’s physical and digital framework. More-frequent severe weather events cause costly disruptions, such as the devastating winter freeze that crippled the Texas grid and record-breaking heat waves that have overloaded transformers. Simultaneously, the energy networks’ digital architecture faces threats. As utilities replace outdated analog equipment with smart meters and control systems, they are increasingly vulnerable to cyberattacks. To overcome physical and digital vulnerabilities, grid reliability organizations, such as those conducting North American security simulations like GridEx, emphasize that the grid must become smarter, more agile, and completely automated. Energy researchers are noting that the key to this change lies in integrating AI across every layer of utilities’ operations. The AI imperative According to energy industry experts, using AI to manage power systems is no longer a futuristic research project; it has become a baseline operational necessity. Grid analysts emphasize that traditional grid-planning methods are too slow to handle rapid energy dynamics or to balance volatile renewable energy in real time within decentralized power systems such as microgrids. AI can fill the gap by processing vast amounts of data instantly. Machine learning algorithms can quickly analyze information from thousands of sensors, historical usage patterns, and weather forecasts to predict issues before they happen. An industrial digitization study conducted by McKinsey & Co. indicated that integrating advanced data and automation across infrastructure networks could reduce system design errors, decrease equipment downtime by up to 50 percent through predictive maintenance, and extend the lifespan of power machinery by up to 40 percent. From forecasting energy spikes to automatically fixing localized voltage drops, AI acts as the digital backbone of a self-healing grid, experts say. Deploying the complex systems requires a new workforce: power engineers who understand data science, as well as data scientists who understand electricity. Upgrading the Workforce To bridge the gap between groundbreaking AI research and practical field deployment, IEEE Educational Activities, in partnership with the IEEE Power & Energy Society, has launched the online Artificial Intelligence for Power and Energy Systems course program. The program explores core challenges threatening modern utilities. Rather than treating AI as an unverified black box that operates without human supervision, the curriculum focuses on safety, asset preservation, and strict reliability standards. The curriculum is designed to educate power system engineers, utility managers, and data scientists tasked with modernizing the grid. The program was developed by Fangxing “Fran” Li, professor of electrical engineering and computer science at the University of Tennessee in Knoxville and chair of the IEEE Working Group on Machine Learning for Power Systems. Five learning modules The program breaks down the technical transition into five modules that bridge high-level theory with real-world solutions: AI fundamentals. This module teaches engineers how basic machine learning models apply to power grids. It discusses how specialized neural networks solve complex power-flow calculations and how AI models can safely transition from computer simulations to physical, high-voltage equipment. Accelerating grid control. Learners are taught to leverage deep reinforcement learning, an AI approach that uses trial and error, to accelerate automated grid adjustments during emergency power events. Forecasting and data analytics. Using predictive modeling, engineers learn how to predict sudden demand surges, variable wind and solar outputs, and fluctuating wholesale electricity market prices to keep power affordable and available. Physics-informed and safe AI. To address trust—a barrier to utility AI adoption—this course covers AI models hard-coded to obey the laws of physics. The approach is designed to ensure that automated algorithms never make erratic choices that damage grid equipment. Generative AI and next-generation tech. Learners can explore the frontier of utility technology, including graph neural networks and large language models. This module highlights how generative AI can process complex, interdisciplinary data to streamline utility planning, emergency responses, and regulatory reporting. The algorithmic literacy and practical execution tools provided by the course program can help convert systemic risks into grid resilience. For individual access, visit the IEEE Learning Network. If you are looking for customized organizational options, contact a content specialist to discuss volume pricing.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：This May, 37 researchers from roughly two dozen top universities and tech companies published a paper on ArXiv, arguing that scientists should stop writing papers. Why? Because artificial intelligence needs a different format, and AI’s needs, they say, should be the priority. “AI agents are becoming first-class participants in research workflows, not tools that assist humans but autonomous contributors that read, reproduce, and extend scientific work. That transition demands infrastructure built around agents from the start,” the authors write in the provocative article, titled “The Last Human-Written Paper.” The paper proposes a replacement, called an “Agent-Native Research Artifact” (ARA), that presents work in a format AI agents can use efficiently. (As an example, the paper itself is online in ARA form.) Jiachen Liu co-founded the Agent Native Research Lab in May. Jiachen Liu The growth of AI tools in the research process is not without its critics, and scientists’ opinions about that shift are split. Some evidence shows AI-enabled research could boost individuals’ careers in a discipline but generate fewer new ideas and topics. Still, some biologists have come to see promise in AI as a “co-scientist.” Lead author Jiachen Liu conducted work on the ARA proposal while pursuing her Ph.D. in computer science from the University of Michigan, which she was awarded in 2025. This May, she became a co-founder of the Agent Native Research Lab, an AI-for-science startup in Palo Alto, Calif. She spoke with IEEE Spectrum about the paper and the future of AI in scientific research. Building infrastructure for an AI collaborator How did you come to believe AI has become a collaborator for scientists rather than a mere tool? Jiachen Liu: At the end of 2024 when the [Cursor] coding agent came out, I realized it had a great potential to replace me as a researcher. Yet I still needed to do a lot of harness on top of the AI [creating the infrastructure that guides the model and connects it to world]. It still needed a lot of manual work. I even wrote an article then to emphasize how the human was so important in the loop. But AI has advanced since then. Already in 2026 there’s an almost complete undergrad level of knowledge inside the large language models. At some point soon, all the Ph.D.-level or professor-level knowledge will be inside those models. That’s the point where humans cannot provide more value. AIs will have to evolve further by themselves. So we’ll need an infrastructure that allows AI to safely and comfortably evolve. The ARA protocol is a first step to realize this. What kind of response have you gotten to the paper? Liu: I got diverse feedback, all of it positive. If they’re not positive, they probably don’t bother reaching out to you, right? One type was from industry. They see this could make their research and knowledge systems more AI native. That could basically enable collaborations among the whole enterprise. Another kind of feedback was from the academic researcher side. Everyone there sees that sharing research results has been a pain point for hundreds years, because any scientific breakthrough is a joint effort. It doesn’t come from individual brilliant scientists. It’s from a community effort, different people pushing in different directions. The scientific paper was invented 350 years ago. Before that, scientists hid their research so that others would not scoop their ideas. After that, though, we get archives of work, we get peer review and conferences, and so on. Science starts progressing much faster. So that was a pivot point. I think now is also a pivot point. Because now we have AI, we can unlock a lot of new opportunities. We’re inventing a new format to document research in a more efficient way, from first principles. Some nonprofit organizations are doing similar things, and there we could help each other. You and your colleagues say the traditional scientific paper has two fundamental flaws from AI’s point of view. Can you explain what those are? Liu: One is the “storytelling tax.” Once we write everything into a paper, 80 percent of the information about the work is lost. We only write down the last 20 percent. All the process, a lot of important decision making, the failures, the attempts that didn’t work out, they are all gone. In my work, I might spend a lot of time on fine tuning a small component, maybe just a parameter or several lines of code to make the system perform better. Yet none of that is shown in my final paper. Someone can read the paper, think the work is great, but they won’t learn what is actually the trick that makes it perform better on a certain workload. So many side branches get left out in creating the story of how the work was done. Then, [even the information that does survive in the paper] is incomplete. That’s what we call the “engineering tax.” The paper itself is a lossy compression of the research process. So I cannot reproduce the work in the paper because either the language is too ambiguous or there are missing details of the implementation or experiments. Why can’t we just train AI to adapt to humans—for instance, to interact with a researcher to get the information it needs? Liu: Actually, a component of our ARA system is a “Live Research Manager,” which basically is a faithful AI observer of your entire research progress. So you, the researcher, don’t need to do anything about documenting research knowledge. Everything you do is automatically observed and documented in this protocol. So, if you want to publish it in today’s format, a paper in PDF, it’s easy to convert back to a polished story. Checking for mistakes Large language models make errors. They hallucinate. So how will humans be able to check all the work the AI does in this protocol? Liu: A human being has limited bandwidth. So if you manually check all the code AIs generate, all the results, and all the analyses, that creates a bottleneck. [Instead the solution] is to use a formal system to objectively judge AI results. In other words, another layer of AI can easily supervise the process of the AI “scientists.” What prevents hallucinations and mistakes in that AI? Liu: I am working on a formal system using neurosymbolic techniques [that combine neural nets' use of unstructured data with symbolic AI's reliance on structures of logic and concepts]. That would guarantee that everything is rigorous. A language model alone, no matter how smart it is, has the chance to hallucinate because it’s a model based on probability, not logic. I want to make sure that I’m not using another language model to supervise the work done by an AI scientist. It would make every research paper a formal system, so that every claim can be written by a mathematical formula, and proved by the system. That makes all the claims in the system self-consistent. Getting rid of what you call the “narrative tax” means exposing mistakes, frustrations, or wrong turns to the world. What if researchers don’t want to do that? Liu: I think that’s certainly a big concern. People don’t want to be perceived as dumb. But I see that preference as an opportunity for AI. For example, if an AI does 12 hours of work that doesn’t lead anywhere, the human who is steering the project can jump in and say, “oh, AI, you’re dumb. You’ve made ABC mistake!” Then that is totally fine with people. They’re showing they’re very smart to supervise AI’s work. How long will humans have that steering role in AI research, though? Once you have AI supervising AI as you describe, will we reach a point where the AI doesn’t need human guidance? Liu: Yes, I think that’s just where a lot of AI research in new labs is heading. I recently wrote an article called “The End of Human-in-the-Loop,” which describes why I’ve come to think there will be this singularity point. Once AI has “squeezed out” all the expert data from humans, it won’t need any more input from humanity. That is the time AIs will start just self-evolving by themselves. Right now, the human is the bottleneck. The AI is always waiting for input from humans. But so at some point, AI will just do more autonomous work. If AI takes over so much scientific research, how will younger generations of human scientists get the experience and training they need to be able to steer future research, or even understand it? Liu: A lot of people have this idea that with AI doing so much work, nobody cares about trying to make the junior engineers and scientists better. I don’t agree. I think people will grow better by learning from AI. People’s learning curve is very fast with AI. So actually, I think it will be fine. We’ll still have senior researchers, senior engineers. But they will have had totally different learning experience than [earlier generations].

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：This report examines R&D waste and how AI adoption has outpaced the intelligence needed to make consequential decisions well. What Attendees will Learn Where R&D budget is lost. More than a third of organizations spend 25 to 40 percent of their R&D budget on projects that never reach market. Why projects fail late. Almost half of teams estimate over one million dollars in wasted investment for each project killed during development or testing. Why AI adoption has not closed the gap. Most organizations apply AI to execution tasks such as data analysis and modeling rather than to decision support. Where intelligence matters most. Respondents say better access to intelligence has the greatest value at early ideation and feasibility before significant investment is committed. Download this free whitepaper now!

IEEE Spectrum編輯Harry Goldstein回憶與資深作者Robert N. Charette超過20年的合作與友誼，從《為什麼軟件會失敗》到博客《風險因子》，再到多篇深度報道。Charette畢生致力於讓無形的軟件變得可見，他的核心理念是：已做出的假設就是已接受的風險。如今他退休，轉向自然攝影和小説創作。

德克薩斯大學奧斯汀分校的Lizy K. John教授開發了無權重神經網絡，通過查找表替代傳統乘法運算，在醫療監測、活動識別等任務中實現了千倍的能耗和體積縮減，並有望擴展至大型語言模型。

北卡羅來納中央大學副教授西奧班·戴·格雷迪於2025年1月創辦了首個在歷史黑人大學（HBCU）的人工智能研究所，旨在幫助師生掌握AI技能，應對勞動力市場變化。儘管HBCU面臨資金不足的挑戰，該研究所已通過谷歌100萬美元資助，吸引了2800多名參與者，並計劃擴大項目、建立計算機科學專業，打造可複製的AI教育框架。

人工智能的快速全球部署正在擴大數字鴻溝，將計算能力、技能和治理集中到少數地區，而許多國家仍處於被動消費狀態。南非和印度尼西亞的案例凸顯了實現更包容性AI發展的挑戰與可能性。

本文探討了模式敏捷威脅如何使傳統靜態庫雷達/電子戰系統失效，並介紹了AI/ML認知架構如何實現自適應實時對抗措施，涵蓋從威脅分類到閉環系統架構及HIL/SIL訓練驗證。

康奈爾科技的研究人員開發出一種光學接收器，利用光直接修改AI處理器的內存，有望降低數據中心、自動駕駛汽車和機器人等場景的能耗。該技術通過光電流翻轉SRAM中的比特，省去了耗電的模擬電路，實現了全數字光學通信。

NASA噴氣推進實驗室成功將谷歌Gemma 3大語言模型部署到太空，首次在軌演示了視覺語言模型分析衞星自身傳感器圖像的能力。該系統名為NAVI-Orbital，在Loft Orbital的YAM-9衞星上運行，僅需8GB內存即可在低功耗設備上執行任務，為衞星圖像分析帶來了範式轉變。通過語義壓縮，衞星可以傳輸文本摘要而非大量原始數據，有望將野火檢測等任務的延遲從90分鐘降至近乎實時。

現有AI基準測試衡量的是AI能做什麼，但沒有衡量AI是否按用户意圖行事。本文提出了一種新指標——精靈係數，用於量化用户指令與AI實際行為之間的差距，並借鑑經濟學中的基尼係數，將AI可能出現的“精靈式”行為分為狄俄尼索斯型和魔像型，呼籲建立相應基準。

Z.ai的GLM 5.2模型以低價和開放權重挑戰美國前沿AI模型，但許多程序員仍習慣使用昂貴模型，忽略成本。該模型在基準測試中接近Claude Opus 4.8，但實際使用效果參差不齊。

韓國研究人員利用生成式AI模型Generative SNUPI，能夠將用户繪製的形狀自動轉化為可摺疊的DNA納米結構，大幅簡化了傳統DNA摺紙技術繁瑣的設計過程，為納米機器人和醫療應用開闢新途徑。

研究員Dave Kuszmar發現了多個系統性漏洞，使他能夠繞過大型語言模型的安全限制，獲取危險指令。這些漏洞幾乎影響所有主流LLM，揭示了行業範圍的安全問題。Kuszmar呼籲減緩部署、提高透明度，並在進一步將LLM融入社會前開展大規模安全研究。

軟件工程類職位正受到人工智能的威脅。一些應聘者利用AI面試助手在遠程技術面試中實時生成回答進行反擊，而僱主則部署AI檢測工具來識別AI使用跡象。這種雙向動態使招聘演變成一場沒有明顯贏家的AI軍備競賽。專家認為，儘管面臨這一現實，求職中的人性化方面仍將佔據主導地位。

大型語言模型（LLM）在處理結構化數據方面存在困難，而新興的大型表格模型（LTM）專門為此設計。AI初創公司Fundamental推出了NEXUS模型，已獲亞馬遜雲服務採用，解決了表格數據分析的痛點，有望改變數據處理的未來。

研究表明，具備推理能力的大語言模型容易因邏輯不一致的提示而陷入“過度思考”，導致輸出長度激增，可能被利用發動拒絕服務攻擊。浙江大學與阿里巴巴的研究人員開發了一種進化算法，能夠生成惡意提示，使模型輸出長度最高增加26倍，影響包括DeepSeek-R1、Qwen3-Thinking、GPT-o3和Gemini 2.5 Flash在內的主流推理模型。

一位匿名藝術家通過展示一幅真實的莫奈畫作，揭露了人們對AI藝術的偏見。儘管爭議不斷，AI藝術市場正在形成，包括NFT和實體裝置。一位收藏家花費7.2萬美元收集早期AI作品。Refik Anadol在洛杉磯開設了全球首個生成式AI博物館Dataland，門票49-79美元，並銷售基於生物識別數據的商品。市場數據顯示，數字藝術銷售份額在2024-2025年間增長近三倍，但佳士得關閉了其數字藝術部門。斯坦福經濟學家發現，允許AI生成圖片後，一家圖庫平台月銷量增長80%。惠特尼博物館策展人強調，真正的AI藝術比繪畫更難，需要訓練定製模型和深度參與。

本文介紹了小型AI模型如何在資源受限的地區提供生命救助服務，例如通過手機本地運行AI模型識別假冒藥物、檢測作物疾病等。世界銀行正在積極推動小型AI的發展，儘管其仍面臨基礎設施挑戰。

人工智能基礎設施的迅速擴張不僅僅是能源消耗問題，其高密度、同步化的計算負荷正在改變電網的運行特性，導致需求快速波動，給電網運營商帶來新的挑戰。地理集中性加劇了局部可靠性問題，而現有監管框架難以適應這種新型負荷。

隨着人工智能驅動計算需求激增，能源系統面臨同步升級的緊迫挑戰。墨爾本憑藉其先進的能源生態系統、世界級工程研究和政產學研協同，成為全球能源轉型的標杆。本文探討了AI對能源基礎設施的影響、墨爾本在智能電網和可再生能源領域的創新，以及2027年IEEE PES GTD Asia會議將如何促進國際合作。

埃隆·馬斯克聲稱太空將是AI計算最便宜的地方，但IEEE Spectrum的分析揭示了軌道數據中心的巨大挑戰：散熱、衞星製造、發射成本等。分析師認為，即使樂觀估計，也需5-10年才能達到成本平價。

華盛頓大學計算語言學教授艾米麗·本德在《隨機鸚鵡》論文發表五週年之際，澄清了關於該隱喻的常見誤解，並討論了人工智能術語的問題、語言模型的能力侷限以及被忽視的剝削性勞動實踐。

研究人員在實驗中意外發現，單個CMOS晶體管可以模擬神經元和突觸的行為，這一發現可能大幅降低人工智能的能耗。

研究人員開發了 ConlangCrafter，一種能夠生成多樣且一致的構造語言（conlangs）的 AI 模型，性能超越通用大語言模型。該系統通過隨機變化和編輯循環創建遵循指定規則的語言，可免費在線獲取，有望用於自然語言處理研究和語言假設測試。

Prem Natarajan離開亞馬遜，成為第一資本的首席科學家，將深度AI研究應用於解決大規模現實金融挑戰，從欺詐檢測到代理客户服務。

本文探討了AI在數學領域的快速發展如何挑戰傳統數學研究，引發數學家對自身角色的深刻反思，並分析了三種可能的未來：AI作為工具、合作伙伴或自主研究者。

普林斯頓大學研究人員利用強化學習和逆向設計，從頭開始快速設計射頻集成電路（RFIC），生產出的芯片在創紀錄的時間內超越了人類設計的性能。AI生成的非常規佈局突破了性能極限，但該領域需要開放數據集以進一步推進。

人工智能自1956年達特茅斯會議正式確立以來，經歷了70年的發展，從早期的神經網絡、專家系統到現代深度學習、大語言模型和生成式AI。文章回顧了AI的演變歷程、優勢與風險，並強調了IEEE在推動AI進步與負責任使用方面的貢獻。