技能之巅，遇见全球匠心
Hy-MultiTurn: A Six-Dimensional Benchmark for Deep Multi-Turn Dialogue Understanding
How Children Design and Reason about Trustworthy AI Chatbots
At AI Day Singapore, NVIDIA and Partners Showcase AI Advancements Across Southeast Asia
NVIDIA AI Day Singapore, which takes place Sept. 22-23 at the Raffles City Convention Centre, is offering attendees opportunities to explore the hands-on training, expert-led sessions and advanced tools to accelerate their work in AI and high-performance computing.
At the event, NVIDIA and its partners are showcasing breakthrough AI advancements across the Southeast Asia region at large.
Read more about these announcements below.
NVIDIA Accelerates Public Sector AI from Pilot to Production in Southeast Asia 
AI is becoming a matter of national strategy, with governments looking to move from pilots to production and deliver impact at scale, while building trusted AI capabilities that reflect local languages, cultures, priorities and economic needs.
NVIDIA is working to enable all nations to be AI nations — providing the technology, infrastructure, ecosystem and expertise needed to make this possible.
To accelerate this transition across Southeast Asia, NVIDIA is helping nations move AI from experimentation to production-scale deployment through open models, developer tools and a broad partner ecosystem.
Together, NVIDIA and its partners are focusing on four key areas:
- Enhancing government operations and service delivery.
- Developing accessible AI-powered citizen services, and empowering local businesses.
- Strengthening critical infrastructure and public safety.
- Supporting startups, developers and researchers to strengthen national AI capabilities and innovation in each country.
Singapore’s HTX (Home Team Science and Technology Agency) is embarking on research using the NVIDIA Nemotron 3 Super and Nemotron 3 Nano Omni models to advance AI for public safety. Nemotron Super has the potential to support the agency’s complex reasoning and agentic workflows, while Omni’s unified vision, audio and language capabilities could help HTX develop multimodal applications grounded in real-world operational data. Together, the models could strengthen HTX’s ability to deploy secure, locally controlled AI across Singapore’s Home Team.
NCS is advancing agentic AI adoption across enterprises and the public sector, using Nemotron models and the NVIDIA Blueprint for video search and summarization (VSS), while advancing physical AI for practical humanoid robotics applications, to address security, responsiveness and data governance requirements. ST Engineering is using NVIDIA NeMo tools and NVIDIA cuOpt software to develop its AI Studio platform and deploy agentic AI solutions across its businesses such as Marine MRO.
Beyond Singapore, similar work is already underway across the region. Malaysia’s YTL AI Labs is fine-tuning Nemotron models for enterprise and citizen services, while Viettel AI is doing the same for Vietnamese-language applications.
In Thailand, the Big Data Institute and iApp Technology, as members of the ThaiLLM Collaboration, are exploring Nemotron as a foundation model. With an initial focus on legal applications, iApp Technology is adapting Nemotron 3 Nano by fine-tuning OpenThai 2.0 Legal with Thai-language legal data using the NVIDIA NeMo framework.
The model is released as open source for the Thai developer community and serves as the engine for Thanoy, the company’s legal-assistant chatbot, which already serves approximately 43,000 users.
In Brunei, Antrique built an AI innovation platform to help boost productivity across the nation’s food sector.
Across the region, NVIDIA Cosmos open world models and the NVIDIA VSS Blueprint are advancing smart city solution development. Malaysia’s ITMAX uses Cosmos with VSS to improve city traffic operations, while Thailand’s AS-TECH applies the same stack to improve passenger flow in airports.
Learn more about NVIDIA
Fact-Checkers Are Fighting Fire With Fire as AI Falsehoods Spread
AI did not invent the false claim. It gave the false claim a faster press — and it handed the fact-checker the same machine. In 2025, the Brazilian fact-checking organization Aos Fatos watched AI-generated content climb from 7% to 16% of the claims it investigated in a single year [1]. The same organization spent that year building its own AI tools. Nothing about the underlying discipline changed. What changed is the volume running through both sides.
That distinction — volume vs. substance. It is worth working through carefully, because the year Aos Fatos recently finished is a test case being played out in newsrooms globally.
What Actually Doubled?
Start with the number itself. Aos Fatos fact-checked 619 claims in 2025. 16% of them involved AI-generated content, up from 7% the year before — more than double in one year, driven by fabricated images and video rather than fabricated text. Reuters Institute researchers, reporting the figure at a 2026 industry gathering on AI and the future of news, framed it as one data point inside a larger shift: AI-powered disinformation in Brazil reached over 32 million views on TikTok in 2025, with 2.1 million additional likes and shares across Facebook and Instagram tied to the same wave of content [1].
Those are real numbers, and they describe a real change. Producing a convincing fake image at scale has become radically cheaper over the past two years. That is not in dispute, and nothing in this piece argues otherwise. Newsrooms that ignore this shift are negligent.
But notice what the number does not show. It does not show that falsehood itself got any easier to sustain once someone actually looked. A fabricated image is not true because it is well-made, any more than a forty-footnote report is more accurate than a four-footnote one. The image still fails the same test it always has: does it hold up when someone traces its origin and checks what it claims? Volume changed the odds that any single fake reaches a reader before it is checked. It did not change what checking finds when it happens.
Same Instrument, Opposite Verdicts
Here is the part of the story that gets less attention than it deserves. Aos Fatos did not just absorb the increase — it built against it. The organization developed Fátima, an AI chatbot that answers audience questions about claims already in its database, and is building a companion tool, Busca Fatos, for real-time verification during live coverage [1]. Two other fact-checking organizations cited at the same gathering, Spain's Maldita and the UK's Full Fact, have gone further: both run large language models that scan and classify claims across millions of sentences of public discourse, flagging the ones worth a human's time before a person ever sees them.
That is the same underlying technology on both sides. One version of it manufactures a fabricated image cheaply enough to flood a feed. Another version of it reads a million sentences overnight and hands a fact-checker the 12 worth investigating. Neither use case is more "natural" than the other. Technology has no preference for truth or falsehood; it accelerates whatever instructions and judgments sit behind it. A newsroom that trains a model to prioritize harmful narratives ends up with a filter. A bad actor who prompts the same class of model to generate a convincing fake gets a pollutant. The mechanism does not decide which one it becomes. The person operating it does.
As is my practice, I give credit where it is due. Maldita, Full Fact, and Aos Fatos are doing genuine, difficult, good-faith work, and the tools they have built represent real progress against a real problem. None of it would exist without practitioners choosing to point the same capability at detection instead of deception. That choice is the entire story — not the capability itself.
Who Still Has to Push Publish?
Chris Morris, CEO of Full Fact, and one of the panelists presenting the Aos Fatos figures, described the risk well: newsrooms are "in danger of getting to a place where no one believes anything they'd read or see or hear anywhere" [1]. That is the failure mode worth taking seriously, and it is not a technology failure. It is a discipline failure, and it can happen with or without AI in the loop.
Consider the two ways this plays out. In the first, a fact-checking team runs a Maldita-style classifier overnight, gets a shortlist of twelve claims worth investigating out of a million sentences, and still sends a person to trace each one back to a source, test the method behind it, and ask who benefits from it being believed. The tool changed how the team spent its morning. It did not change what counted as verification. In the second, a newsroom — or a reader, or a company evaluating a vendor's claim — sees an AI-generated shortlist, a citation, a polished report, and treats the artifact's existence as proof of the claim it contains. The tool did the same work in both cases. Only one of those newsrooms is still doing its job.
That test applies at any scale. A reader deciding whether to share a video, a fact-checking desk triaging a million sentences, a company evaluating a benchmark slide before it reaches the board — all three are answering the same question with the same three ingredients: is the origin traceable, does the reasoning hold up under a second look, and does whoever is asking benefit from the answer being believed. AI has made all three questions faster to ask and faster to dodge. It has not changed which one actually needs to be answered.
The Point
Aos Fatos's number will keep climbing. So will the sophistication of the tools built to answer it, on both sides of that fight. Betting that better detection technology eventually wins the race outright is a mistake — Maldita and Full Fact's own systems still hand the final judgment to a person, by design, because the technology was never built to make that call alone [1]. The organizations doing this well are not the ones with the most advanced models. They are the ones who never stopped treating the model's output as a lead to check rather than a verdict to publish.
That is not a hopeful prediction or a cautious hedge. It is the same principle discussed in previous articles … opened with and now tested against a full year of real data from a newsroom that lived through both sides of it at once: the questions we ask determine the answers we get, and no volume of AI-generated content — real or fabricated — changes who is responsible for asking them.
Citation(s)
Citation 1 — Reuters Institute for the Study of Journalism, "AI and the Future of News 2026: what we learned about its impact on newsrooms, fact-checking and news coverage." Primary source for all Aos Fatos figures (619 claims fact-checked in 2025; AI-generated content share rising from 7% to 16%; the 32 million TikTok views / 2.1 million Facebook–Instagram interactions figures), the Fátima and Busca Fatos tools, Maldita's and Full Fact's claim-classification systems, and the Chris Morris quote.
This UN week, Bill Gates is touting a new coalition that aims to make AI more inclusive
Bill Gates told his foundation’s Goalkeepers gathering that “increased generosity” and “the smart use of AI” together could accelerate the fight against inequality, urging the world’s richest countries to expand international aid and apply artificial intelligence toward what he considers “lifesaving innovations.”
“We failed to build in the right human values and ways of keeping our moral code controlling what these AIs do,” Gates, the foundation’s chair, said during a Monday conversation with political commentator Van Jones at Jazz at Lincoln Center in New York.
The Gates Foundation’s annual event, usually hosted the week of the United Nations’ high-level meetings, opened with a reminder: 26 of the world’s 34 wealthiest nations decreased development aid to poorer countries last year. Meanwhile, the U.N. has warned that progress on its sustainable development goals—lofty targets to reduce poverty, improve health and protect the environment by 2030—is “slow and uneven.”
The contraction comes at a time when Gates suggested “we have some of the best innovations ever to improve people’s lives.” His foundation spent much of its two-hour event elevating educators, farmers and health workers who are using AI to identify students’ learning gaps, combat pest infestations and better recall patients’ medical records.
The Gates Foundation also touted its new coalition of Anthropic, Google, OpenAI Foundation and dozens of other organizations working to make AI more accessible in underrepresented languages.
The partners—60 in total, according to Monday’s announcement—include frontier AI labs, corporations, philanthropies and others already working to expand the number of languages available in AI tools that the Gates Foundation believes could combat global inequality. The coalition aims to reach more than 3 billion people over five years by better coordinating existing efforts.
It’s work that must continue “full speed ahead,” according to Gates Foundation CEO Mark Suzman, even as some of the largest AI companies urge their industry to slow down development of advanced models. Governments must regulate the impacts on cybersecurity or children’s development, he said, while they extend AI’s humanitarian applications to poor communities currently “shut out” from the technology.
“Even if AI was frozen right now—which I don’t expect and I am not calling for—we would want to be building these language sets and making them usable with the tools that we have available right now,” Suzman told The Associated Press.
The collaboration follows last week’s Goalkeepers report, where the foundation committed $1 billion toward AI-focused efforts to improve health outcomes, upgrade educational tools and inform small farmers’ practices around the world. Unrepresentative language data could lead a model to mistranslate a pregnant Malawi woman saying her “water has broken” into the direct English that she’d “thrown away water,” the report warned.
The “original sin” is that many AI tools were trained on language data scraped from the internet, according to Mozilla Data Collective CEO E.M. Lewis-Jong, whose company was incubated by the Mozilla Foundation and is now a subsidiary of the not-for-profit Mozilla.org. The data-sharing platform seeks to help communities upload cultural and linguistic data sets on their own terms rather than have that data taken from the web without their explicit consent.
“The internet is not a representative space,” she said. “Why would you think that you were going to get a culturally diverse and representative system out of something that was predominantly trained on Reddit?”
Details such as the coalition’s governance are still getting hammered out, according to Suzman. A secretariat will track each signatory’s commitments and the Gates Foundation may nudge partners to fill larger gaps when necessary.
Google, a member of the coalition, has been funding an effort to collect more than 150,000 hours of audio across every district in India. Known as Project Vaani, the push underscores the need to gather speech data on dialects within languages, according to Google senior vice president James Manyika. To get there, he said, Google is working with local partners to record speech in the field.
Anthropic, whose CEO published an essay calling for industrywide cooperation on decelerating advancements, was already working with the foundation to speed up vaccine developments and bolster its chatbot’s data set of local crops. Elizabeth Kelly, the head of beneficial deployments at Anthropic, acknowledged the company’s products “lag in many African languages in particular.”
“We’re acutely aware that we can’t achieve any of the benefits we want to see in terms of improving patient outcomes or improving literacy and numeracy for students across the globe unless we actually get this language piece right,” she told AP.
Associated Press coverage of philanthropy and nonprofits receives support through the AP’s collaboration with The Conversation US, with funding from Lilly Endowment Inc. The AP is solely responsible for this content. For all of AP’s philanthropy coverage, visit https://apnews.com/hub/philanthropy.
—James Pollard, Associated Press
Amazon blocks Metas Muse AI agent from shopping
Amazon is reportedly blocking Meta's new AI agent Muse from shopping. As spotted by GeekWire, people attempting to use Muse to buy products from Amazon are now being served with a warning that doing so violates the online marketplace's terms of use.
Meta introduced its personal AI agent Muse earlier this month, billing it as a digital assistant that "doesn't just answer questions, it actually does the work."
"People just tell Muse what needs to get done, and it takes action, powered by Muse Spark, Meta’s most capable model to date, built for real-world agentic work like this," Meta wrote in its Sept. 8 announcement.
Such tasks include shopping, with Muse capable of paying for purchases using Link by Stripe. Meta further stated that it has plans to add Shop Pay as an alternate payment method (as well as 1Password support so you can give its AI the ability to use all your login credentials across all websites, if that was something you wanted to do).
However, at least for now, Amazon won't be one of Muse's preferred retailers. GeekWire reports that consumers trying to use Muse to shop on Amazon have recently begun to see a popup stopping them from doing so.
"Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed," the warning reads.
Amazon's Conditions of Use contains a section specifically relating to autonomous agents such as Muse, stating that they are barred from using its services unless they fulfil several transparency requirements. These include identifying themselves as a specific agent, not imitating or disguising themselves as humans, and not trying to circumvent Amazon's limits regarding how they interact with it. Amazon further reserves the right to limit or block agents' access at its own discretion.
"We think it's fairly straightforward that third-party applications that offer to make purchases on behalf of customers from other businesses should operate openly and respect service provider decisions about whether or not to participate," an Amazon spokesperson said in a statement to Mashable. "Agentic third-party applications such as Muse have the same obligations, and we've requested that Meta remove Amazon from the experience."
According to Amazon, Meta's Muse not only fails to identify itself, but also captures and stores Amazon users' credentials and scrapes their account data, which is a potential security concern. Meta further did not inform Amazon that Muse would access its storefront, and as such received no authorisation to do so.
Mashable has reached out to Meta for comment.
Amazon is currently beta testing its own AI shopping agent Buy for Me, which is capable of purchasing products on users' behalf on its own platform as well as from third-party vendors.
UPDATE: Sep. 23, 2026, 12:30 p.m. AEST This article has been updated with a statement from Amazon.
Big banks are worried AI agents could increase the risk of scams and fraud
While AI true believers worry about doom-and-gloom scenarios of AI destroying the world, some major financial institutions are sounding the alarm over a more grounded AI-related concern: scams.
On Tuesday, a consortium of big global banks put out a "principles paper" detailing their apprehensions regarding AI assistants as it relates to commerce. It also details principles that they would like to see followed as AI companies develop and continue to work on the technology.
Namely, the banks are concerned that AI agents empowered to act on behalf of customers will lose money to scams and fraud and, in turn, hurt merchants.
"As highly regulated financial entities, we are focused on managing risk effectively as emerging technologies arise," reads the opening of the paper. "Consumers are unclear if AI agents will act in their interests."
"They are concerned that AI agents may buy the wrong thing or spend too much – or even worse, lose their money to scams and fraud," the paper continues. "They are not sure whether they will be protected or who they will need to go to if things go wrong."
The paper, called Building Trust in Agentic Commerce, was published by Bank of America, Capital One, ASB Bank, Commonwealth Bank of Australia, ING Group, and NatWest Group.
Agentic commerce introduces new safety risks, with potential for higher rates of scams, fraud and disputes. There may be mismatched expectations between consumers, AI agents and merchants on what product or service should be provided, or when and how, or who would be liable if an AI agent exceeds its authority. Some providers may engage in unsafe practices, including requesting consumer card details and entering them directly into websites, prioritizing payment methods with lower protections, and not complying with payment processing standards and payment scheme rules. Malicious actors may attempt new attack vectors for scams and fraud, including compromising or impersonating AI agents and merchants, and engaging in new forms of social engineering.
The groups also voiced their concern for merchants and business owners, as the rise of AI shopping chatbots can likely lead to an uptick in credit card disputes and chargebacks due to the actions of the AI agent.
The paper outlines five principles — Transparency, Safety, Privacy & Data, Choice, and Interoperability — which the banks would like to see the AI industry follow.
For example, the financial institutions believe that the AI companies should prioritize consumer and merchant safety and consent when it comes to users' data and privacy. The banks would also like the AI companies to respect consumers' choices in the market without locking them in with restrictions.
The paper's release date certainly had interesting timing. It was released within 24 hours of the reported discovery of a serious zero-day vulnerability within Meta's Muse agentic AI assistant. The paper was also dropped within hours of Amazon's announcement that the e-commerce giant would be blocking Muse from making purchases on its platform.
UK cops arrest 2 EvilTokens suspects, Microsoft seizes 50 phishing kit websites
How to Use AI With Your Privacy Intact
A New Chatbot Wants to Unlock the Secrets in Tattered Ancient Greek Records
A New Tool Found Malware That’s Guided by an AI Hive Mind—No Humans in Sight
I Built AI Clones of My Coworkers. Things Got Weird
Microsoft disrupts AI-assisted platform that compromised 12,000 accounts
Microsoft said Tuesday that it led an industry-wide disruption of a subscription-based scam platform that used an AI chatbot to compromise 12,000 Microsoft accounts over a few-month span.
Named EvilTokens, the platform was introduced over a Telegram channel in February and charged an initial $1,500 fee and a recurring $500 charge each month after that. EvilTokens provided a single service for streamlining most steps required to compromise email accounts in large numbers. From there, the platform helped customers analyze inboxes, select targets that would provide the biggest potential payouts, and draft follow-up emails that provided realistic ruses for tricking company employees into transferring funds to attacker-controlled accounts.
Minutes, not days
“While EvilTokens helped cybercriminals access email accounts, at the center of the service was an AI-style chatbot that could analyze a victim’s inbox and help criminals identify trusted relationships, payment authorizations, and sensitive responsibilities, as well as other circumstances where fraud was most likely to succeed,” Microsoft said. “The platform could even recommend fraud strategies, including drafting messages that impersonated trusted contacts to help criminals trick victims into taking action.”
[程序员] 注册 meta 的 muse 智能体 使用 google gemini pro 的 spark [绕过 ip 问题与排队]
1.打开 gemini pro 的 spark-->输入:打开 https://muse.ai/join-->授予权限-->在 remote 浏览器中进行注册-->输入邮箱-->接受验证码-->要求验证年龄-->选择第一个验证方式(信用卡验证,可以使用国内 visa 信用卡.我用的是招行的 visa)-->注册成功.
进入正常的桌面端浏览器,使用刚刚的注册信息进入 muse.ai,然后点击左下角设置-->在通用-->使用情况中进行邀请码填写,我的邀请码是: DBADDI
使用邀请码我们各自可以获得 10 亿 token 额度,而且永不过期.
然后对话框的右上角有礼物按钮,点击后可以分享你们的邀请码,被使用同样可以获得 10 亿 token 额度
** 各位 v2exer,注册起来吧!
