By Adam Lidgett ( October 24, 2024, 9:45 PM EDT) — The Patent Trial and Appeal Board has fully thrown out an Orckit Corp. link aggregation patent and mostly invalidated another telecommunications patent following challenges from Cisco, finding their claims were too obvious to warrant patent protection…. Law360 is on it, so you are, too. A Law360 subscription puts you at the center of fast-moving legal issues, trends and developments so you can act with speed and confidence. Over 200 articles are published daily across more than 60 topics, industries, practice areas and jurisdictions. A Law360 subscription includes features such as Daily newsletters Expert analysis Mobile app Advanced search Judge information Real-time alerts 450K+ searchable archived articles And more! Experience Law360 today with a free 7-day trial. source

Younger generations and the growing influence of artificial intelligence are reshaping the business buying landscape. According to Forrester’s Buyers’ Journey Survey, 2024, over two-thirds of buyers involved in large and complex transactions valued over $1 million are Millennials and Generation Z buyers. This generational shift, combined with the rapid rise of generative AI (genAI), is fundamentally altering the way that business buyers make their purchases, and we can expect rapid changes in 2025. Younger buyers not only know that they can get access to more information than ever without directly engaging with a provider, but they also see providers consistently missing the mark in the buying cycle. They are hungering for more: real insight into providers and solutions; better ways to collect and analyze the information they find and receive from others; and, ultimately, a better buyer experience. They’re neither afraid to take advantage of the latest technology to improve their buying experience nor are they afraid to walk away if they don’t get what they want. Three key predictions highlight the changes that we can expect in business purchasing in 2025: GenAI will drive buyers to consider five or more providers for large purchases. GenAI is revolutionizing the way business buyers approach their purchase decisions. AI tools not only expedite the buying process but also enhance overall business outcomes for organizations. Over 90% of buyers who used genAI to inform purchases of $1 million or more reported positive results. These tools enable buyers to conduct extensive research, reduce biases, and evaluate a wider range of providers. As buyers become more adept researchers, we expect them to continue to expand their consideration set and include more providers because they can analyze and assess them with greater ease. Dissatisfaction will drive two-thirds of business buyers to seek new solutions. Younger buyers are highly dissatisfied with their buying experience. From technical implementation issues to concerns about diversity and inclusion programs, these buyers are pushing providers to meet their technical requirements while engaging them as valued partners. Younger buyers know that they have alternatives and are willing to evaluate them if they can replace their current solution. Half of younger buyers will include 10 or more external influencers in their purchase. As the influence of Millennials and Generation Z buyers continues to grow, they increasingly rely on external sources, including their value network, to make decisions. Today, almost one-third of younger buyers bring in 10 or more individuals outside of their organization to the decision-making process. These include online community members and industry conference attendees. Social media platforms, which give access to a host of new influencers, already rank among the top three preferred interaction types among young buyers, and their influence continues to grow. Marketers And Sellers Must Take Control Where They Can Old ways of doing business are no longer sufficient — marketers and sellers need to meet this new generation of buyers where they are. Marketers should identify the digital signals that buyers give off so that they can best leverage the increasingly rare, direct, and personal interactions that buyers may have with them. They’ll also need to improve both their hard and soft skills to engage with customers who are dissatisfied and willing to go back on the market to seek out better options. Finally, marketers will need to extend their outreach to include influencers, especially those on social media, who are engaging with buyers and delivering insight. Read our full Predictions 2025: Business Buyers report to learn more about these predictions and two additional ones. Schedule a Forrester guidance session to discuss these predictions and learn how you can prepare your sales, marketing, and product teams to adapt their strategies. If you aren’t a client, you can download our complimentary B2B Predictions guide, which covers all of our top predictions for 2025. Additional resources, including webinars, are available on the Predictions 2025 hub. source

The 2024 presidential election will surely have far-reaching consequences in many areas — and artificial intelligence is no exception. EY’s latest technology pulse poll, published in October, revealed that 74% of 503 tech leaders expect the election to impact AI regulation and global competitiveness. Although tech leaders said they plan to significantly increase AI investments in the next year, future growth of AI may hinge on the outcome of the election. Respondents believe the outcome of the election will mostly impact regulation related to cybersecurity/data protections, AI and machine learning, and user data and content oversight. “Of course, all of these are closely tied to innovation, growth and global competitiveness,’’ James Brundage, EY global & Americas technology sector leader, told TechRepublic. “The U.S. is the world’s tech innovation leader, so future tech policy should strike a balance that supports U.S. innovation while establishing guardrails where they are needed,” such as in data privacy, children’s online safety, and national security. SEE: Year-round IT budget template (TechRepublic Premium) Greater investments in AI Notably, tech companies will continue to make significant investments in AI regardless of the outcome of the presidential election, according to the survey. However, the result may impact the direction of fiscal, tax, tariff, antitrust, and regulatory policies as well as interest rates, mergers and acquisitions, initial public offerings, and AI regulations, the survey said. “We were surprised that trade/tariffs were not higher up on the minds of these executives,’’ Brundage observed. On the heels of a sluggish tech market in 2024, he said that “the 2025 trajectory is bullish, as companies focus on raising capital to invest in growth and emerging technologies like AI.” The majority of tech leaders (82%) said their company plans to increase AI investments by 50% or more in the next year. In the next year, AI investments will focus on key areas including AI-specific talent (60%), cybersecurity (49%), and back-office functions (45%). With an eye on innovation, most tech industry leaders surveyed also plan to allocate resources toward AI investments in the next six to 12 months, with 78% of tech leaders reporting their company is considering divesting non-core assets or businesses as part of their growth strategy during that time. More must-read AI coverage Big organizations struggling with AI initiatives Brundage also finds it surprising that 63% of tech leaders report their organization’s AI initiatives have successfully moved to the implementation phase. “That number seems high, but several factors could explain it,’’ he noted. “First, companies may be focusing on short-term, low-hanging fruit AI projects, which are easier to implement, have higher success rates, but may not be the opportunities with maximum impact.” Further, use of “quick-buy solutions like ChatGPT or Copilot, which are relatively simple to deploy and drive productivity, may inflate this percentage.” Also, successful implementation “likely means moving from proof of concept (POC) to implementation,” Brundage said, adding that “real challenges such as data quality, scaling, governance, and infrastructure still lie ahead.” Additionally, size matters — the report observed that organizations with more employees are finding less success moving AI initiatives to the implementation phase. Data quality issues (40%) and talent/skills shortages (34%) are the most common reasons for AI initiatives failing to progress to the next stage, according to those who indicated that fewer than half of their AI initiatives have been implemented successfully. How the election’s impact on AI could be felt Regardless of who takes office in 2025, there could be a continuation of current regulatory and enforcement trends related to AI given that the Federal Trade Commission and Department of Justice have been very active and may remain so, Brundage said. Given that “some legislative proposals are bipartisan … we expect that they will advance in 2025 or 2026,” such as children’s online safety. But he pointed out that state legislatures and attorneys general also impact policy, “so it’s a nuanced playing field. We expect these changes to be measured in years, not months.” Tech leaders must realize that the U.S. is experiencing a new geopolitical environment compared with five to 10 years ago, Brundage said. “New government industrial policy in the U.S. and around the globe is driving business action — both in the tech sector and in the industries and supply chains that it relies upon. These global tech businesses are particularly at the forefront of geopolitics as countries seek to de-risk from one another.” AI capabilities have also become highly competitive and geopolitically significant across the globe, he said. “There is a dual race to innovate and regulate here in the U.S. and elsewhere. We see a need to have business models that account for the different regulatory approaches like sovereign frontier models.” Wanted: AI tech talent search intensifies As organizations continue to integrate more AI functionality into their businesses, the need to hire AI-specific talent will increase, as well as the need to restructure or reduce headcount from legacy job functions, according to the survey. Eighty percent of tech leader respondents foresee reducing or restructuring headcount from legacy functions to other in-demand functions, and 77% anticipate an increase in hiring for AI-specific talent, according to the survey. Additionally, 40% of technology leaders said human capital efforts such as training will be the focus of their company’s AI investments next year. AI’s impact on national security and foreign policy Meanwhile, the Biden administration on Thursday released the first-ever AI-focused national security memorandum (NSM) to ensure that the U.S. continues to lead in the development and deployment of AI technologies. The memorandum also prioritizes how the country adopts and uses AI while preserving privacy, human rights, civil rights, and civil liberties so the technology can be trusted. The NSM also calls for the creation of a governance and risk management framework for how agencies implement AI and requiring them to monitor, assess, and mitigate AI risks related to those issues. source

“Welcome on board. I have been tasked with taking you to Mäntsälä — in the middle of nowhere,” the minivan driver greets us in the characteristic clear and unhurried intonations of a Finnish native speaker. Mäntsälä is, indeed, in the middle of nowhere. But this kind of location is often where you find collections of some of the most powerful machines of today, humming away behind doors along unpretentious corridors. This includes Nebius’ AI data centre, taking shape in the small community an hour’s drive or so north of Helsinki. Amsterdam-based Nebius is labelling itself an AI cloud infrastructure company. Its proprietary platform, it says, has been optimised for AI training and inference without performance bottlenecks. “The AI cloud is different from the ‘regular’ cloud. In the set of tools, in the applications, yes, but the people who use it are also different,” says Nebius’ head of product and infrastructure, Andrey Korolenko. The would-be European AI infrastructure force has begun amassing a tremendous amount of compute. Today, it announced that it will triple the Mäntsälä capacity to up to 60,000 Nvidia GPUs. Precisely, this entails deployment of Nvidia’s H200 GPUs, available from November, in addition to already installed H100s. Nebius is also one of the launch partners for Nvidia’s upcoming Blackwell platform. The Blackwell GB200 will enter mass production in December. Korolenko states that whenever the first unit ships (currently slated for Q1 2025), Nebius will also have it “in a matter of weeks.” Filling gaps in the AI training and inference market While others are toiling away to close the gap to Nvidia, the latter’s hardware (along with its CUDA platform) is still the gold standard for AI training and inference. Nebius’ core business is to offer time on its GPUs to everyone from app developers and companies optimising foundation models for their own businesses to AI model tuners and builders, from pre-training to inference, with different levels of support for different levels of skill. The company has 400+ engineers in its employ, and customers already include the likes of Mistral AI, Genesis Therapeutics, Recraft, and Jetbrains. Nebius has built custom racks for its Nvidia hardware. “We are doing it [building the data centre] from the ground up,” Korolenko says. “If you build it, you can just adjust it,” he adds, addressing the evolving nature of chips along with all the infrastructure requirements this entails. During our visit, a large batch of racks has just arrived from the manufacturer in Taiwan, and Nebius crew are in the process of unpacking them from their boxes. At the mention of Taiwan, red flags concerning supply chains immediately pop up, and it is not long before the question of “What will happen to you in the event of a ‘reunification’ in the South China Sea?” arises. “What will happen to the world in this case?” Korolenko muses, sounding rather stoic. Nebius is busy setting up new server halls to host its GPUs. Credit: Linnea Ahlgren/TNW The data centre also houses ISEG, which currently ranks as the 19th fastest supercomputer in the world, and the fastest commercially available in Europe. With 35.26 GFlops per watt-second, it has also made it to number 24 on the Green Top500 list. One billion USD in European AI infrastructure The build-out of Mäntsälä is part of a plan to splurge a total of $1bn across Europe by mid-2025, including opening an additional three data centres across the continent (as well as one in an as-of-yet undecided location in North America). This includes a recent addition in Paris, a colocation deployment based at Equinix’s PA10 campus. The site, located in the Saint-Denis district, has an urban farm on the roof, and heat from its servers was used to heat the Olympic pool during the 2024 summer games. At its Mäntsälä data centre, Nebius provides heating for about 2,000 homes. A feature that, if replicable, makes the company’s presence an attractive proposition for other remote locations eager to up their green credentials. Nebius’ data centre currently only utilises air to cool its servers. This will change with the addition of the later GPU models. What will not change is the amount of power that will go back towards heating the neighbouring town — the percentage, approximately 70%, may even increase when deploying a liquid cooling system. In fact, with the expansion, Nebius will export more heat than the village of Mäntsälä requires. The Yandex legacy You may wonder how come you have not heard of a company with 400 engineers able to get a hold of tens of thousands of highly coveted Nvidia GPUs up until now. Nebius recently emerged from the European remnants of Yandex (which had a long-term relationship with the GPU maker) following the company’s high-profile divestment from Russia. One of the things to come out of that laborious process, other than the Mäntsälä data centre, was a few billion dollars in cash. This is currently fueling the rollout of what could end up being a global force in AI infrastructure. As a result of the legacy from Yandex, Nebius Group is listed on NASDAQ. The group also encompasses data business Toloka, upskilling edtech platform TripleTen, and autonomous driving technology unit Avride. However, its shares are not currently trading. Earlier this month, it announced it had enlisted Goldman Sachs as its financial advisor with a view to recommence trading further down the line. Yandex founder Arkady Volozh is Nebius’ CEO, who has publicly criticised Russia’s war in Ukraine. However, along with the Mäntsälä data centre, Nebius has likely also inherited a certain degree of suspicion due to its origins. The company has moved its employees out of Russia, essentially evacuating thousands of people including whole families to locations outside of the country in 2022. The company does not allow its employees to work from inside Russia should they go back and visit. It has also had to go through rigorous vetting from EU authorities to receive approval for the divestment deal. Its execs, who had to

In many different areas, talent is an important yet often elusive goal. Just ask anyone whose piano keyboard skills have never moved beyond pecking out the first few measures of “Heart and Soul.” When it comes to generative AI, large language models (LLMs), trained on massive quantities of data, supply the capabilities needed to drive multiple use cases and applications, as well as handle an almost endless array of tasks. To get the most out of generative AI, think of it as a tool rather than a replacement, suggests Daniel Wu, an AI research fellow at Stanford University in an email interview. He notes that LLMs can already do great work. “They’re being used in coding assistance and customer service, but they work best with clear prompting.” Every organization produces large amounts of text as part of its normal business operations, observes Manfred Kügel, data scientist and IoT industry advisor for AI and analytics provider SAS, via email. Before LLMs, organizations needed to perform complex text analytics in order to get value out of unstructured text data, such as maintenance records or shift logs in a production environment. “LLMs can be used to structure text data and prepare it as inputs for machine learning models used for production optimization and predictive maintenance.” Related:How Generative AI Is Changing the Nature of Cyber Insurance Pushing it to the Max To gain maximum value from generative AI, users need to clearly define their problems and objectives, says Kevin Ameche, president of ERP software provider RealSteel, in an email interview. “Identify-specific use cases, such as content generation, data analysis, or automation,” he advises. “Then, ensure you can access high-quality data for training the AI model.” Ameche recommends collaborating with internal or external AI experts to fine-tune and customize their model to align with specific needs. “Continuously evaluate and refine the model’s performance and stay updated with the latest advancements in generative AI technology to maximize its potential for your organization.” To maximize generative AI’s value, users should first understand its inherent capabilities and shortcomings, Kügel says. “We are still in the early days of realizing the full potential of generative AI,” he states. Kügel believes that everyone involved in core business processes should interact with models in the same way they interact with their colleagues. “This will drive quick adoption and encourage organizations to provide the necessary and user-friendly generative AI tools to overcome any structural or cultural hurdles.” Related:Can Generative AI and Data Quality Coexist? Achieving Effectiveness Generative AI’s effectiveness lies in its ability to automate creative processes, generate content, and provide data-driven insights at scale, Ameche explains. “It can handle repetitive tasks, freeing-up human resources for more strategic work.” Meanwhile, the technology’s adaptability and capacity to learn from data make it a valuable tool in various industries. An AI agent can’t read minds. “If you ask a poorly defined question, you’ll get one of any number of valid responses,” Wu says. “But by giving AI a stronger sense of what you’re searching for, either through clear prompts, data, or even model fine-tuning, you’ll get more useful responses.” To empower team members, organizations should invest in generative AI training and development programs, Ameche says. “Start by identifying the specific skills and knowledge needed for working with AI,” he recommends. “Consider partnering with AI vendors or educational institutions for tailored training.” Ameche believes that it’s also important to encourage employees to experiment with AI tools in real-world projects to gain hands-on experience. “Create an environment of continuous learning and provide access to resources, such as online courses, webinars, and AI communities,” he suggests. “Collaboration and knowledge sharing within the team can also accelerate the learning process, helping team members harness the maximum value from generative AI.” Related:10 IT Trends to Watch for This Year Common Mistakes Wu notes there’s a common saying in AI research: junk in, junk out. “Users may inadvertently harm their projects by providing biased datasets or creating poor prompts,” he explains. “Model outputs should always be taken with a hint of salt,” Wu recommends. Both over- and underestimating generative AI’s potential is a serious concern, Kügel says. “So is seeing AI as a threat when an AI model produced insights that we didn’t see ourselves.” As with any breakthrough technology, Kügel sees skepticism among many IT leaders. He highlights that it’s important to clearly show that generative AI augments and supports, rather than replaces, human experts. He recommends taking a balanced approach to AI adoption by deploying guardrails and plausibility checks. “The model should report on its own when it drifts too far from reality,” Kügel says. Final Thought Generative AI holds immense potential for enterprises across many domains, Ameche says. “However, successful implementation requires careful planning, ongoing training, and vigilance to avoid pitfalls.” He believes that organizations should view generative AI as a tool to augment human capabilities, not as a replacement. “When used strategically and responsibly, generative AI can transform efficiency, creativity, and decision-making, driving innovation and competitive advantage.” source

By Christopher Cole ( October 25, 2024, 8:35 PM EDT) — The Federal Communications Commission said Friday that AT&T has agreed to pay almost $2.3 million to resolve claims it broke the rules for two major federal broadband subsidy programs…. Law360 is on it, so you are, too. A Law360 subscription puts you at the center of fast-moving legal issues, trends and developments so you can act with speed and confidence. Over 200 articles are published daily across more than 60 topics, industries, practice areas and jurisdictions. A Law360 subscription includes features such as Daily newsletters Expert analysis Mobile app Advanced search Judge information Real-time alerts 450K+ searchable archived articles And more! Experience Law360 today with a free 7-day trial. source

Threat actors accessed the private health information of more than 100 million people in the February breach of Change Healthcare — the largest-ever health care data breach reported to federal regulators — the U.S. Office for Civil Rights revealed on Oct. 22. The hack, information about which was revealed in June, could affect up to one-third of Americans. It has proven to be one of the most significant cyberattacks of the year and shows how ransomed data can lead to physical harms such as belated delivery of essential medication. SEE: Nation-state attackers may search for “target-rich, cyber-poor” organizations like public infrastructure or health care, said CISA advisor Nicole Perlroth. What was the Change Healthcare cyberattack? In February, UnitedHealth Group, the parent company of Change Healthcare, found out that an attacker had introduced ransomware into Change Healthcare’s systems. The group ALPHV, sometimes called BlackCat, claimed responsibility for the breach. By March, Change Healthcare had determined attackers accessed their systems from Feb. 17 to 20. The company brought in “leading cybersecurity and data analysis experts,” Mandiant personnel among them, and obtained a copy of the stolen records, analyzing the dataset. United Healthcare released a more thorough accounting of the incident in April. In a Senate hearing on the matter in May, UnitedHealth Group CEO Andrew Witty said the company had paid a ransom of $22 million in Bitcoin to release the stolen data. Cybersecurity experts don’t recommend paying ransoms because it rewards threat actors, can cause significant financial harm to the business, and does not guarantee the return of the data. The U.S. government has considered the controversial idea of banning ransom payments. Change Healthcare said it can’t specify what data has been affected for each individual. In general, the stolen data included: First and last name, address, date of birth, phone number, and email. Health information such as diagnoses, medical record numbers, images, and test results. Billing, claims, and payment information Other personal information that may be associated with medical records, such as Social Security numbers, driver’s licenses or state ID numbers, or passport numbers. Full medical histories or doctors’ charts have not been found among the stolen data. The attack delayed prescription deliveries and led to a business disruption impact of $705 million. Overall, Change Healthcare’s financial outlook for next year is lower than expected. Must-read security coverage Change Healthcare offers resources for affected customers United Healthcare says their investigation of the attack is still ongoing but in its final stages. The company is still sending notifications to those affected. Change Healthcare offers two years of complimentary credit monitoring and identity theft protection services from IDX to eligible customers. They provided “trained clinicians to provide emotional support services” through a dedicated call center. The call center cannot provide information about what specific data may have been exposed from individual accounts. United Healthcare recommends impacted patients monitor their bank accounts and medical insurance statements. Unusual activity should be reported to their financial institution or health care provider as appropriate. Ransomware attacks on health care have far-reaching consequences Cyberattacks on health care data are a perfect storm of potentially lucrative random opportunities for threat actors and heightened mistrust among affected customers. Patients can lose access to necessary medications and care can be delayed if operations are disrupted. In May, a ransomware attack at hospital system Ascension slowed down care. Around the same time, the U.S. Advanced Research Projects Agency for Health announced its intention to invest more than $50 million in tools for information technology professionals in hospital settings to improve their cybersecurity. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Improving the capabilities of large language models (LLMs) in retrieving in-prompt information remains an area of active research that can impact important applications such as retrieval-augmented generation (RAG) and in-context learning (ICL). Microsoft Research and Tsinghua University researchers have introduced Differential Transformer (Diff Transformer), a new LLM architecture that improves performance by amplifying attention to relevant context while filtering out noise. Their findings, published in a research paper, show that Diff Transformer outperforms the classic Transformer architecture in various settings. Transformers and the “lost-in-the-middle” phenomenon The Transformer architecture is the foundation of most modern LLMs. It uses an attention mechanism to weigh the importance of different parts of the input sequence when generating output. The attention mechanism employs the softmax function, which normalizes a vector of values into a probability distribution. In Transformers, the softmax function assigns attention scores to different tokens in the input sequence. However, studies have shown that Transformers struggle to retrieve key information from long contexts. “We began by investigating the so-called ‘lost-in-the-middle’ phenomenon,” Furu Wei, Partner Research Manager at Microsoft Research, told VentureBeat, referring to previous research findings that showed that LLMs “do not robustly make use of information in long input contexts” and that “performance significantly degrades when models must access relevant information in the middle of long contexts.” Wei and his colleagues also observed that some LLM hallucinations, where the model produces incorrect outputs despite having relevant context information, correlate with spurious attention patterns. “For example, large language models are easily distracted by context,” Wei said. “We analyzed the attention patterns and found that the Transformer attention tends to over-attend irrelevant context because of the softmax bottleneck.” The softmax function used in Transformer’s attention mechanism tends to distribute attention scores across all tokens, even those that are not relevant to the task. This can cause the model to lose focus on the most important parts of the input, especially in long contexts. “Previous studies indicate that the softmax attention has a bias to learn low-frequency signals because the softmax attention scores are restricted to positive values and have to be summed to 1,” Wei said. “The theoretical bottleneck renders [it] such that the classic Transformer cannot learn sparse attention distributions. In other words, the attention scores tend to flatten rather than focusing on relevant context.” Differential Transformer Differential Transformer (source: arXiv) To address this limitation, the researchers developed the Diff Transformer, a new foundation architecture for LLMs. The core idea is to use a “differential attention” mechanism that cancels out noise and amplifies the attention given to the most relevant parts of the input. The Transformer uses three vectors to compute attention: query, key, and value. The classic attention mechanism performs the softmax function on the entire query and key vectors. The proposed differential attention works by partitioning the query and key vectors into two groups and computing two separate softmax attention maps. The difference between these two maps is then used as the attention score. This process eliminates common noise, encouraging the model to focus on information that is pertinent to the input. The researchers compare their approach to noise-canceling headphones or differential amplifiers in electrical engineering, where the difference between two signals cancels out common-mode noise. While Diff Transformer involves an additional subtraction operation compared to the classic Transformer, it maintains efficiency thanks to parallelization and optimization techniques. “In the experimental setup, we matched the number of parameters and FLOPs with Transformers,” Wei said. “Because the basic operator is still softmax, it can also benefit from the widely used FlashAttention cuda kernels for acceleration.” In retrospect, the method used in Diff Transformer seems like a simple and intuitive solution. Wei compares it to ResNet, a popular deep learning architecture that introduced “residual connections” to improve the training of very deep neural networks. Residual connections made a very simple change to the traditional architecture yet had a profound impact. “In research, the key is to figure out ‘what is the right problem?’” Wei said. “Once we can ask the right question, the solution is often intuitive. Similar to ResNet, the residual connection is an addition, compared with the subtraction in Diff Transformer, so it wasn’t immediately apparent for researchers to propose the idea.” Diff Transformer in action The researchers evaluated Diff Transformer on various language modeling tasks, scaling it up in terms of model size (from 3 billion to 13 billion parameters), training tokens, and context length (up to 64,000 tokens). Their experiments showed that Diff Transformer consistently outperforms the classic Transformer architecture across different benchmarks. A 3-billion-parameter Diff Transformer trained on 1 trillion tokens showed consistent improvements of several percentage points compared to similarly sized Transformer models. Further experiments with different model sizes and training dataset sizes confirmed the scalability of Diff Transformer. Their findings suggest that in general, Diff Transformer requires only around 65% of the model size or training tokens needed by a classic Transformer to achieve comparable performance. The Diff Transformer is more efficient than the classic Transformer in terms of both parameters and train tokens (source: arXiv) The researchers also found that Diff Transformer is particularly effective in using increasing context lengths. It showed significant improvements in key information retrieval, hallucination mitigation, and in-context learning. While the initial results are promising, there’s still room for improvement. The research team is working on scaling Diff Transformer to larger model sizes and training datasets. They also plan to extend it to other modalities, including image, audio, video, and multimodal data. The researchers have released the code for Diff Transformer, implemented with different attention and optimization mechanisms. They believe the architecture can help improve performance across various LLM applications. “As the model can attend to relevant context more accurately, it is expected that these language models can better understand the context information with less in-context hallucinations,” Wei said. “For example, for the retrieval-augmented generation settings (such as Bing Chat, Perplexity,

Marketers have juggled many unknowns in the last year, including media chapters that were left unfinished. There is the looming ban on TikTok in the US (or is there?) and the mad dash to gain generative AI-based efficiencies. And let’s not forget Google’s third-party-cookie pivot and the ensuing question of whether it helps or hurts marketers. In 2025, we’ll see many of these chapters close and marketers gain confidence as a result. But don’t get too comfortable: Just as some stories wind down, others will ramp up (Google antitrust cases, anyone?). To help you prepare for the coming year, Forrester predicts that: Privacy Sandbox will erode as Google bulldozes the Topics API and Attribution Reporting API. Google faces pressure on multiple fronts: antitrust scrutiny, nonviable Privacy Sandbox tests, and marketers long believing that Google was bluffing on third-party cookie deprecation (only 13% of global B2C marketing decision-makers believed that Google would actually go ahead with it, Forrester survey data shows). These pressures will force the company to abandon the Topics API and Attribution Reporting API, two hallmarks of its Privacy Sandbox. Doubters don’t believe that a Google-built, -owned, and -self-measured solution will actually work. Despite the inevitable disintegration of Privacy Sandbox, marketers should still test alternative targeting and measurement methods. Marketers will shift 10% of their performance media budgets into social commerce. While social commerce is already established in APAC, other regions have been slower to adopt the behavior. 2025 will be the year when social commerce becomes a staple within media plans. This will be a sharp turn given that in 2021, 60% of online adults in the UK, 51% in the US, and 46% in France hadn’t even seen a shoppable ad on social media. TikTok Shop changed the narrative with a massively improved user experience. Integrated payment and retail systems such as Apple Pay and Shop have also made checkout and tracking experiences on social media platforms significantly more seamless. Consumers will embrace social media channels as viable shopping platforms, making these channels critical for marketers with goods to sell. TikTok will not get banned or divest in the US in 2025. Despite intense pressure from the US government to divest TikTok from Chinese-owned ByteDance, the immensely popular platform isn’t going anywhere anytime soon. Aside from the fact that a ban would be a political land mine for either Kamala Harris or Donald Trump, TikTok’s enormous investment in lobbying and legal fees would push the appeals process beyond 2025. But should our prediction be wrong, Meta and YouTube stand to gain both consumers and media dollars. Forrester survey data indicates that if TikTok gets banned in the US, TikTok users would turn to Reels and Shorts instead — which is exactly what happened when India banned TikTok in 2020. Continue investing in this high-performing media channel and build a multidimensional creator strategy to support it. Although it might feel like things are settling down, don’t be fooled. Marketers should stay observant and adaptable in 2025. The new media chapter that we’re entering is bound to keep you on your toes. Read our full Predictions 2025: Media And Advertising report to get more detail about each of these predictions, plus two more bonus predictions. Set up a Forrester guidance session to discuss these predictions or plan out your 2025 media and communications strategy. If you aren’t yet a client, you can download our complimentary Predictions 2025 guide for B2C marketing and customer experience professionals, which includes more of our top predictions for 2024. Get additional complimentary resources, including webinars, on the Predictions 2025 hub. source

港市早段摘要 大型新盤推售帶動一手樓交投 ; 美國逾10,000億基建案仍存分歧 港股美國買賣的預託證券造價較本港收市低 滙豐收市折合45.89元—-比本港收市跌1元3毫1仙中國人壽15.33元——–跌2毫3仙騰訊控股600.69元——-跌2元3毫1仙長和58.99元————跌6毫1仙港交所464.87元———跌4元9毫3仙中國石油股份3.46元—–跌1毫2仙中國石油化工3.93元—–跌4仙工商銀行5元————跌5仙建設銀行6.16元———跌4仙友邦保險94.55元——–跌6毫半 中國汽車工業協會料今年汽車銷量較原先估計好 中國汽車工業協會副總工程師判斷，內地今年汽車銷量可能增長 6.5 % ，增速較年初預測 4 % 大，推算相當近2700萬輛，乘用車銷量估計增長 10 % ，新能源車可能升 46 % ，至超過200萬輛。 大型新盤推售帶動一手樓交投 新盤成交近 300 宗，過去周末二手成交略為放慢 。 大型新盤帶動一手樓交投，過去周末新盤成交超過290宗，集中在新世界大圍站上蓋柏傲莊III，第三輪推售173伙，收到超過30,000張票，成為歷來收票最多的新盤，單日出售近170伙，其中有買家以超過5,400萬元購入3個分別509和797呎單位，平均呎價超過25,000元。 何文田萬科VAU Residence過去周末開售55伙，單日只剩下1伙，發展商隨即加推17伙，介乎209至388呎，預計周三推售。 人行主管《 金融時報 》發文強調市場不應過分擔心資金面 人行主管的《 金融時報 》發文，強調貨幣政策 ” 穩字當頭 ” ，保持流動性合理充裕不是空話，所以市場毋須對流動性，產生不必要的擔憂，更不宜沒有根據地猜測，流動性會收緊和波動，誤導市場預期，人為製造波動，建議外界應該更多關注政策利率的水平，而非單日逆回購等市場操作的規模。人行亦已通過《貨幣政策執行報告 》等渠道多次強調，將引導貨幣市場利率圍繞人行短期政策利率波動運行，以及要多關注人行政策利率水平而不是操作量 。 美國逾10,000億基建案仍存分歧 逾 10,000 億美元基建方案白宮和議員仍有分歧。國際金融協會憂供應樽頸和通脹持續 。 美國總統拜登推動20,000 億美元基建計劃，遭佔參議院半數議席共和黨反對，上周 21 名共和、民主兩黨議員提出 12,000 億美元的反建議，但雙方在如何提供資金仍然有分歧。 “華爾街日報” 消息指，白宮反對向電動車徵費，建議主力加強執法、打擊避稅、瞞稅等提升稅收。拜登周一會正式研究議員方案，另一位共和黨參議員格雷厄姆形容，跨黨派支持的基建方案已完成，拜登要決定是否接受。 即使加強投資，亦面對供應短缺的樽頸，因為疫情影響供應鏈，亦缺乏足夠勞工重返崗位，國際金融協會發表報告，預計問題會持續至明年，估計觸發物價持續上升。 阿克曼SPAC入股環球音樂 阿克曼SPAC作價約40億美元入股環球音樂 阿克曼的SPAC同意收購環球音樂 10 % 股權，作價約 40億美元。有關SPAC表明，會繼續尋找新業務併購。雖然一般SPAC是與收購對象合併，但今次公司還餘下約15億美元，還有阿克曼旗下基金等承諾提供多14億美元資金。 ……………………………………………………..????????讚好Facebook 專頁：https://www.facebook.com/VeriMedia.io……………………………………………………..更多資訊:https://veri-media.io……………………………………………………. LinkedIn Email Facebook Twitter WhatsApp source