Despite US sanctions, Chinese tech companies demonstrated their latest AI advancements, emphasizing their ability to thrive in the sector at the World AI Conference in Shanghai
The World AI Conference (WAIC) 2024 concluded in Shanghai with groundbreaking updates in AI and technology. The event marked historic levels of attendance and digital engagement, showcasing the latest trends with over 300,000 onsite visits and 2 billion online views. Over 1,300 delegates from 50 countries discussed AI advancements, while 500 companies exhibited 1,500 products, including 50 world premieres. The ceremony saw the signing of 24 AI projects worth 40 billion yuan. Important launches included the Shanghai AI Security Governance Lab, and plans for improved 5G-A networks. The event underscored the global momentum towards integrating AI in various sectors, with Shanghai reinforcing its commitment to AI with plans to host WAIC 2025.
Shanghai authorities call for “reviewable and trustworthy” AI - The Shanghai Declaration on AI Governance, announced at the World Artificial Intelligence Conference, emphasizes the importance of creating AI systems that are accountable, transparent, and traceable. Highlighted by Shanghai's party secretary, Chen Jining, the statement aligns with the exhibit of cutting-edge advancements by tech giants and startups. It advocates for AI companies to adhere to local laws and promotes global collaboration in AI development under a philosophy of openness and sharing.
A Hacker Stole OpenAI Secrets, Raising Fears That China Could, Too - Last year, a hacker infiltrated OpenAI's internal communications but didn't access its AI development systems. The breach, disclosed internally and to the board in April 2023, wasn't announced publicly as no customer data was compromised. Despite no evident national security threat, concerns among employees grew about potential foreign espionage and OpenAI's security measures. Post-incident, an OpenAI technical program manager called for stricter safeguards to protect the company's AI technology from foreign theft, highlighting internal discord on AI risks and security protocols. The incident did not involve law enforcement notification.
Humanoid robots powered by AI turn heads at the World Artificial Intelligence Conference - The World Artificial Intelligence Conference in China is a hub for AI advancements, presenting a vast array of over a hundred AI products and innovations. Of particular interest are approximately 25 robots, including brand-new humanoid models that are capturing the attention of attendees. This event underscores the latest in robotic technology and AI's burgeoning presence.
NVIDIA will make over $12bn from China alone by selling AI chips, despite US sanctions - In the latest quarter, China's contribution to NVIDIA's total revenue fell to 9% from the previous year's 22%, yet revenue from China still grew by over 50% to $2.5 billion. NVIDIA is expected to achieve $12 billion in sales of AI chips, including the new H20 chip designed to comply with US export controls, doubling expected sales of Huawei's competing product. US restrictions aimed at preventing military use of advanced AI technology have affected Chinese tech giants, giving NVIDIA a competitive edge in AI chip sales. Despite the controls, robust demand for NVIDIA's H20 chip remains, priced between $12,000 - $13,000 per unit, indicating resilient performance and strategic agility within the tightly regulated market.
Canva CEO Melanie Perkins thinks the design world needs more alternatives to Adobe - Canva CEO and co-founder Melanie Perkins discusses the company's trajectory and future plans on The Verge's Decoder podcast. Canva, known for democratizing design, provides web-based design tools that cater to non-traditional design users such as small businesses and social media marketers. The platform recently pushed into the enterprise market with Canva Enterprise, aiming to capture a larger user base among organizations. Perkins also spoke about the company's culture, being Australia-based, and the balance between growth and maintaining company values. Canva's integration of AI tools reflects its commitment to ease and efficiency for users, with features like the background remover tool from the acquired company Kaleido. Despite concerns within the design community regarding AI's impact on creative work, Perkins emphasizes Canva's role in streamlining idea execution and not replacing creativity itself. Lastly, the acquisition of Affinity positions Canva as a stronger Adobe alternative for professional designers, expanding its reach and potential impact in the design software industry.
Google considered blocking Safari users from accessing its new AI features, report says - As part of an ongoing antitrust investigation by the Justice Department, Google's strategy to diminish its reliance on Apple's Safari browser has been highlighted. A key issue is a lucrative arrangement wherein Google pays Apple to be the iPhone's default search engine—costing upwards of $20 billion annually. Efforts to encourage iPhone users to prefer Google's or Chrome's apps for searches have seen modest gains but have not reached Google's aim of 50% by 2030. The preinstalled convenience of Safari impedes Google's progress. Newly hired executive Robby Stein is tasked with spearheading the transition towards Google's apps, possibly leveraging generative AI to enhance app appeal. Despite contemplation, Google chose not to restrict its AI Overviews feature to its apps, avoiding further alienation of Safari users. The outcome of the antitrust case, expected in the coming months, looms large over Google's strategy.
Telecoms companies now using scambaiting AI systems - Telecom companies have deployed advanced scambaiting AI systems to shield customers from burgeoning scam calls that cost Americans over $25 billion annually. These AI bots, akin to an enhanced version of the previously used 'Lenny', entertain and prolong conversations with fraudsters, effectively hamstringing their ability to target others. Scambaiting practices also gather intelligence on scam operations, helping to dismantle them. The AI system 'Apate', born from Macquarie University's research, not only engages scammers with sophisticated conversation but also learns their methods to assist in public awareness and law enforcement strategies. While engaging with scammers is left to these AIs, consumers are recommended to simply disconnect calls suspected of fraud.
Apple Intelligence and a better Siri may be coming to iPhones this spring - Apple is poised to upgrade Siri with the release of iOS 18.4 in spring 2025, incorporating the new Apple Intelligence AI system. Bloomberg’s Mark Gurman reports that the update will enable Siri to perform advanced tasks such as sorting through large photo libraries to retrieve specific items. The incremental iOS update will introduce these AI capabilities, though not all features, like AI-powered emoji generation or writing assistance, may be ready simultaneously. Also expected are a revamped Siri interface and enhanced product knowledge for basic tech support in the fall of 2023. Furthermore, future Apple AI developments could include the integration of AirPods with embedded cameras, capable of hand gesture and motion detection, similar to Meta's Ray-ban glasses, although this technology might still be a couple of generations away.
Stable Diffusion 3 License Revamped Amid Blowback, Promising Better Model - Stable Diffusion's licensing terms caused an upheaval among its users, leading to a revised Community License for the Stable Diffusion 3 (SD3) model from Stability AI. While the new license allows for free use under certain non-commercial and research criteria, as well as for businesses under $1 million in revenue, it prohibits the creation of new foundational models using SD3-generated images. The updated terms aim to maintain open-source principles and have removed previous restrictions on the number of generated media files. Despite these changes, skepticism remains within the AI community, with some criticizing the license for potential complications and the ease with which Stability AI can terminate the agreement.
The deluge of bonkers AI art is literally surreal - Facebook is often flooded with images that lean towards the surreal, ranging from absurdly edited animals to impossible scenarios, generating a blend of genuine and automated reactions. Khan Schoolcraft curates a Facebook group that compiles such striking AI-generated examples. Jonathan Gilmore, a philosophy professor, deems these images tech spam that flirt with surrealism, raising philosophical questions on AI's role in art creation. These images, while potentially intriguing, generally exploit emotional themes for engagement. Despite their lowbrow appeal, they prompt debate about authenticity and artistry in the digital age. Polina Kostanda, an AI artist, distinguishes her work by focusing on evocative prompts and photographic quality, moving beyond the "slop" to create meaningful AI art. The discourse is further complicated by perceptions of effort, originality, and the evolving definition of what constitutes art in the context of AI advancements.
AI industry needs to earn $600 billion per year to pay for massive hardware spend — fears of an AI bubble intensify in wake of Sequoia report - Despite large investments in AI by tech giants, AI's profitability is questionable, as analyst David Cahn suggests a $600 billion annual revenue need to sustain AI infrastructure costs. Nvidia reports significant datacenter hardware revenue, with Microsoft being a notable client. Yet, even with promising new Nvidia processors and OpenAI's revenue growth, there's a sizable gap in achievable AI revenue versus investment. The industry faces commoditization threats due to new entrants, which could drive intense competition and devalue existing technologies. Cahn warns of speculative investment risks and the necessity for sustained innovation to avoid an economic crisis from a potential AI investment bubble.
CIOs' concerns over generative AI echo those of the early days of cloud computing - At the MIT Sloan CIO Symposium, CIOs discussed generative AI, revealing concerns similar to those they had during the early days of cloud computing, including governance, security, and responsible use. Unlike the past, when IT departments often said no to new technologies, today's CIOs acknowledge that employees will use AI regardless of restrictions. They now focus on responsible use, enhancing customer experiences, and educating employees to create an AI-literate workforce. This shift in approach highlights a more realistic and proactive stance toward integrating new technologies safely and effectively.
For AI Giants, Smaller Is Sometimes Better - Tech companies are shifting focus from large AI models to smaller, more efficient ones to cut costs and improve performance. Smaller models, trained on specific tasks with less data, are cheaper to develop and operate, and often perform tasks nearly as well as larger models. Companies like Microsoft, Google, and Apple are releasing these models, which can run on devices without requiring massive cloud resources. This trend reflects the need for practical, cost-effective AI solutions as the pace of advancement in large models slows.
Nvidia AI Chip Supply Is a ‘Huge Bottleneck,’ EU’s Vestager Warns" - EU competition chief Margrethe Vestager highlighted a significant bottleneck in Nvidia's AI chip supply, which has drawn regulatory attention due to Nvidia's dominance in the market. While Nvidia's GPUs are crucial for AI model development, their high demand has led to supply challenges. Although the European watchdogs are still in the preliminary stages of their inquiry and have not yet decided on regulatory actions, Vestager noted that dominant companies might face future restrictions to ensure fair competition and innovation in the AI chip market.
Awesome Research Papers
A False Sense of Safety: Unsafe Information Leakage in 'Safe' AI Responses - Large Language Models (LLMs) are susceptible to methods that prompt them to produce harmful content, known as jailbreaks. Current safety mechanisms like output filters and alignment fine-tuning are inadequate, as they don't counter the crafty use of models to indirectly extract sensitive information or combine harmless outputs into dangerous ones. The paper introduces a novel threat model, inferential adversaries, who cleverly piece together information from separate outputs to serve harmful purposes. A new defense approach is suggested, involving an information censorship criterion to limit information leakage, highlighting an inevitable balance between safety and utility. This research provides a theoretical framework for developing LLMs that can be released with greater safety assurances.
MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases - This paper introduces MobileLLM, an efficient large language model (LLM) designed specifically for mobile devices with less than a billion parameters. The study challenges the notion that data and size are the only determinants of model quality, highlighting the importance of architecture. MobileLLM employs a deep, thin structure, embedding sharing, and grouped-query attention to outperform existing models of comparable size. Furthermore, MobileLLM-LS, which introduces immediate block-wise weight-sharing, delivers additional accuracy improvements with minimal latency impact. The MobileLLM family shows superior performance on chat benchmarks and competes closely with a much larger 7B model in API task accuracy, proving its effectiveness for on-device applications.
Mixture of A Million Experts - The paper introduces PEER (parameter efficient expert retrieval), a novel design for transformer models that employs a sparse mixture-of-experts approach. By utilizing the product key technique, PEER enables efficient retrieval from over a million tiny experts, overcoming limitations of existing MoE models. Experiments show that PEER layers significantly improve performance and computational efficiency in language modeling tasks compared to traditional dense feedforward layers and coarse-grained MoEs. This advancement allows for further scaling of transformer models while maintaining computational feasibility.
Data Curation via Joint Example Selection Further Accelerates Multimodal Learning - The paper introduces JEST, a method for selecting batches of data in multimodal learning, showing that joint example selection accelerates training more effectively than independent example selection. JEST leverages the dependencies in multimodal contrastive learning to create efficient batches, leading to significant reductions in training iterations and computational costs. By scoring batches based on learnability and utilizing model approximation techniques, JEST achieves up to 13 times fewer iterations and 10 times less computation, surpassing the performance of state-of-the-art models on various benchmarks.
SenseTime Unveils SenseNova 5.5 - At the WAIC 2024 AI Forum, SenseTime unveiled its SenseNova 5.5 Large Model featuring a significant upgrade from its predecessor, including China's first real-time multimodal model, SenseNova 5o. This update offers enhanced AI interaction capabilities akin to GPT-4o's streaming interaction model, with a 30% performance improvement and robust features like advanced reasoning and multilingual support. Additionally, SenseTime introduced cost-effective edge-side large models, reduced inference costs enabling widespread IoT deployment, and launched "Project $0 Go" for enterprises transitioning from OpenAI. New applications include Vimi, an AI avatar video generator, and upgrades to SenseTime's generative AI and productivity tools, advancing various sectors including finance, agriculture, and cultural tourism. Celebrating its 10th anniversary, SenseTime emphasizes these improvements as a step forward in multimodal large models, predicting transformative changes in human-AI interaction.
AI model unveiled for finance and customer service sectors - At the 2024 World Artificial Intelligence Conference in Shanghai, CLOOPEN, a Beijing-based provider, revealed the Rongxi Intelligent Model, a new AI foundation model tailored for the finance, marketing, and customer service sectors. This innovative range of AI solutions is designed to boost business operations, such as enhancing marketing strategies and improving service efficiency. Concurrently, Kupasi, a Shanghai tech leader, launched a high-quality data product for AI training to advance the application of AI across various industries. The event also introduced the "Corpus Operation Platform 1.0" to standardize the data supply chain and released industry-specific guidelines and white papers on corpus construction to standardize high-quality data sets for sectors like finance, education, and health.
Krea AI announces Scene Transfer - Krea introduces Scene Transfer, a feature designed to quickly produce new video scenes or transform existing ones with consistent lighting and color. Aimed at delivering professional-quality outcomes, it's highlighted for its seamless transitions, enhancing video production quality. Notably, this advanced tool is made available at no cost, widening access to high-end video production capabilities. Krea's Scene Transfer positions itself as an essential asset for effortless video project improvements, inviting users to explore its creative potential.
Life-Size Humanoid Robot With Dexterous Hands Unveiled at World AI Conference - China-based Humanoid Robots introduced Qinglong, a 5'9" general-purpose humanoid robot with dexterous hands, at WAIC 2024. Equipped with AI to gauge the necessary force for object manipulation, Qinglong demonstrated its ability to handle delicate items with precision. Developed by a Shanghai research lab designated to create a national humanoid robot innovation center, the robot features a bionic torso, 43 degrees of freedom for motion, and "athletic" capabilities. An annual release model following the Chinese zodiac has been planned. The lab further supports the field by launching OpenLoong, an open-source platform sharing Qinglong's hardware specs and forthcoming software, aiming to consolidate resources and advance embodied intelligence technology.
YouTube’s improved eraser tool easily removes copyrighted music from videos - YouTube has introduced a refined Erase Song feature, designed to help creators excise copyrighted music from their videos without disrupting other sounds. Leveraging an AI-powered algorithm, this enhanced tool aims for greater precision in identifying and stripping out copyrighted tunes. However, YouTube acknowledges that perfection isn't guaranteed and provides an alternative to mute all sound in instances where the tool struggles to isolate the music. This aims to assist creators in resolving content ID claims without removing the entire audio track.
Schneider Electric unveils its AI for GREEN strategy at WAIC 2024 - At WAIC 2024, Schneider Electric revealed its "AI for GREEN" strategy, highlighting the five key value propositions AI offers to businesses: Growth, Reliability, Efficiency, Environment, and New Horizon. The strategy emphasizes AI's role in driving industrial innovation and sustainability by enhancing energy and resource management, and potentially transforming business models. Schneider Electric showcased its EcoStruxure AI engine, which has been recognized with the Super AI Leader Award, demonstrating capabilities such as data management, model development, and application deployment.
AI-Driven Behavior Change Could Transform Health Care - OpenAI and Thrive Global are funding Thrive AI Health to develop a personalized AI health coach aimed at tackling the growing burden of chronic diseases, that account for 90% of America's $4.1 trillion health care spending. This mobile app leverages AI for hyper-personalization, understanding each individual's unique medical and behavioral data to offer tailored recommendations on sleep, diet, activity, stress, and social connections. Unlike generic health prompts, this AI coach will suggest specific, actionable lifestyle adjustments to facilitate healthier choices and address health inequities. The collaboration underscores the potential of AI to enhance health spans and lifespans by promoting healthy daily behaviors, with a focus on privacy and security in handling personal health data.
Microsoft Minference Demo on Hugging Face - MInference introduces a sparse calculation method to reduce the computational challenges of long-sequence LLM inference, which is hindered by the quadratic complexity of attention computation. By identifying efficient patterns in attention matrices, MInference significantly reduces latency in the pre-filling stage on GPUs without needing model modifications. It achieves up to 10 times faster inference while maintaining accuracy across various tasks and models.
Check Out My Other Videos:
Claude
