Рубрика: AI

Huawei unveiled its newest smartphone, the Mate 60 Pro, on August 30, 2023. While this may seem like just another smartphone launch, the Mate 60 Pro carries significant geopolitical implications. The latest smartphone by Huawei is powered by a 7 nm processor, manufactured by state-owned ​​Semiconductor Manufacturing International Corporation (SMIC).

The recent development has made many in the West jittery and question whether US sanctions to cripple China’s chip production have really worked? Ironically, the launch also aligned with the US Secretary of Commerce Gina Marie Raimondo’s visit to the country. Notably, the US Department of Commerce is responsible for the majority of the sanctions imposed on China.

China bypasses sanctions

In 2018, when Donald Trump was still the POTUS, the US blacklisted Huawei and leaned on its allies to do the same on the ground of espionage. Furthermore, on May 15, 2019, the Department of Commerce included Huawei and other Chinese firms in its ‘Entity List’, citing allegations of unlawful trade practices and resulting in legal indictments against the company.

However, despite short-term setbacks, Huawei, from being a telecom equipment company, has emerged as a leading AI company in China. Besides building Large Language Models (LLMs), Huawei has developed an analog GPU that is comparable to the NVIDIA A100 GPU. Considering the current GPU shortage and China’s robust manufacturing capabilities, China may indeed offer a potential solution to alleviate this shortage. Moreover, the Huawei’s Mate 60 Pro launch only solidifies China’s claims that the US sanctions have been a complete failure.

Moreover, with pressure from the US, Netherland-based chip-making equipment maker ASML is also limiting the sale of its highly sophisticated chip-making machines to China. Almost every chip in use today, be it in automobiles, smartphones, or laptops, are being developed using ASML’s equipment. ASML’s TWINSCAN NXE:3600D is currently one of its most advanced EUV lithography systems, supporting volume production of chips. However, reportedly, China is in the process of making its own AI chip-making equipment.

Shanghai Micro Electronics Equipment (SMEE), a leading lithography machine maker, is working to deliver its first system based on 28 nm technology later this year. Moreover, in 2020, local media reported that a Chinese research institute had made a breakthrough in a new type of 5 nm laser lithography technology. This could mean China could produce advanced chip-making equipment within this decade.

AI war

The US-China trade conflict, which evolved from a trade war to a cold war, and then transitioned into a tech war, is now manifesting as an AI battleground, spanning nearly half a decade. In 2021, Raimondo also stated that the US needs to work with Europe to slow China’s innovation rate. As per reports, the US government is contemplating additional export controls targeting AI hardware made by the likes of NVIDIA, which play a vital role in AI model training and data centre operations.

Today, both China and the US are locking horns for AI supremacy and much of the sanctions on China are to curb Chinese innovation in AI. But, China does not appear to be much behind the US. Huawei, the company making GPUs and AI chips, is also working on LLMs to compete with the likes of OpenAI, Microsoft and Google. Other Chinese companies such as Tencent and Alibaba have already developed their own LLM.

Moreover, despite the sanctions, Chinese companies have been finding a way one way or the other to get their hands on GPUs. The Financial Times reported that Chinese companies are accessing high-end US chips through intermediaries, revealing potential loopholes in Washington’s strategy to curb the sale or transfer of AI hardware to China.

SenseTime, a firm blacklisted by the US, was employing intermediaries to circumvent export controls. Reports also suggest that certain Chinese AI companies are leveraging NVIDIA’s processors via cloud servers located in different countries to navigate the restrictions. Not only this, chipmakers such as AMD and NVIDIA are building modified versions of their products and shipping them to the Chinese market, which, undoubtedly, is one of their biggest. This further validates China’s point that the sanctions, indeed, have been a failure.

Are sanctions actually backfiring?

Given the US is restricting NVIDIA to sell its advanced AI hardware to China, it will heavily impact the Santa Clara-headquartered company’s revenue given China is one of their most important markets. “Over the long term, restrictions prohibiting the sale of our datacentre graphic processing units to China, if implemented, would result in a permanent loss of opportunities for the US industry to compete and lead in one of the world’s largest markets and impact on our future business and financial results,” said Colette Kress, NVIDIA’s chief financial officer.

Moreover, the US significantly puts pressure on its allies to align their policies and impose similar sanctions on targeted countries, thereby increasing the collective impact of economic and political pressure. But, on the contrary, it could be pushing its allies away towards China, which projects itself as a friend for the global south. Hence Huawei’s substantial growth, China’s economic growth, advancement in AI, and semiconductors as well as China’s global influence are indicators that the US sanctions on China have been ineffective.

The post US Sanctions to Cripple Chinese Innovation Backfires appeared first on Analytics India Magazine.

How would you feel if you spent months working on a research and when you try to publish it for the world, the people in authority pull it down because of some “anonymity” policy? This same heartbreaking and aggravating situation has been happening for some time with researchers submitting their papers to the Association of Computational Linguistics (ACL) conference because of its review and citation policy.

Naomi Saphra, PhD researcher at University of Edinburgh recently took to X to share her experience with this policy. “Just got a desk rejection, post-rebuttals, for a paper being submitted to arXiv <30 min late for the anonymity deadline,” she said. Sapra posted this after the ACL conference rejected her because she was just half an hour late from changing the anonymity, which means removing her name from the paper, which was published on open-access archive platform, arXiv.

It is the first paper that the lead, @ZackAnkner, ever submitted. I told him that EMNLP would be better than an ML venue because of the short paper option, which is a great feature for a conference. But increasingly, like many others, I don’t consider ACL venues worth the hassle.

— Naomi Saphra (@nsaphra) September 4, 2023

Vitthal Bhandari also shared the same reason for getting his paper rejected, which he submitted in May for the ACL conference. Undoubtedly, people on X say that this is making the ACL venue less and less desirable for NLP work.

just got a reject, post-rebuttals, for a paper i submitted in may. talking about the ACL anonymity policy hurting junior researchers, since i submitted pretty close to the deadline, i couldn't upload my work on arxiv. now i have nothing to show for my grad apps 🙁 https://t.co/0y9wGYkxH0 pic.twitter.com/wrFCWER7Dx

— Vitthal Bhandari (@Vit_Bhandari) September 5, 2023

Sebastian Raschka from LightningAI replied to the post, “Wait, first they don’t allow people to talk about their arXiv manuscripts online and now they also restrict when someone can and cannot upload to arxiv?”

It seems like ACL is saying that your paper is a very valuable contribution for the scientific community, “but since you published it on another website 30 minutes after us, your work is not worthy to appear in our prestigious venue.”

Let’s get this clear that ACL believes that if you want to publish your paper for its conference, you have to be anonymous, get your paper reviewed, and not publish on any other website such as arXiv, as long as the committee tells you to. ACL wants to have complete control over the science and research community. This whole conversation is directed against arXiv.

Ethicists getting jealous

ACL is not the only platform that is against arXiv. The conversation actually gets interesting as soon as Emily M Bender, a professor and member of the Distributed Artificial Intelligence Research Institute (DAIR), which is founded by Timnit Gebru, decided to put in her thoughts. Bender said, “arXiv is a cancer that promotes dissemination of “junk science” in a format indistinguishable from real publication..”

arXiv is a cancer that promotes the dissemination of junk "science" in a format that is indistinguishable from real publications. And promotes the hectic "can't keep up" + "anything older than 6 months is irrelevant" CS culture.
>>

— @emilymbender@dair-community.social on Mastodon (@emilymbender) August 29, 2023

Obviously, this take from an academic researcher was not received well by the whole research community. “Coming from a person at a position of considerable academic power, this is an unbelievably careless & elitist statement that goes against a lot of what (I hope) you believe in terms of equity & level playing field for people outside of prestigious institutions,” replied Stanislav Fort to Bender.

Members of the DAIR community have been the most vocal when talking about making AI more ethical and democratising the field by making it more inclusive. Interestingly, this comment is antithetical to the “inclusive philosophy” of their organisation.

It seems that the organisation holds itself on a higher pedestal and aims to decide who can actually contribute to open source. Moreover, the organisation wishes to promote peer reviewed publications, but do not realise it actually kills the open source spirit and affects the junior researchers the most.

To prove the point and rebut, Bender took on the medium and published a blog – Scholarship should be open, inclusive and slow, where admittedly she discusses a few points and merits about the policy she proposes. “Peer review, when it is working well, doesn’t guarantee truth or correctness, but it does mean: This was examined critically and thoughtfully by 2–4 independent people with relevant knowledge who found it sufficiently solid (given their own knowledge) and worthy of others’ attention.”

She further tries to explain how sloppy work by famous researchers often gets promoted because it is done by famous researchers, which shouldn’t be the case. Further, bringing anonymity levels the playing field for outside researchers. This seems like a fair point but later, continuing with the elitist view of being part of the ACL’s Executive Committee, and next year’s President, Bender says that arXiv promotes biases, while also calling research papers that would not be relevant in a few months, as “not all that interesting.”

A simple interface, free access, no frills, what’s not to love about arxiv!?

— Sebastian Raschka (@rasbt) September 5, 2023

Research is about reproducibility, not only reviewing

There is merit in the conversation that anyone publishing on a website which is believed to be the truest source of information for science needs some assessment. But arXiv has been handling it pretty well by having a single URL all this while post publication for as much peer review as possible. Moreover, pushing anonymity and embargoing papers that need to be reviewed by an “expert committee” is contrary to the principles of open and inclusive science.

“It sounds like a gatekeeper who is mad that people found a path to publishing that bypasses their gate,” said Thomas Steinke.

There is not even an argument against the fact that arXiv has contributed more to the science and engineering community and research more than any other publishing portal. The open access nature of the website is possibly pushing every field in the best forward direction that there is. Boaz Barak, a theoretical computer science professor, said that, “I’ve said it before, arXiv has done much more to advance science, and expand participation in it, than all the anonymity interventions ever will..”

Even Yann LeCun, the Meta AI chief, agrees that any policy that obstructs arXiv is just silly.

"Any policy that obstructs ArXiv is just silly."
100% https://t.co/R026bF8ajh

— Yann LeCun (@ylecun) September 5, 2023

All in all, it seems like ACL and DAIR are just afraid of how much arXiv is praised, as being the best platform for scientific research, and they only want to be the gatekeeper of the research, obsessed with control, more than making it better, not actually making it inclusive. There are valid concerns about the publishing culture for sure, but arXiv is not the culprit, thus not needing ACL or DAIR’s opinion.

The post arXiv Doesn’t Need Ethicists’ Opinion appeared first on Analytics India Magazine.

At the B20 Summit last week, IBM CEO, Arvind Krishna asked Finance Minister Nirmala Sitharaman for advice and encouragement for multinational companies who wish to have a strong presence in India and from India to serve the world.

Taking a dig, Sitharaman added that everyone knows that AatmaNirbhar is already in progress, and “IBM is probably outside of that realm.”

While IBM might be waiting for favourable policies or tax slabs to invest heavily in AI research and development in the country, it has been doing its bit to upskill and foster growth at the grassroots level, upskilling children of the country.

Three years ago, CBSE, in partnership with IBM, introduced an AI curriculum for high school students in Grades XI & XII in 2020. This initiative is part of CBSE’s SEWA program and has been implemented in about 200 schools across 13 Indian states, including Karnataka, Tamil Nadu, Uttar Pradesh and others.

The curriculum covers AI basics, history, and applications, along with skills like design thinking, data fluency, and critical thinking. It also emphasises ethical decision-making and bias awareness in AI. Developed with Macquarie University and local partners like Learning Links Foundation and 1M1B, this program aims to equip students with AI knowledge and skills to build AI models for real-life use cases.

Manav Subodh, the founder of 1M1B, a non-profit management company, told AIM, that they want to help students practise and do things that would benefit their careers ahead. They implement curriculums related to AI, data science, AR/VR and data safety, amongst others—-and have partnered with IBM, Meta, Adobe and others to avail their programs in over 3,000 schools across India.

“We train people as per the National Education Policy 2020 standards. We make it project-based. The core problem besides tech within the course is, that, we are imparting problem-solving (abilities) using technology. Because there is no point in studying technology for technology’s sake, technology has to solve problems,” he added.

Further, he said that technology has to empower people and have an impact on the community. To implement this. “We have skilling programs and have curriculums which are implemented in skills from 6th standard onwards,” he explained. The company claimed that it conducts skilling programs and olympiads like ‘Future Tech Olympiad’ and ‘AI Startup School’ to foster prototype development and build critical workforce skills.

AIM recently visited one such event in Bangalore last week, organised by 1M1B, which saw a slew of students, parents, teachers, industry representatives and policymakers gather. At the event, students showcased the prototypes they built for agriculture, linguistics, and e-commerce use cases.

Showing the prototype, a class X student named Swaminathan told AIM that he had developed a model to identify the crop which would generate maximum yield based on updated data on soil health and condition from government portals.

Besides schools and colleges, on 6 September, IBM also renewed its research collaboration with IIT Bombay, and IISc Bangalore, to advance innovations in hybrid cloud and AI. This partnership, part of IBM’s AI Horizon Network, aims to tap into the intellectual talent of students, faculty, and researchers to address global challenges. The collaboration will focus on areas like natural language processing, generative AI, machine learning for time series data, fake news detection, hybrid cloud optimization, and sustainable computing.

IIT Bombay joined IBM’s AI Horizon Network in 2018 for AI research, while in 2021, IBM and IISc Bangalore launched the IBM-IISc Hybrid Cloud lab for hybrid cloud technology research and innovation.

What About Others?

Adobe, in collaboration with India’s Ministry of Education, this week launched a creative and digital literacy initiative called Adobe Express. Announced during the ongoing G-20 Summit, the program will impact 20 million students & 500,000 educators in India by 2027. AI-powered content creation app Adobe Express will be central to this effort, and the company will offer a curriculum that covers creativity, generative AI, design, animation, and emerging tech, aiming to empower students with future-ready skills.

Meta has also entered a three-year partnership with India’s Ministry of Education and Ministry of Skill Development and Entrepreneurship. This initiative aims to equip students, educators, and entrepreneurs with digital skills and training. Meta will work with organisations like NIESBUD, AICTE, and CBSE to provide digital marketing skills, diploma courses, and training in AR, VR, AI, and Digital Citizenship.

Additionally, Meta’s ongoing partnership with CBSE, which began in December 2021, aims to train 10 million students and 1 million educators in AR, VR, AI, and Digital Citizenship until 2026.

In August, the Directorate General of Training announced a collaboration with AWS in India to upskill students in cloud computing, data annotation, AI, and machine learning.

In 2018 Microsoft also launched Intelligent Cloud Hubs aimed at providing students with upskilling opportunities in AI and cloud technologies.

Google also partnered with CBSE, to train 1 million teachers by the end of 2020. Meanwhile, Google DeepMind is also doing something similar in the UK and Europe region partnering with Raspberry Pie to launch ‘Experience AI’—a learning programme to introduce AI/ML concepts and promote critical thinking. However, the programme is yet to be available in India.

These developments indicate that a lot of public partnership has been helping to upskill the Indian students. However, while all these initiatives are beneficial for the country and the ecosystem, the country needs additional support in the form of direct investment in research and AI development.

The post IBM Catches India’s AI Talent Young appeared first on Analytics India Magazine.

Amid the fear of AI replacing jobs, Andrew Ng, the founder of DeepLearning.AI has been instrumental in AI literacy. Back in his iconic pale blue shirt, he added six new short courses that cover the current AI topics with a fresh set of partnerships, including Microsoft and Google. And the best part? These courses are offered free of charge (unless you want a certificate) and can typically be completed within one to two hours, making them easily accessible and time-efficient. Let’s delve into the details of these courses.

Large Language Models with Semantic Search

In partnership with Canadian startup Cohere, this course is tailored for beginners with a basic grasp of Python, this course imparts valuable skills in augmenting keyword search using Cohere Rerank. It delves into the realm of advanced search techniques, instructing learners on seamlessly integrating LLMs into search systems.

The course introduces the concept of dense retrieval, a powerful NLP tool that leverages embeddings to elevate the relevance of search results beyond conventional keyword-based methods. Participants will also gain insight into the intelligent reranking process, infusing LLMs’ intelligence into search systems for heightened efficiency and faster response times.

Upon course completion, you will have a comprehensive idea about foundational principles of keyword search, transforming search systems using innovative reranking methods, embedding-based semantic understanding, and practical hands-on experiences. The course instructors, Jay Alammar and Luis Serrano from Cohere, ensure a comprehensive education in integrating language model-driven search functionalities into websites and projects to enhance user engagement and interactions

Finetuning Large Language Models

“Finetuning Large Language Models” is a short course, led by Sharon Zhou, the co-founder and chief executive of Lamini, who has also instructed in the GANs Specialisation and How Diffusion Models Work. Upon course completion, participants will understand when to employ finetuning techniques on LLMs, adeptly prepare data for this purpose, and successfully train and assess LLMs on their datasets. Finetuning, a central focus of the course, enables individuals to customise LLMs using their own data, allowing for the adaptation of model weights and thus differentiation from alternative methods like prompt engineering and retrieval augmented generation. This process equips the model to acquire style, form, and incorporate new knowledge to enhance overall performance. Ideal candidates for this course possess Python proficiency and a solid understanding of deep learning frameworks, particularly PyTorch.

Building Generative AI Applications with Gradio

Ng has teamed up with Hugging Face to offer this new, concise course for beginners and will be instructed by Apolinário Passos, an ML Art Engineer at Hugging Face. Participants will delve into a range of tasks, such as image generation, image captioning, and text summarisation, using Gradio, an open-source Python library.

Gradio empowers rapid development of user-friendly and adaptable UI components for machine learning models or APIs, enabling the creation of user-friendly applications even for those without coding expertise. Through Gradio, individuals can effortlessly construct intuitive graphical elements to interact with their models or APIs, ensuring accessibility and customization for users. By the course’s end, you will have acquired practical skills for developing interactive apps and demos, streamlining project validation and implementation.

Evaluating and Debugging Generative AI Models Using Weights and Biases

This course equips individuals with the skills to evaluate programs using LLMs and generative image models with platform-independent tools. Participants will discover how to instrument a training notebook, incorporating essential elements such as tracking, versioning, and logging. Moreover, they will gain expertise in monitoring and tracing LLMs’ performance over time in complex interactions. The course addresses the challenges of managing data sources, extensive data volumes, model development, parameter tuning, and experimentation in the realm of machine learning and AI projects. By introducing learners to Weights & Biases platform tools, this course simplifies experiment tracking, data running, and collaboration within a team.

Key lessons cover Jupyter notebook instrumentation, hyperparameter configuration management, run metric logging, dataset and model versioning artifact collection, and experiment result logging. As a result, participants will develop a structured workflow that enhances productivity and expedites progress toward groundbreaking outcomes. Suitable for those with Python and PyTorch familiarity and an interest in streamlining, versioning, and debugging their machine learning workflow.

The instructor, Carey Phelps, is the Founding Product Manager at Weights & Biases, bringing a wealth of expertise to guide learners through this transformative course.

Understanding and Applying Text Embeddings with Vertex AI

In this beginner-level course lasting one hour, instructed by Nikita Namjoshi, a Developer Advocate for Generative AI at Google Cloud and Ng, participants can access it for free for a limited time. The course focuses on leveraging text embeddings to capture the essence of sentences and paragraphs. You’ll learn how to utilize these numerical text representations for tasks like text clustering, classification, and identifying outliers.

Furthermore, the course delves into constructing a question-answering system using Google Cloud’s Vertex AI. Participants will gain insights into word and sentence embeddings, measuring semantic similarity, text generation adjustments, and efficient semantic search using the ScaNN library. Upon completion, learners will have a solid grasp of text embeddings and their integration into Language Model applications. Basic Python knowledge is the only prerequisite for joining.

How Business Thinkers Can Start Building AI Plugins With Semantic Kernel

In this latest course with Microsoft, you can delve into the world of Microsoft‘s open-source orchestrator, the Semantic Kernel. Offered for free for a limited time, learners can enhance their business planning and analysis skills while harnessing the potential of AI tools. Throughout the course, students will advance their knowledge of LLMs, exploring techniques like using memories, connectors, chains, and more. Participants will be equipped to create sophisticated business applications using LLMs, effectively leverage LLM building blocks, and integrate the Semantic Kernel into their applications, streamlining AI service interactions without the need to learn multiple APIs. This course is suitable for anyone interested in learning Semantic Kernel, with basic Python skills and an understanding of APIs recommended but not mandatory. Instructor John Maeda, VP of Design and Artificial Intelligence at Microsoft, guides learners through this invaluable learning journey.

Read more: Key Highlights of Google Cloud Next ‘23

The post 6 Brilliant New Free Courses by Andrew Ng on Generative AI appeared first on Analytics India Magazine.

Microsoft and other investors have poured $110 million into d-Matrix, an artificial intelligence chip company, Reuters revealed on Tuesday. d-Matrix is remarkable because it focuses on chips for inference. Put simply, AI inference is the process of improving the accuracy of a generative AI or large language model’s predictions. It occurs after training.

Support for inference gives d-Matrix a valuable niche and avoids competition with NVIDIA, the wide-ranging technology company that makes GPUs and system-on-chip units, among other software and hardware.

Jump to:

• What is d-Matrix?
• What is d-Matrix building?
• Why d-Matrix stands out among the AI chip landscape
• How chiplets fit into the global chip shortage

What is d-Matrix?

d-Matrix is a Silicon Valley-based company that produces compute platforms (chips) for generative AI and large language models. Its flagship product is Corsair, an in-memory compute engine for AI inference. The design’s ability to hold an AI model entirely in-memory is novel and builds on d-Matrix’s previous Nighthawk, Jayhawk-I and Jayhawk II chiplets.

What is d-Matrix building?

With the new round of funding, d-Matrix will work on commercializing Corsair. It wants to fix the problem of AI and LLM companies not having enough compute power to run the workloads they need. To solve this memory bottleneck, d-Matrix made chiplet-based Digital Memory In Compute platforms that can, d-Matrix says, reduce the total cost of ownership of the inference process.

Corsair is expected to launch next year, in 2024.

Why d-Matrix stands out among the AI chip landscape

d-Matrix stands out because chip-making is competitive, and many smaller companies are having trouble finding funding. NVIDIA has pressured many smaller companies and investors out of the AI chip market. In particular, NVIDIA’s dominance in both hardware and software makes it hard for other companies to squeeze in, Reuters said.

NVIDIA declined to comment on the investments in d-Matrix.

The $110 million investment in d-Matrix comes from a Series B funding round from investment firms Temasek and Playground Global as well as M12, Microsoft’s venture capital fund. Prior to this, d-Matrix had raised $44 million in a funding round with Playground Global.

“The current trajectory of AI compute is unsustainable as the TCO to run AI inference is escalating rapidly,” said Sid Sheth, cofounder and CEO at d-Matrix, in a press release. “The team at d-Matrix is changing the cost economics of deploying AI inference with a compute solution purpose-built for LLMs, and this round of funding validates our position in the industry.”

“D-Matrix is the company that will make generative AI commercially viable,” Sasha Ostojic, partner at Playground Global, stated in the same press release.

“We’re entering the production phase when LLM inference TCO becomes a critical factor in how much, where, and when enterprises use advanced AI in their services and applications,” said Michael Stewart from M12, Microsoft’s Venture Fund, in the press release.

How chiplets fit into the global chip shortage

The generative AI industry, which has taken off in leaps and bounds since the commercialization of ChatGPT in November 2022, faces two major problems today. First, running generative AI is extremely costly — training an LLM costs as much as $4 million as of March 2023.

Second, graphics processing units, which are required for AI training and which NVIDIA produces, can still be hard to find. They’re so short in supply that countries around the world are starting initiatives to boost the chip industry. For example, in early September, China put $40 billion toward its chip industry; although, there’s no indication that those chips aren’t specifically targeting generative AI or LLM products.

SEE: Here’s everything you need to know about the chip shortage, including why it started. (TechRepublic)

The DIMC engines and chiplet solutions d-Matrix makes are alternatives to GPU-based solutions, so d-Matrix could be poised to provide a solution to a major problem.

Subscribe to the Daily Tech Insider Newsletter

Stay up to date on the latest in technology with Daily Tech Insider. We bring you news on industry-leading companies, products, and people, as well as highlighted articles, downloads, and top resources. You’ll receive primers on hot tech topics that will help you stay ahead of the game.

Delivered Weekdays Sign up today