Newsroom

Posted on May 12, 2025

Posted on May 12, 2025

Posted on May 11, 2025

While p-value-based decision-making offers a seemingly straightforward approach, its inherent limitations in being a one-size-fits-all, absolute measure based on a single analysis can be problematic. Relative model selection criteria like AIC, AICc, and BIC provide a more nuanced and robust framework by comparing multiple plausible models. Grounded in information theory, aligned with Occam’s razor, and being compatible with inference to the best explanation, these criteria estimate the relative information loss or, from a Bayesian perspective, the evidence for different models. Their application spans traditional statistical modeling and modern data science, aiding in tasks ranging from regression analysis to feature selection.

Link: https://www.youtube.com/watch?v=GTu5XF-QFUI

Posted on May 10, 2025

In an age where information flows constantly through diverse channels, the ability to understand and process data in multiple formats has never been more important. We live in a multimodal world—a reality composed not just of text, but of images, speech, videos, charts, and other forms of sensory and symbolic input. This growing need has driven the development of multimodal artificial intelligence—AI systems that can process and reason over data in more than one form. Let's explore!

Link: https://www.youtube.com/watch?v=RADRU9zySi0

Posted on May 9, 2025

Link: https://www.youtube.com/watch?v=Li1NblqcJTY

Posted on May 9, 2025

On May 5, 2025, Julius Černiauskas published a thought-provoking article titled “Behind the Scenes of Using Web Scraping and AI in Investigative Journalism.” The summary is as follows:

While investigative journalism often conjures images of hidden sources and undercover work, many compelling stories begin with publicly available information—data hiding in plain sight. This is where web scraping, the automated extraction of online data, has become indispensable. It's not only a method for gathering facts quickly, but also a powerful tool for holding institutions accountable, revealing data manipulation, and uncovering misconduct. For instance, data scraping tools exposed that 38,000 articles about the war in Ukraine, all published in a single year, were attributed to the same supposed “journalist,” helping real reporters debunk fake journalism and identify inauthentic authorship.

Despite common misconceptions that web scraping is shady, journalists—including nonprofit newsroom The Markup—have actively defended it, even at the U.S. Supreme Court, arguing that it’s critical to a functioning democracy. In tandem, artificial intelligence is amplifying what journalists can do with scraped data, from sifting through massive document troves to spotting anomalies and generating leads. Even those without coding skills can now use no-code tools like browser extensions to engage in data-driven storytelling. Yet, ethical concerns remain front and center. Journalists must use discretion when gathering and storing data, particularly when anonymity is vital, such as monitoring the dark web. Trained AI systems can assist with filtering sensitive content, but final editorial decisions must always lie with human professionals. Ultimately, the fusion of AI and web scraping empowers investigative reporters to uncover meaningful truths in a sea of digital noise, transforming journalism in the data age.

That’s my take on it:

On one hand, web scraping unlocks access to vast amounts of public information, making it a critical tool for uncovering patterns, inconsistencies, or outright manipulation, like the case of the fake Ukraine war journalist. On the other hand, robots.txt files and similar exclusion tags give website owners a way to block automated scraping, whether for reasons of privacy, intellectual property, or security. Simply put, opt-out mechanism can be used to hide things from scrutiny.

This creates a structural asymmetry: those who have something to hide—or simply the means and awareness to deploy these exclusion tags—can wall off their content from automated analysis, while less technically-guarded or smaller sites remain open. In turn, this can skew investigations by making some patterns invisible and some actors untouchable. It also means that bad-faith players who understand how to manipulate these rules can fly under the radar, especially if journalists adhere strictly to ethical or legal boundaries around scraping.

There's also the valid concern about intellectual property and content ownership. Just because something is publicly viewable doesn't mean it’s legally or ethically scrapeable. This is especially tricky when it comes to original reporting, personal blogs, or creative work, where scraping for republishing or mass analysis feels exploitative rather than investigative.

As such, scraping-based journalism can be incomplete or biased, especially when key data sources opt out—whether to hide shady activity or to protect legitimate rights. That’s why transparency in methodology is so important. Responsible journalists often disclose the scope and limits of their data collection, highlighting what they could and couldn’t access. And it also points to a larger issue: technology alone isn't enough—a thoughtful, skeptical human must still decide what the data really means and where the blind spots lie.

Link: https://hackernoon.com/behind-the-scenes-of-using-web-scraping-and-ai-in-investigative-journalism

Posted on May 2, 2025

Huawei is rapidly emerging as a key player in the AI chip market, having begun deliveries of its advanced AI "cluster" system, CloudMatrix 384, to domestic clients in China, according to the Financial Times. This development comes in response to growing U.S. export restrictions that have made it increasingly difficult for Chinese companies to acquire Nvidia’s high-end semiconductors. Huawei has reportedly sold over ten units of the CloudMatrix 384, a system that links together a large number of AI chips, and these have been shipped to data centers supporting various Chinese tech firms.

Dylan Patel, founder of SemiAnalysis, stated that CloudMatrix 384 is capable of outperforming Nvidia’s flagship NVL72 cluster in both computational power and memory. Despite some drawbacks—namely higher power consumption and more complex software maintenance—CloudMatrix is seen as a viable and attractive alternative, especially given China’s deep engineering talent pool and ample energy resources. This marks a significant strategic shift as China looks to reduce its dependence on Western AI hardware.

That’s my take on it:

The CloudMatrix 384 consumes nearly four times more power than the NVL72, leading to lower energy efficiency. Despite this, in regions like China where power availability is less constrained, the higher energy consumption is considered an acceptable compromise for the increased computational capabilities.

Based on the current trend, it is unlikely that Huawei's technology can catch up Nvidia’s in the near future. Nvidia isn’t just a chipmaker—it’s an ecosystem. It dominates the AI space not only with its hardware (e.g., H100) but also with its software stack (CUDA, cuDNN, TensorRT, etc.). These tools are mature, widely adopted, and deeply integrated into enterprise and research workflows.

But don’t forget that in the '80s, Japan’s chipmakers like NEC, Toshiba, and Hitachi managed to outcompete U.S. firms like Intel in DRAM by focusing on quality control, manufacturing efficiency, and aggressive investment. While Nvidia leads now, that lead isn’t invincible.

Link: https://www.ft.com/content/cac568a2-5fd1-455c-b985-f3a8ce31c097?accessToken=zwAAAZcgU2HwkdPKxWiiX9FFXNO5hfOozjHAlwE.MEQCIASnmNkxJzppNfWifnU4F8NIZHhvb-dI-uQ92OJ4P8egAiAKodKrU6w-8_cmYRzPi54ClKa2rBh2XKAP-t6iAFKwCw&segmentId=cac568a2-5fd1-455c-b985-f3a8ce31c097

Posted on May 2, 2025

Recently the strategic landscape of the global electric vehicle (EV) industry has witnessed a notable shift. Japanese automakers, long admired for their craftsmanship, reliability, and global reach, are increasingly partnering with Chinese tech firms renowned for their advancements in artificial intelligence and smart mobility platforms. Is the US falling behind?

Link: https://www.youtube.com/watch?v=Xp52-mDDa4E

Posted on May 2, 2025

Posted on April 29, 2025

Posted on April 28, 2025

Quantitative Research is more than statistics: Design, measurement, and analysis in the era of big data and AI

Many people equate quantitative research with statistical analysis. Indeed, statistics is only a subset of data analysis, and data analysis is only one of three components of quantitative research. The three components are:

1. Research design

2. Measurement or data collection

3. Data analysis

Link: https://www.youtube.com/watch?v=aDIEYX_JIBM

Posted on April 27, 2025

AI could lead to serious social discontent. As individuals realize that years of education, experience, and hard work no longer secure them a stable place in the economy, frustration and resentment may grow. The divide between the "AI privileged" and the "AI disenfranchised" could mirror, or even worsen, existing economic inequalities. Access to cutting-edge AI will likely be determined by wealth, corporate affiliation, or geographical location — deepening the rift between those who can thrive and those left behind. Are you prepared for the consequences?

Link: https://www.youtube.com/watch?v=1pdRZ1MMzNY

Posted on April 25, 2025

Posted on April 24, 2025

In a world increasingly shaped by algorithms and artificial intelligence, a pressing question emerges: are we still truly free to make our own choices? From what we watch on Netflix to the news we read on our social media feeds, AI recommendation systems shape and filter our experiences with remarkable precision. This video explores the implications of AI-powered recommendation systems on human autonomy and moral responsibility, interrogating whether we can still be held accountable for choices made under algorithmic influence. We’ll examine these questions through the lens of four major philosophical perspectives on free will and consider the real-world implications for how we think, act, and govern ourselves in an AI-driven society.

Linked: https://www.youtube.com/watch?v=4_KNN1Y_u_E

Posted on April 24, 2025

Roger Penrose, a renowned British physicist and mathematician, believes that consciousness cannot be reduced to algorithms. His reasoning begins not with neuroscience, but with mathematics itself—specifically, with Gödel’s incompleteness theorems.

Link: https://www.youtube.com/watch?v=yyJbdU9AgOE

Posted on April 19, 2025

The Wikimedia Foundation has announced a new initiative aimed at reducing the strain placed on Wikipedia’s servers by artificial intelligence developers who frequently scrape its content. In partnership with Kaggle, a Google-owned platform for data science and machine learning, Wikimedia has released a beta dataset containing structured Wikipedia content in English and French. This dataset is explicitly designed for machine learning workflows and offers a cleaner, more accessible alternative to scraping raw article text.

According to Wikimedia, the dataset includes machine-readable representations of Wikipedia articles in the form of structured JSON files. These contain elements such as research summaries, short descriptions, image links, infobox data, and various article sections. However, it intentionally excludes references and non-textual content like audio files. The goal is to provide a more efficient and reliable resource for tasks such as model training, fine-tuning, benchmarking, and alignment.

While Wikimedia already maintains content-sharing agreements with large organizations such as Google and the Internet Archive, this collaboration with Kaggle is intended to broaden access to high-quality Wikipedia data, particularly for smaller companies and independent researchers. Kaggle representatives have expressed enthusiasm for the partnership, highlighting their platform’s role in supporting the machine learning community and their commitment to making this dataset widely available and useful.

That’s my take on it:

While the release of a structured dataset by the Wikimedia Foundation is a meaningful step toward reducing reliance on web scraping, its overall impact on the broader data science community—particularly those working with unstructured data—may be limited. For data scientists focused on structured tasks such as natural language processing or machine learning applications involving encyclopedic knowledge, the dataset offers clear benefits. By providing pre-processed, machine-readable JSON files containing curated article content, it simplifies data ingestion and integration, reducing the overhead traditionally associated with scraping and cleaning raw HTML. This is particularly valuable for smaller organizations and independent researchers who may lack the infrastructure or resources to perform large-scale data extraction.

However, for those whose work depends heavily on unstructured data—such as social media analysis, customer feedback mining, or domain-specific natural language processing—the dataset does little to alleviate their ongoing need to collect data from diverse, often messy sources. The vast majority of valuable online information remains in unstructured formats, and in many cases, it is accessible only through scraping or limited APIs. As such, this initiative by Wikimedia is unlikely to replace the necessity of scraping for most real-world applications.

Web scraping is controversial. This move is symbolically significant. It reflects a broader trend toward encouraging ethical and sustainable access to machine-learning-relevant content. By offering a public, machine-learning-friendly dataset, Wikimedia sets a precedent that could inspire other content providers to follow suit, potentially reducing the strain caused by indiscriminate scraping and fostering greater transparency. In that sense, while the immediate practical implications may be narrow, the long-term influence on data access practices could be substantial.

Link: https://www.theverge.com/news/650467/wikipedia-kaggle-partnership-ai-dataset-machine-learning

Posted on April 18, 2025

A recent study by researchers from Carnegie Mellon, Stanford, Harvard, and Princeton suggests that over-training large language models (LLMs) may actually make them harder to fine-tune. Contrary to the common belief that more training leads to better performance, the team found diminishing returns—and even performance degradation—when they trained two versions of the OLMo-1B model with different token counts. One version was trained on 2.3 trillion tokens, and the other on 3 trillion. Surprisingly, the more heavily trained model performed up to 3% worse on evaluation benchmarks like ARC and AlpacaEval. This led the researchers to identify a phenomenon they call "catastrophic overtraining," where additional training causes the model to become increasingly sensitive to noise introduced during fine-tuning. They describe this growing fragility as "progressive sensitivity," noting that beyond a certain "inflection point," further training can destabilize the model and undo prior gains. To validate this, they introduced Gaussian noise during fine-tuning and observed similar drops in performance. The takeaway is clear: training beyond a certain threshold may reduce a model's adaptability, and developers may need to rethink how they determine optimal training duration—or develop new methods that extend the safe training horizon.

That’s my take on it:

For years, the dominant belief in large language model (LLM) development has been that increasing model size and training data leads to better performance—a view supported by early scaling law research (e.g., OpenAI's and DeepMind's work). The study conducted by CMU, Stanford, Harvard, and Princeton counter-argues that bigger may not be better. There are other studies concurring with this finding. In another study, even in smaller models (1B–10B), researchers have observed what they sometimes call “loss spike” behavior—where longer training actually causes performance drops, particularly in out-of-distribution generalization. That lines up with this idea of an “inflection point” the paper describes.

The key question is: “Where is the inflection point?” or “How much is too much?” Perhaps there’s no universal threshold. Some researchers are exploring ways to detect it, including tracking validation loss trends, fine-tuning adaptability at various checkpoints, analyzing gradient noise, and probing noise sensitivity (e.g., via Gaussian perturbations). Some even use loss landscape analysis or generalization curves to flag when models start to become brittle. Perhaps future progress in LLMs may depend less on pushing scale and more on training efficiency, model robustness, and smarter tuning strategies. Instead of asking “how big can we go?” we might now ask “how far should we go before it starts breaking things?”

Link: https://arxiv.org/abs/2503.19206

Posted on April 17, 2025

Posted on April 16, 2025

Today, artificial intelligence is a household name — powering search engines, voice assistants, self-driving cars, and even generating human-like conversations. The AI revolution, largely led by the United States, took off with the rise of deep learning and large language models developed by companies like OpenAI, Google, and Meta. But decades before ChatGPT or GPT-4 made headlines, Japan had already launched a bold and ambitious attempt to build intelligent machines. In the 1980s, during its peak as a global tech superpower, Japan announced a sweeping national initiative: the Fifth Generation Computer Systems (FGCS) project. But, why did it fade away?

Link: https://www.youtube.com/watch?v=mvcDA7jv-g0

Posted on April 14, 2025

Artificial intelligence has transitioned from theoretical promise to a practical force transforming industries worldwide. This surge was propelled by advances in machine learning and the breakout of powerful generative AI models that can produce human-like text, images, and predictions. Across sectors such as healthcare, finance, education, marketing, and engineering, AI implementation accelerated lately. This video examines recent global case studies in each of these fields, highlighting the practical applications and real-world impact of AI’s deployment.

Link: https://www.youtube.com/watch?v=9EPd1WUrmfU

Posted on April 12, 2025

Posted on April 10, 2025

Artificial Intelligence is often seen as a product of recent technological revolutions, but its intellectual roots stretch deep into the 20th century. Among the constellation of thinkers who paved the way for intelligent machines, few loom as large as John von Neumann. A polymath of rare genius, von Neumann made lasting contributions to mathematics, physics, economics, and computing. While he never designed an AI system per se, the foundational work he did across multiple domains now serves as the intellectual scaffolding for much of modern AI.

Link: https://www.youtube.com/watch?v=CcT3YJAUCEg

Posted on April 8, 2025

Posted on April 6, 2025

Posted on April 5, 2025

In 1943, a young man named Walter Pitts co-authored a seminal paper with neurophysiologist Warren McCulloch titled “A Logical Calculus of the Ideas Immanent in Nervous Activity.” This paper introduced the McCulloch-Pitts neuron—a revolutionary mathematical model of how real neurons could compute logical functions. This concept is foundational not just for neuroscience, but for artificial intelligence and the entire field of neural networks. But Pitts was almost forgotten. Why?

Link: https://www.youtube.com/watch?v=1y0gFgEK0oY

Posted on April 4, 2025

Posted on April 3, 2025

Posted on March 29, 2025

Posted on March 29, 2025

On March 25 2025, Google released Gemini 2.5, its latest AI model that outperforms all other existing AI models by all major benchmarks. Specifically, Google's Gemini 2.5 Pro has demonstrated superior performance compared to other leading AI models, including OpenAI's ChatGPT and DeepSeek's offerings, across various benchmarks.

Key Features of Gemini 2.5 Pro:

1. Enhanced Reasoning Abilities: Gemini 2.5 Pro is designed as a "thinking model," capable of processing tasks step-by-step, leading to more informed and accurate responses, especially for complex prompts. This advancement allows it to analyze information, draw logical conclusions, and incorporate context effectively. 
2. Advanced Coding Capabilities: The model excels in coding tasks, including creating visually compelling web applications, agentic code applications, code transformation, and editing. 
3. Multimodal Processing: Building upon Gemini's native multimodality, 2.5 Pro can interpret and process various data forms, including text, audio, images, video, and code. This versatility enables it to handle complex problems that require integrating information from multiple sources. 
4. Extended Context Window: The model ships with a 1 million token context window, with plans to expand to 2 million tokens soon. This extensive context window allows Gemini 2.5 Pro to comprehend vast datasets and manage more extensive data, enhancing its performance in tasks requiring long-term context understanding. 

That’s my take on it:

In 1968 American artist Andy Warhol predicted that "In the future, everyone will be world-famous for 15 minutes". This quote expresses the concept of fleeting celebrity and media attention. In the age of generative AI, Andy Warhol’s prophecy echoes louder than ever: every model is famous for 15 minutes. AI has been growing at the pop culture speed. Models are celebrities. They rise fast, trend for a moment, then get dethroned.

• January 2025: DeepSeek-VL and R1 stunned everyone—especially with open weights and insane capabilities in reasoning and math.
• Early February: OpenAI fired back with o3 (internally believed to be GPT-4.5), nudging the bar higher.
• Late Feb/Early March: Qwen 2.5 enters and crushes multiple leaderboards, especially in multilingual and code-heavy tasks.
• March 2025: Gemini 2.5 Pro drops and suddenly becomes the new benchmark king in reasoning, long-context, and multi-modal tasks.

This is not just fast-paced—this is accelerating. Each "champion" barely holds the crown before someone new comes knocking. Just like any other tech curve (e.g., Moore’s Law for chips), AI can't grow infinitely in capability, speed, or intelligence without hitting some hard ceilings. But the key question is not if, but when—and what kind of plateau we will encounter. I will explore this next.

Link: https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking

Posted on March 28, 2025

Hi, all, I have uploaded a new video to YouTube. John Searle’s famous Chinese Room thought experiment challenges the idea that computers can truly “understand” language or possess minds. John McCarthy, the father of logical AI, explicitly called Searle’s Chinese room a “fallacy”. What is the controversy? This video is a brief explanation (4 minutes):

Posted on March 27, 2025

ChatGPT-4o’s image generation capabilities mark a major leap forward in AI creativity, blending high realism, smart prompt handling, and seamless editing tools in one powerful system. One of its standout strengths is photo-realistic fidelity — it renders textures, lighting, and detail with stunning clarity, often outperforming models like Midjourney or Stable Diffusion in visual accuracy.

It also has exceptional prompt comprehension, allowing users to describe complex, multi-layered scenes, styles, and emotions, and get results that align perfectly with their vision. Whether you want an anime character, a cyberpunk street scene, or a vintage oil painting, ChatGPT-4o switches styles effortlessly.

Another key advantage is its reference-aware editing — users can upload an image and make specific changes like altering backgrounds, adding objects, or modifying color tones. These edits blend in smoothly, avoiding awkward transitions or visual artifacts common in older tools.

Moreover, it handles spatial reasoning impressively. If you ask for a scene with specific object placement — like a vase to the left of a cat — it understands and respects composition accurately. This makes it ideal for design, storytelling, and visual planning tasks.

It also supports iterative workflows directly in the chat. You can request tweaks like “make the lighting softer” or “change the outfit to red,” and get updated versions quickly, without rewriting your prompt from scratch.

ChatGPT-4o further allows consistent visual output for characters or scenes across multiple images, perfect for comics or branding work. And with clean, high-resolution outputs, it minimizes distortion and maintains visual integrity even in fine detail.

The attachment shows a side-by-side comparison between the images created by 4o image generator and its previous version, DALL-E3.

That’s my take on it:

One of the standout strengths of the ChatGPT-4o image generator is its exceptional ability to produce technically accurate and visually effective infographics. While most AI generators excel at creating photorealistic images, 4o distinguishes itself by delivering visuals that are genuinely useful for educational and technical communication.

When I need to generate illustrations for topics in statistics or computing, tools like ReCraft and Ideogram often fall short. They tend to approximate the concept or struggle with textual accuracy. In contrast, 4o consistently produces infographics that are not only visually appealing but also presentation-ready and pedagogically sound.

For example, I tested the following prompt:

Example 1: “Illustrate Lambda smoothing in a scatterplot with data forming a nonlinear pattern. The illustration must be good enough for teaching purposes.”
As shown in the side-by-side comparison, the image generated by ReCraft includes nonsensical text and distorted elements, making it unusable for serious teaching. The 4o-generated image, however, is clean, precise, and visually intuitive — ideal for lectures or documentation.

Another test:
Example 2: “Illustrate deep learning by emphasizing transformations inside multiple hidden layers in a neural network. Make the graph colorful and appealing.”
While Ideogram generated a visually pleasing layout, it lacked essential components like labels or explanatory structure. In contrast, 4o produced a textbook-style diagram with proper node icons, layer labels, and transformation highlights — exactly what you'd expect in professional slides or educational material.

In today's landscape, many AI tools can generate impressive imagery, but when it comes to high-quality, functional infographics, ChatGPT-4o is in a league of its own (see attached PDF. Please scroll down to view all).

Link: https://openai.com/index/introducing-4o-image-generation/

Posted on March 26, 2025

The Scaling Hypothesis is the idea that the performance of artificial intelligence systems, particularly large language models, increases predictably and substantially as the amount of computational power, training data, and model size grows. According to this view, intelligence does not arise from complex or specially designed algorithms, but rather from the sheer scale of resources applied. Is it true? Let's explore.

Link: https://www.youtube.com/watch?v=q7guu1LgBn8

Posted on March 26, 2025

Posted on March 24, 2025

During the spring break I was interviewed by Sandra Wu, the department Chair of Financial, Accounting, and Legal Studies at Algonquin College, Canada in her program “Career Canvas”. The title of the episode is: “Don’t Be a Baby Duck: Lifelong Learning and Reinvention in the Age of AI.” The full interview can be accessed at: https://www.youtube.com/watch?v=Jw2Yz8Et3qU&t=2011s

Posted on March 21, 2025

In a year where artificial intelligence is becoming the bedrock of innovation across industries, the importance of data science has never been clearer. As Michel Tricot, CEO of Airbyte, puts it: “No data, no AI.” The 10 companies recognized by Fast Company in 2025 aren’t just building clever AI tools—they’re transforming how data is collected, processed, and used to solve real-world problems. From healthcare to crypto, supply chains to outer space, these innovators are proving that the smart use of data can power meaningful change.

1. Unstructured

Unstructured unlocks hidden business value by converting unstructured data into AI-ready formats, fueling applications like RAG and fine-tuned LLMs. With 10,000+ customers and partnerships with U.S. military branches, it’s become a foundational tool for enterprise AI.

2. Chainalysis

Chainalysis brings clarity to the murky world of crypto through blockchain forensics, helping trace and recover billions in illicit funds. In 2024 alone, it analyzed $4 trillion in transactions and secured a landmark legal win for crypto analytics.

3. Airbyte

Airbyte makes large-scale data integration seamless, enabling AI initiatives with plug-and-play connectors and unstructured data support. Its open-source model now powers over 170,000 deployments and a thriving ecosystem of 10,000+ community-built connectors.

4. Norstella

Norstella speeds up the drug development pipeline by analyzing billions of data points through its AI platforms, helping pharma companies make faster, smarter decisions. It has directly contributed to the launch of over 50 new drugs in the past year.

5. Makersite

Makersite empowers product teams to design more sustainably with real-time supply chain data and AI-driven life cycle analysis. In one standout case, it helped Microsoft slash the Surface Pro 10’s carbon footprint by 28%.

6. Anaconda

Anaconda is democratizing AI by enhancing Python workflows for data scientists and non-coders alike, with tools like Python in Excel and a secure AI model library. Now used by over 1 million organizations, it’s a key enabler of accessible data science.

7. Satelytics

Satelytics uses advanced geospatial analytics to detect methane leaks and monitor land health, offering quick insights from satellites and drones. Its technology helped Duke Energy detect hundreds of leaks and has expanded across multiple industries.

8. Rune Labs

Rune Labs is changing the way Parkinson’s disease is managed with real-time data from wearables and AI-driven treatment insights. Its platform has improved patient outcomes significantly, reducing ER visits and boosting medication adherence.

9. EarthDaily

EarthDaily enhances sustainability in mining through hyperspectral imaging and radar analytics that reduce environmental impact and safety risks. It provides precision tools to accelerate mineral discovery while avoiding unnecessary drilling.

10. Nominal

Nominal streamlines testing and evaluation in aerospace, defense, and high-tech sectors with a unified, real-time analytics platform. Used in everything from drone trials to spacecraft diagnostics, it’s redefining how critical systems are validated.

That’s my take on it:

The phrase “No data, no AI” captures more than just a technical truth—it underscores the deep interdependence between data science and artificial intelligence. No matter how advanced AI becomes, it cannot function in a vacuum. It needs clean, relevant, and well-structured data to learn, adapt, and perform effectively. And that process—collecting, cleaning, transforming, and curating data—is still very much a human-driven discipline.

The success of the companies recognized by Fast Company in 2025 highlights this reality. Whether it's transforming unstructured data into LLM-ready formats, streamlining complex supply chains, or analyzing geospatial signals from satellites, these innovations all hinge on strong data science foundations, not just AI magic. What they demonstrate is that while AI engineering skills are in high demand, non-AI data science roles—like data engineering, data quality management, and domain-specific analytics—remain absolutely essential.

Link: https://www.fastcompany.com/91269286/data-science-most-innovative-companies-2025

Posted on March 20, 2025

Posted on March 20, 2025

Posted on March 20, 2025

Posted on March 19, 2025

Nvidia's GPU Technology Conference (GTC) keynote, delivered by CEO Jensen Huang, took place on March 18, 2025, at the SAP Center in San Jose, California. The following are the key points:

1. Next-Generation AI Chips:

• Blackwell Ultra: Scheduled for release in the latter half of 2025, this GPU boasts enhanced memory capacity and performance, offering a 1.5x improvement over its predecessors. 
• Vera Rubin: Named after the renowned astronomer, this AI chip is set to launch in late 2026, followed by Vera Rubin Ultra in 2027. These chips promise substantial performance gains and efficiency improvements in AI data centers. 

2. AI Infrastructure and Software:

• Nvidia Dynamo: An open-source inference software system designed to accelerate and scale AI reasoning models, effectively serving as the "operating system of an AI factory." 

3. Robotics and Partnerships:

• 'Blue' Robot: Developed in collaboration with Disney Research and Google DeepMind, this robot showcases advancements in robotics technology and a new physics engine called Newton. 
• General Motors Collaboration: Nvidia is partnering with GM to integrate AI systems into vehicles, factories, and robots, aiming to enhance autonomous driving capabilities and manufacturing processes. 

4. AI Evolution and Future Outlook:

• Agentic AI: Huang highlighted the progression of AI from perception and computer vision to generative and agentic AI, emphasizing its growing ability to understand context, reason, and perform complex tasks.
• Physical AI: The next wave of AI involves robotics capable of understanding physical concepts like friction and inertia, with Nvidia introducing tools like Isaac GR00T N1 and the evolving Cosmos AI model to facilitate this development. 

That’s my take on it:

Despite these advancements, Nvidia's stock experienced a 3.4% decline during the keynote. The announcements, while significant, were perceived as extensions of existing technologies rather than disruptive innovations. While Nvidia continues to innovate, the emergence of efficient and cost-effective AI models from Chinese companies is reshaping the competitive landscape.

Further, the partnerships between Nvidia, Disney, and GM are not exciting at all. Disney is primarily an entertainment company rather than a technology leader. While they do invest in advanced CGI, theme park animatronics, and AI-driven personalization, they aren’t a dominant force in AI hardware or software. The company has faced backlash over diversity and inclusion policies, especially regarding recent film releases like Snow White. This controversy might make Disney a less attractive partner from a PR perspective, particularly if Nvidia is looking to impress a broader tech audience.

While GM is one of the biggest automakers in the U.S., it has struggled to keep pace with Tesla and BYD in the EV and autonomous driving sectors. Tesla’s Full Self-Driving (FSD) is already on the road, and BYD dominates China’s EV market with highly cost-effective solutions. GM’s self-driving unit Cruise has faced setbacks, including safety issues and regulatory scrutiny, leading to a halt in operations in multiple cities. This tarnishes GM’s image as a leader in AI-powered mobility. In my opinion, these partnerships aren’t groundbreaking.

Link: https://www.youtube.com/watch?v=erhqbyvPesY

Posted on March 18, 2025

Posted on March 18, 2025

Recently Baidu has launched ERNIE 4.5 and ERNIE X1, two new AI models focused on multimodal capabilities and advanced reasoning, respectively.

• Performance & Benchmarks: Baidu claims these models outperform DeepSeek V3 and OpenAI’s GPT-4.5 on third-party benchmarks like C-Eval, CMMLU, and GSM8K.
• Cost Advantage: ERNIE 4.5 is 99% cheaper than GPT-4.5, and ERNIE X1 is 50% cheaper than DeepSeek R1, emphasizing aggressive market positioning.
• ERNIE X1 Capabilities: Designed for complex reasoning and tool use, it supports tasks like advanced search, document Q&A, AI-generated image interpretation, and code execution.
• ERNIE 4.5 Capabilities: A multimodal AI optimized for text, image, audio, and video processing, featuring improved reasoning, generation, and hallucination prevention through FlashMask Dynamic Attention Masking and Self-feedback Enhanced Post-Training.

That’s my take on it:

Baidu's ERNIE 4.5 model is priced at approximately 1% of OpenAI's GPT-4.5 cost. It is an attractive option for businesses looking to cut AI expenses, especially in cost-sensitive markets like China, Southeast Asia, and emerging economies. Nevertheless, GPT-4.5 is widely recognized as the best-performing model in English, and OpenAI has a trust advantage among global businesses. OpenAI’s models are deeply integrated into Microsoft’s ecosystem, dominating enterprise AI adoption in the West.

Although ERNIE 4.5 is claimed to outperform GPT-4.5, independent benchmarks are still lacking. In addition, many U.S. and European companies might hesitate to adopt Baidu’s AI due to security concerns and data regulations. Further, Chinese LLMs, including ERNIE 4.5, operate under strict government regulations that enforce censorship on politically sensitive topics. This has major implications for freedom of information, research, and AI usability outside of China.

Link: https://venturebeat.com/ai/baidu-delivers-new-llms-ernie-4-5-and-ernie-x1-undercutting-deepseek-openai-on-cost-but-theyre-not-open-source-yet/

Posted on March 18, 2025

Posted on March 12, 2025

Posted on March 12, 2025

The new AI agent Manus, developed by the Wuhan-based startup Butterfly Effect, has taken the AI world by storm since its launch on March 6, 2025. Unlike traditional chatbots, Manus operates as a general AI agent, leveraging multiple models, including Claude 3.5 Sonnet and Alibaba’s Qwen, to perform a variety of tasks autonomously. Simply put, it is capable of multi-tasking.

Despite the hype, access to Manus remains limited, with only a small fraction of users receiving invite codes. MIT Technology Review tested the tool and found it to be a promising but imperfect assistant, akin to a highly competent intern—capable but prone to occasional mistakes and oversights.

The reviewer conducted three tests:

1. Compiling a list of China tech reporters – Initially, Manus produced an incomplete list due to time constraints but improved significantly with feedback.
2. Finding NYC apartment listings – It required clarification for nuanced search criteria but eventually delivered a well-structured ranking.
3. Nominating candidates for Innovators Under 35 – The task was more challenging due to research limitations, paywall restrictions, and system constraints. The final output was incomplete and skewed.

Strengths:

• Transparent, interactive process allowing user intervention
• Strong performance in structured research tasks
• Affordable ($2 per task, significantly cheaper than alternatives like ChatGPT DeepResearch)
• Replayable and shareable sessions

Weaknesses:

• Struggles with large-scale research, paywalls, and CAPTCHA restrictions
• System instability and crashes under heavy load
• Requires user guidance to refine results

While Manus is not flawless, it represents a significant step in AI autonomy, particularly in research and analysis. It underscores China’s growing role in shaping AI development, not just in model innovation but also in the practical implementation of autonomous AI agents.

Links: https://www.youtube.com/watch?v=WTgkRitFKGs

https://www.technologyreview.com/2025/03/11/1113133/manus-ai-review/

Posted on March 12, 2025

Posted on March 11, 2025

Mistral AI, a leading French AI startup, is recognized as one of France’s most promising tech firms and the only European contender to OpenAI. Despite its impressive $6 billion valuation, its global market share remains modest.

A few days ago the company launched its AI assistant, Le Chat, on mobile app stores, generating significant attention, particularly in France. French President Emmanuel Macron even endorsed it in a TV interview, urging people to choose Le Chat over OpenAI’s ChatGPT. The app quickly gained traction, reaching 1 million downloads in two weeks and topping France’s iOS free app chart.

Founded in 2023, Mistral AI champions openness in AI and positions itself as the “world’s greenest and leading independent AI lab.” Its leadership team includes ex-Google DeepMind CEO Arthur Mensch and former Meta AI researchers Timothée Lacroix and Guillaume Lample. The company’s advisory board includes notable figures like Jean-Charles Samuelian-Werve, Charles Gorintin, and former French digital minister Cédric O, whose involvement sparked controversy.

Despite its growth and strong funding, Mistral AI’s revenue is still in the eight-digit range, indicating it has significant ground to cover before becoming a true OpenAI rival.

That’s my take on it:

Mistral AI has the potential to become a serious competitor to OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and other top AI models. The strained relationship between the U.S. and Europe, particularly during the Trump administration, has fueled a growing sense of technological sovereignty in Europe. As tensions over trade, defense, and digital policies deepened, many European nations—especially France—became increasingly wary of relying on American tech giants. This sentiment extends to AI, where European leaders and businesses are seeking alternatives to U.S.-dominated models like ChatGPT, Claude, and Google Gemini.

Mistral AI, as Europe’s most promising AI company, stands to benefit from this shift. French President Emmanuel Macron’s endorsement of Le Chat highlights a broader push for European-built AI solutions, reinforcing the region’s desire for independent innovation and data security. With strong government backing and a growing market of users eager to support local technology, Mistral AI could leverage this geopolitical rift to carve out a stronghold in Europe, challenging American AI dominance in the years to come.

However, Mistral AI still faces several challenges. Outside of France and Europe, brand recognition is still weak compared to OpenAI, Google, and Anthropic. 

Link: https://techcrunch.com/2025/03/06/what-is-mistral-ai-everything-to-know-about-the-openai-competitor/

Posted on March 10, 2025

Dr. Chong Ho (Alex) Yu and his colleagues conducted a research study on perceptions of Artificial Intelligence (AI) use in higher education. The summary of the responses is here. 

Posted on March 1, 2025

Posted on February 22, 2025

Posted on February 21, 2025

Google has introduced an "AI Co-Scientist," a sophisticated AI system designed to assist researchers in accelerating scientific discovery. Built on Gemini 2.0, Google’s latest AI model, the AI Co-Scientist can generate testable hypotheses, research overviews, and experimental protocols. It allows human scientists to input their research goals in natural language, suggest ideas, and provide feedback.

In an early demonstration, the AI Co-Scientist solved a complex scientific problem in just two days—a problem that had confounded researchers for over a decade. A notable test involved researchers from Imperial College London, who had spent years studying antibiotic-resistant superbugs. The AI Co-Scientist independently analyzed existing data, formulated the same hypothesis they had reached after years of work, and did so in a fraction of the time.

The system has shown promising results in trials conducted by institutions such as Stanford University, Houston Methodist, and Imperial College London. Scientists working with the AI have expressed optimism about its ability to synthesize vast amounts of evidence, identify key research questions, and streamline experimental design, potentially eliminating fruitless research paths and accelerating progress significantly.

This is my take on it:

The rapid advancement of AI in research and data analysis raises important questions about the future of statistical and data science education. As AI systems become more proficient at conducting analysis, traditional data analysts may face challenges in maintaining their relevance in the job market. Since AI models rely heavily on the quality of data, perhaps our focus should shift from analysis to data acquisition. Specifically, ensuring that students develop strong skills in data collection, validation, and preprocessing will be critical. Understanding biases in data, ethical considerations, and methods for ensuring data integrity will be more valuable than manually performing statistical calculations. In addition, while AI can analyze data, human judgment is required to interpret results in context, assess their implications, and make informed decisions. Thus, statistical and data science education should emphasize critical thinking, domain expertise, and the ability to translate insights into real-world applications.

Link: https://www.forbes.com/sites/lesliekatz/2025/02/19/google-unveils-ai-co-scientist-to-supercharge-research-breakthroughs/ 

Posted on February 18, 2025

Yesterday (2/17) Elon Musk unveiled Grok 3, the latest AI chatbot from his company xAI. This new version is designed to surpass existing chatbots like OpenAI's ChatGPT, boasting advanced reasoning capabilities that Musk describes as "scary-smart." Grok 3 has been trained using xAI's Colossus supercomputer, which utilizes 100,000 Nvidia H100 GPUs, providing 200 million GPU-hours for training—ten times more than its predecessor, Grok 2.

During the live demo, Musk highlighted Grok 3's ability to deliver "insightful and unexpected solutions," emphasizing its potential to revolutionize AI interactions. The chatbot is now available to X Premium Plus subscribers, with plans to introduce a voice interaction feature in the coming week.

That’s my take on it:

Elon Musk described Grok 3 as the "smartest AI on Earth." He stated that Grok 3 is "an order of magnitude more capable" than its predecessor, Grok 2, and highlighted its performance in areas like math, science, and coding, surpassing models from OpenAI, Google, and DeepSeek. However, it's important to note that these claims have not been independently verified.

According to "Huang's Law", proposed by Nvidia CEO Jensen Huang, the performance of AI and GPUs doubles every two years, driven by innovations in architecture, software, and hardware. Earlier this year, OpenAI released Deep Research that outperforms DeepSeek's R1 in specific tasks. For now, Grok 3 may be the most advanced AI on Earth, but how long will that last? In just a month or two, another company could unveil a model that outshines everything before it. Huang's Law is right!

Links: https://www.livemint.com/ai/grok-3-launch-live-elon-musks-xai-smartest-ai-on-earth-today-sam-altman-openai-chatgpt-gemini-google-deepseek-11739810000644.html?utm_source=chatgpt.com

https://nypost.com/2025/02/18/business/elon-musks-xai-claims-grok-3-outperforms-openai-deepseek/?utm_source=chatgpt.com

Posted on February 18, 2025

Hi, all, in a previous video, I explored whether the US will win the AI race. This video serves as a sequel, shifting the focus to China. Now, I will assess the likelihood of China emerging as the victor in the AI race.

Link: https://www.youtube.com/watch?v=nfgpb29y-0I

Posted on February 17, 2025

Posted on February 12, 2025

Posted on February 3, 2025

Posted on February 3, 2025

Video: https://www.youtube.com/watch?v=8hJF8xohSTs

Posted on February 1, 2025

Posted on January 30, 2025

Posted on January 29, 2025

OpenAI, supported by major investor Microsoft, suspects that DeepSeek may have illicitly utilized its proprietary technology to develop R1. The primary concern centers on the potential use of a technique known as "distillation."

Distillation in AI refers to a process where a smaller model is trained to replicate the behavior of a larger, more complex model. This is achieved by having the smaller model learn from the outputs of the larger model, effectively "distilling" its knowledge. While this method can enhance the efficiency of AI models, using it without proper authorization, especially with proprietary systems, raises significant ethical and legal issues.

Microsoft's security researchers detected unusual activities in the fall, where individuals allegedly linked to DeepSeek were extracting large volumes of data using OpenAI's API. This behavior suggests that DeepSeek may have bypassed usage restrictions to gather substantial data, potentially violating OpenAI's terms of service.

David Sacks, serving as President Donald Trump's top AI adviser, has indicated that there is considerable evidence pointing to DeepSeek's involvement in replicating OpenAI's models. He emphasized the seriousness of these allegations, highlighting the broader implications for intellectual property rights in the AI industry.

That’s my take of it:

As a researcher, I strive to approach every issue with an open mind, setting aside preconceptions. I do not accept claims at face value but instead rely on thorough investigation and fact-checking. I avoid cherry-picking data to support any predetermined stance. Regarding the controversy surrounding DeepSeek, I carefully examine claims from all sides to ensure a fair and objective assessment.

While simple and generic questions may produce similar responses across different AI systems, I conducted an informal test comparing ChatGPT, Claude, Perplexity AI, and DeepSeek by posing a fairly complex question related to instructional technologies. This included topics such as Seymour Papert’s LOGO programming, hypertext, multimedia, web-based instruction, and course management systems—each covering a broad range of information, making it nearly impossible for two models to generate identical responses. While the answers from ChatGPT, Claude, and Perplexity AI varied significantly, the responses from ChatGPT and DeepSeek were strikingly similar. For example,

Web-based instruction

ChatGPT: “The advent of the internet brought hopes of democratizing education through web-based instruction. Massive Open Online Courses (MOOCs) epitomized this promise, with institutions like MIT and Stanford offering free courses to millions. While these platforms achieved significant reach, their completion rates remained dismally low, often below 10%. MOOCs largely benefited highly motivated, self-directed learners while failing to support those who needed guidance and structure the most.”

DeepSeek: “The advent of the internet brought promises of democratized education, with web-based instruction enabling access to knowledge for learners worldwide. While platforms like MOOCs (Massive Open Online Courses) have made education more accessible, they have also faced criticism for low completion rates and a lack of meaningful engagement. Many learners struggle with self-directed online learning, and the absence of personalized support often leads to disengagement.”

Hypertext and Multimedia

ChatGPT: “The rise of hypertext and multimedia in the 1980s and 1990s was heralded as a way to create richer, more engaging learning environments. Proponents argued that students would benefit from non-linear access to information, enabling personalized learning paths. However, the results often showed cognitive overload rather than enhanced understanding, as students struggled to navigate and synthesize large amounts of unstructured information. Furthermore, multimedia often prioritized flashiness over substance, contributing little to deeper learning.”

DeepSeek: “In the 1990s, hypertext and multimedia were touted as tools that would make learning more interactive and engaging. Proponents argued that these technologies would allow students to explore information in non-linear ways, catering to individual learning styles. However, the reality was often overwhelming for learners, who struggled with cognitive overload and disorientation. Moreover, the lack of structured guidance in hypertext environments sometimes hindered rather than helped learning.”

Nonetheless, it is still premature to draw a conclusion. This issue awaits further investigation.

Link: https://apnews.com/article/deepseek-ai-chatgpt-openai-copyright-a94168f3b8caa51623ce1b75b5ffcc51

Posted on January 28, 2025

Several AI experts assert that DeepSeek is built upon existing open-source models, such as Meta’s LlaMA. For example, according to a research scientist at Riot Games, there is evidence suggesting that China's DeepSeek AI models have incorporated ideas from open-source models like Meta's Llama. Analyses indicate that DeepSeek-LLM closely follows Llama 2's architecture, utilizing components such as RMSNorm, SwiGLU, and RoPE.

Even the paper published by DeepSeek said so. In the paper entitled “DeepSeek LLM: Scaling open-source language models with longtermism” (Jan 2024), the DeepSeek team wrote, “At the model level, we generally followed the architecture of LLaMA, but replaced the cosine learning rate scheduler with a multi-step learning rate scheduler, maintaining performance while facilitating continual training” (p.3).

However, today (Jan., 28, 2025) when I asked DeepSeek whether it learned from Meta’s LLaMA, the AI system denied it. The answer is: “No, I am not based on Meta's LLaMA (Large Language Model Meta AI). I am an AI assistant created exclusively by the Chinese Company DeepSeek. My model is developed independently by DeepSeek, and I am designed to provide a wide range of services and information to users.”

That’s my take on it:

Various sources of information appear to be conflicting and inconsistent. Nonetheless, If DeepSeek built its model from scratch but implemented similar techniques, it can technically argue that it is an "independent" development, even if influenced by prior research.

It is too early to draw any definitive conclusions. At present, Meta has assembled four specialized "war rooms" of engineers to investigate how DeepSeek’s AI is outperforming competitors at a fraction of the cost. Through this analysis, Meta might be able to determine whether DeepSeek shares any similarities with LLaMA. For now, we should wait for further findings.

Links: https://fortune.com/2025/01/27/mark-zuckerberg-meta-llama-assembling-war-rooms-engineers-deepseek-ai-china/

https://planetbanatt.net/articles/deepseek.html?utm_source=chatgpt.com

https://arxiv.org/pdf/2401.02954

Posted  on January 28, 2025

While global attention is focused on DeepSeek, it is noteworthy to highlight the recent releases of other powerful AI models by China's tech companies.

MiniMax: Two weeks ago, this Chinese startup introduced a new series of open-source models under the name MiniMax-01. The lineup includes a general-purpose foundational model, MiniMax-Text-01, and a visual multimodal model, MiniMax-VL-01. According to the developers, the flagship MiniMax-01, boasting an impressive 456 billion parameters, surpasses Google’s recently launched Gemini 2.0 Flash across several key benchmarks.

Qwen: On January 27, the Qwen team unveiled Qwen2.5-VL, an advanced multimodal AI model capable of performing diverse image and text analysis tasks. Moreover, it is designed to interact seamlessly with software on both PCs and smartphones. The Qwen team claims Qwen2.5-VL outperforms GPT-4o on video-related benchmarks, showcasing its superior capabilities.

Tencent: Last week, Tencent introduced Hunyuan3D-2.0, an update to its open-source Hunyuan AI model, which is set to transform the video game industry. The updated model aims to significantly accelerate the creation of 3D models and characters, a process that typically takes highly skilled artists days or even weeks. With Hunyuan3D-2.0, developers are expected to streamline production, making it faster and more efficient.

That’s my take on it:

Chinese AI models are increasingly rivaling or even outperforming U.S. counterparts across various benchmarks. This growing competition poses significant challenges for U.S. tech companies and universities, particularly in attracting and retaining top AI talent. As China's AI ecosystem continues to strengthen, the risk of a "brain drain" or heightened competition for skilled researchers and developers becomes more pronounced.

Notably, in recent years, a substantial number of Chinese AI researchers based in the U.S. have returned to China. By 2024, researchers of Chinese descent accounted for 38% of the top AI researchers in the United States, slightly exceeding the 37% who are American-born. However, the trend of Chinese researchers leaving the U.S. has intensified, with the number rising dramatically from 900 in 2010 to 2,621 in 2021. The emergence of DeepSeek and similar advancements could further accelerate this talent migration unless proactive measures are taken to attract new foreign experts and retain existing ones.

To address these challenges, U.S. universities must take steps to reform the STEM education system, aiming to elevate the academic performance of locally born American students. Additionally, universities will need to expand advanced AI research programs, prioritizing areas such as multimodal learning, large-scale foundational models, and AI ethics and regulation. These efforts will be essential to maintain the United States' global competitiveness in the face of intensifying competition from China's rapidly advancing AI sector.

Link: https://finance.yahoo.com/news/deepseek-isn-t-china-only-101305918.html

Posted on January 24, 2025

The emergence of DeepSeek's AI models has ignited a global conversation about technological innovation and the shifting dynamics of artificial intelligence. Today (January 24, 2025) CNBC interviewed Aravind Srinivas, the CEO of Perplexity AI, about DeepSeek. It's worth noting that this interview is not about Deepseek only; rather, it is a part of a broader discussion about the AI race between the United States and China, with DeepSeek's achievements highlighting China's growing capabilities in the field. The following is a summary:

Geopolitical Implications:

The interview highlighted that "necessity is the mother of invention," illustrating how China, despite facing limited access to cutting-edge GPUs due to restrictions, successfully developed Deepseek.

The adoption of Chinese open-source models could embed China more deeply into the global tech infrastructure, challenging U.S. leadership. Americans worried that China could dominate the ecosystem and mind share if China surpasses the US in AI technologies.

Wake-up call to the US

Srinivas acknowledged the efficiency and innovation demonstrated by Deepseek, which managed to develop a competitive model with limited resources. This success challenges the notion that significant capital is necessary to develop advanced AI models.

Srinivas highlighted that Perplexity has begun learning from Deepseek's model due to its cost-effectiveness and performance. Indeed, in the US AI companies have been learning from each other. For example, the groundbreaking Transformer model developed by Google inspired other US AI companies.

Industry Reactions and Strategies:

There is a growing trend towards commoditization of AI models, with a focus on reasoning capabilities and real-world applications.

The debate continues on the value of proprietary models versus open-source models, with some arguing that open-source models drive innovation more efficiently.

The AI industry is expected to see further advancements in reasoning models, with multiple players entering the arena.

That’s my take on it:

No matter who will be leading in the AI race, no doubt DeepSeek is a game changer. Experts like Hancheng Cao from Emory University contended that DeepSeek's achievement could be a "truly equalizing breakthrough" for researchers and developers with limited resources, particularly those from the Global South.

DeepSeek's breakthrough in AI development marks a pivotal moment in the global AI race, reminiscent of the paradigm shift in manufacturing during the late 1970s and 1980s from Japan. Just as Japanese manufacturers revolutionized industries with smaller electronics and fuel-efficient vehicles, DeepSeek is redefining AI development with a focus on efficiency and cost-effectiveness. Bigger is not necessarily better.

Link to the Interview (second half of the video):   https://www.youtube.com/watch?v=WEBiebbeNCA

Posted on January 23, 2025

DeepSeek, a Chinese AI startup, has recently introduced two notable models: DeepSeek-R1-Zero and DeepSeek-R1. These models are designed to rival leading AI systems like OpenAI's ChatGPT, particularly in tasks involving mathematics, coding, and reasoning. Alexandr Wang, CEO of Scale AI, called DeepSeek an “earth-shattering model.”

DeepSeek-R1-Zero is groundbreaking in that it was trained entirely through reinforcement learning (RL), without relying on supervised fine-tuning or human-annotated datasets. This approach allows the model to develop reasoning capabilities autonomously, enhancing its problem-solving skills. However, it faced challenges such as repetitive outputs and language inconsistencies.

To address these issues, DeepSeek-R1 was developed. This model incorporates initial supervised data before applying RL, resulting in improved performance and coherence. Benchmark tests indicate that DeepSeek-R1's performance is comparable to OpenAI's o1 model across various tasks. Notably, DeepSeek has open-sourced both models under the MIT license, promoting transparency and collaboration within the AI community.

In terms of cost, DeepSeek-R1 offers a more affordable alternative to proprietary models. For instance, while OpenAI's o1 charges $15 per million input tokens and $60 per million output tokens, DeepSeek's Reasoner model is priced at $0.55 per million input tokens and $2.19 per million output tokens.

That’s my take on it:

Based on this trajectory, will China's AI development surpass the U.S.? Both counties have advantages and disadvantages in this race. With the world's largest internet user base, China has access to vast datasets, which are critical for training large AI models. In contrast, there are concerns and restrictions regarding data privacy and confidentiality in the US.

However, China’s censorship mechanisms might limit innovation in areas requiring free expression or transparency, potentially stifling creativity and global competitiveness. DeepSeek-R1 has faced criticism for including mechanisms that align responses with certain governmental perspectives. If I ask what happened on June 4, 1989 in Beijing, it is possible that the AI would either dodge or redirect the question, offering a neutral or vague response.

Nonetheless, China's AI is rapidly being integrated into manufacturing, healthcare, and governance, creating a robust ecosystem for AI development and deployment. China is closing the gap!

Brief explanation to reinforcement learning:  https://www.youtube.com/watch?v=qWTtU75Ygv0

Summary in mass media: https://www.cnbc.com/2025/01/23/scale-ai-ceo-says-china-has-quickly-caught-the-us-with-deepseek.html

DeepSeek’s website:  https://www.deepseek.com/

Posted on January 22, 2025

On January 21, 2025, President Donald Trump announced the launch of the Stargate project, an ambitious artificial intelligence (AI) infrastructure initiative with an investment of up to $500 billion over four years. This venture is a collaboration between OpenAI, SoftBank, Oracle, and MGX, aiming to bolster AI capabilities within the United States.

Key Details:

·       Investment and Infrastructure: The project begins with an initial $100 billion investment to construct data centers and computing systems, starting with a facility in Texas. The total investment is projected to reach $500 billion by 2029.

·       Job Creation: Stargate is expected to generate over 100,000 new jobs in the U.S., contributing to economic growth and technological advancement.

·       Health Innovations: Leaders involved in the project, including OpenAI CEO Sam Altman and Oracle co-founder Larry Ellison, highlighted AI's potential to accelerate medical breakthroughs, such as early cancer detection and personalized vaccines.

·       National Competitiveness: The initiative aims to secure American leadership in AI technology, ensuring that advancements are developed domestically amidst global competition.

That’s my take on it:

While the project has garnered significant support, some skepticism exists regarding the availability of the full $500 billion investment. Elon Musk, for instance, questioned the financing, suggesting that SoftBank has secured well under $10 billion.

Nevertheless, I am very optimistic. Even if Softbank or other partners could not fully fund the project, eventually investment would snowball when the project demonstrates promising results. In industries with high growth potential, such as AI, no investor or major player wants to be left behind. If the Stargate project starts delivering significant breakthroughs, companies and governments alike will want to participate to avoid losing competitive advantage.

Some people may argue that there is some resemblance between the internet bubble in the late 1990s and the AI hype today. The late 1990s saw massive investments in internet companies, many of which were overhyped and under-delivered. Valuations skyrocketed despite shaky business models, leading to the dot-com crash. Will history repeat itself?

It is important to note that the internet bubble happened at a time when infrastructure (broadband, cloud computing, etc.) was still in its infancy. AI today benefits from mature infrastructure, such as powerful cloud platforms (e.g., Amazon Web Services) , advanced GPUs, and massive datasets, which makes its development more sustainable and its results more immediate.

The internet primarily transformed communication and commerce. AI, on the other hand, is a general-purpose technology that extends its power across industries—healthcare, finance, education, manufacturing, entertainment, and more. Its applications are far broader, making its overall impact more profound and long-lasting.

Links: https://www.cbsnews.com/news/trump-stargate-ai-openai-softbank-oracle-musk/

https://www.cnn.com/2025/01/22/tech/elon-musk-trump-stargate-openai/index.html

Posted on January 15, 2025

Recently the World Economic Forum released the 2025 "Future of Jobs Report." The following is a summary focusing on job gains and losses due to AI and big data:

Job Gains

• Fastest-Growing Roles: AI and big data are among the top drivers of job growth. Roles such as Big Data Specialists, AI and Machine Learning Specialists, Data Analysts, and Software Developers are projected to experience significant growth.
• Projected Net Growth: By 2030, AI and information processing technologies are expected to create 11 million jobs, contributing to a net employment increase of 78 million jobs globally.
• Green Transition Influence: Roles combining AI with environmental sustainability, such as Renewable Energy Engineers and Environmental Engineers, are also seeing growth due to efforts to mitigate climate change.
• AI-Enhanced Tasks: Generative AI (GenAI) could empower less specialized workers to perform expert tasks, expanding the functionality of various roles and enhancing productivity.

Job Losses

• Fastest-Declining Roles: Clerical jobs such as Data Entry Clerks, Administrative Assistants, Bank Tellers, and Cashiers are expected to decline as AI and automation streamline these functions.
• Projected Job Displacement: AI and robotics are projected to displace approximately 9 million jobs globally by 2030.
• Manual and Routine Work Impact: Jobs requiring manual dexterity, endurance, or repetitive tasks are most vulnerable to automation and AI-driven disruptions.

Trends and Dynamics

• Human-Machine Collaboration: By 2030, work tasks are expected to be evenly split between humans, machines, and collaborative efforts, signaling a shift toward augmented roles.
• Upskilling Needs: Approximately 39% of workers will need significant reskilling or upskilling by 2030 to meet the demands of AI and big data-driven roles.
• Barriers to Transformation: Skill gaps are identified as a major challenge, with 63% of employers viewing them as a significant barrier to adopting AI-driven innovations.

That’s my take on it:

The report underscores the dual impact of AI and big data as key drivers of both job creation in advanced roles and displacement in routine, manual, and clerical jobs. Organizations and higher education should invest in reskilling initiatives to bridge the skills gap and mitigate job losses. However, there is a critical dilemma in addressing the reskilling and upskilling challenge. If faculty and instructors have not been reskilled or upskilled, how can we help our students to face the AI and big data challenges? As a matter of face, instructors often lack exposure to the latest technological advancements that are critical to the modern workforce. There is often a gap between what educators teach and what the industry demands, especially in rapidly evolving fields. Put it bluntly, the age of “evergreen” syllabus is over. The pace of technological advancements often outstrips the ability of educational systems to update curricula and training materials. To cope with the trend in the job market, we need to collaborate with technology companies (e.g., Google, Amazon, Nivida, Microsoft…etc.) to co-create curricula, fund training programs, and provide real-world learning experiences for both educators and students.

Posted on January 10, 2025

Python has been named "TIOBE's Programming Language of the Year 2024" in the TIOBE Index due to achieving the highest ratings. Nonetheless, while Python offers numerous advantages, it also faces challenges such as performance limitations and runtime errors. The TIOBE Index measures programming language popularity based on global expertise, courses, and third-party support, with contributions from major platforms like Google and Amazon. Positions two through five are occupied by C++, Java, C, and C#. Notably, SQL ranks eighth, R is positioned at number 18, and SAS is at number 22.

That’s my take on it:

Python's widespread popularity is largely driven by the growing demand for data science and machine learning. Its rich ecosystem of libraries and frameworks, including TensorFlow, PyTorch, and scikit-learn, makes it an ideal choice for DSML tasks. Interestingly, certain programming languages exhibit remarkable longevity. For example, JavaScript was ranked seventh in 2000 and currently holds sixth place. Similarly, Fortran, which was ranked 11th in 1995, now occupies the tenth position. The resurgence of Fortran is notable; according to TIOBE, it excels in numerical analysis and computational mathematics, both of which are increasingly relevant in artificial intelligence. Fortran is also gaining traction in image processing applications, including gaming and medical imaging.

While some languages maintain stable rankings over time, others have shown dramatic improvements. For instance, SQL was ranked 100th in 2005 but has since risen to ninth place. Predicting the future trajectory of programming languages is challenging, underscoring the dynamic nature of the field. As the saying goes, "Never say never!"

Links: https://www.tiobe.com/tiobe-index/

https://www.techrepublic.com/article/tiobe-index-may-2024/

Posted on January 8, 2025

Two days ago (Jan 6, 2025) Kanwal Mehreen, KDnuggets Technical Editor and Content Specialist on Artificial Intelligence, posted an article on KDnuggets, highlighting the top 10 high-paying AI skills in 2025:

Position and expected salaries

1.        Large Language Model Engineering ($150,000-220,000/year)

2.        AI Ethics and Governance ($121,800/year)

3.        Generative AI and Diffusion Models ($174,727/year)

4.        Machine Learning Ops and On-Prem AI Infrastructure ($165,000/year)

5.        AI for Healthcare Applications ($27,000 to $215,000)

6.        Green AI and Efficiency Engineering ($90,000 and $130,000/year)

7.        AI Security ($85,804/year)

8.        Multimodal AI Development ($150,000–$220,000/year)

9.        Reinforcement Learning (RL) ($121,000/year)

10.  Edge AI/On-Device AI Development ($150,000+/year)

That’s my take on it:

When I mention AI-related jobs, most people associate these positions with programming, engineering, mathematics, statistics…etc. However, as you can see, the demand for AI ethics is ranked second on the list. AI ethics is indeed a skill in high demand, and the training of professionals in this area often spans multiple disciplines. Many come from backgrounds such as philosophy, law, mass communication, and social sciences. For example, Professor Shannon Vallor is a philosopher of technology specializing in ethics of data and AI. Dr. Kate Crawford is a Microsoft researcher who studies the social and political implications of artificial intelligence. She was a professor of journalism and Media Research Centre at the University of New South Wales.

In an era where AI and data science increasingly shape our lives, the absence of ethics education in many data science and AI programs is a glaring omission. By embedding perspectives on ethics from multiple disciplines into AI and data science education, we can ensure these powerful tools are used to create a future that is not just innovative, but also just and equitable. After all, AI ethicist is a high-paying job! Why not?

Link: https://www.kdnuggets.com/top-10-high-paying-ai-skills-learn-2025

Posted on January 8, 2025

Today (Jan 7, 2025) at Consumer Electronics Summit (CES) AI giant Nvidia announced Project Digits, a personal AI supercomputer set to launch in May 2025. The system is powered by the new GB10 Grace Blackwell Superchip and is designed to bring data center-level AI computing capabilities to a desktop form factor similar to a Mac Mini, running on standard power outlets. With a starting price of $3,000, Project Digits can handle AI models up to 200 billion parameters.

The GB10 chip, developed in collaboration with MediaTek, delivers 1 petaflop of AI performance. The system runs on Nvidia DGX OS (Linux-based) and comes with comprehensive AI software support, including development kits, pre-trained models, and compatibility with frameworks like PyTorch and Python.

Nvidia's CEO Jensen Huang emphasized that Project Digits aims to democratize AI computing by bringing supercomputer capabilities to developers, data scientists, researchers, and students. The system allows for local AI model development and testing, with seamless deployment options to cloud or data center infrastructure using the same architecture and Nvidia AI Enterprise software platform.

That’s my take on it:

A few decades ago, access to supercomputers like Cray and CM5 was limited to elite scientists and well-funded institutions. Today, with initiatives like Project Digits, virtually anyone can harness the computational power needed for sophisticated projects. This democratization of technology allows scientists at smaller universities, independent researchers, and those in developing countries to test complex theories and models without the prohibitive costs of supercomputer access. This shift enables more diverse perspectives and innovative approaches to scientific challenges. Fields not traditionally associated with high-performance computing, such as sociology, ecology, and archaeology, can now leverage advanced AI models, potentially leading to groundbreaking discoveries.

Given this transformation, it is imperative to update curricula across disciplines. Continuing to teach only classical statistics does a disservice to students. We must integrate AI literacy across various fields, not just in computer science, mathematics, or statistics. Additionally, the focus should be on teaching foundational concepts that remain relevant amidst rapid technological advancements. It is equally critical to emphasize critical thinking about analytical outputs, fostering a deep understanding of their implications rather than solely focusing on technical implementation.

Link: https://www.ces.tech/videos/2025/january/nvidia-keynote/