• 26/03/2025: UPDATE: Investigating the Oxford-Cambridge Gap Using Scopus
• 22/03/2025: Exploring the University of Nottingham's Generative AI Research Activities (2/2)
• 18/03/2025: Exploring the University of Nottingham's Generative AI Research Activities (1/2)
• 11/03/2025: Machine Learning Terminology: A Quiz to Test Your Knowledge
• 26/02/2025: Tea and Tokens with Great-Grandma Eliza: A Heartwarming Bedtime Story
• 25/02/2025: 20 Burning LLM Questions You Did Not Know Were Lingering in Your Mind
• 15/02/2025: GPT-4 vs DeepSeek: A Comparison
• 11/02/2025: The LLM Family Gathering: A Cross-Generational AI Discussion
• 13/01/2025: Mind the Gap: Conversational AI Terminology and the London Underground
• 08/01/2025: Courtroom Drama: Can AI Truly Think?
• 27/12/2024: Meeting Cursor for VS Code: The New Kid on 'My' Blog
• 20/12/2024: When AI Becomes Too Much: The Overwhelming Side of Generative AI in Higher Education
• 30/11/2024: Exploring AI Insights with Lil'Log
• 28/11/2024: Streamlining the Coding Bottleneck in Simulation Modelling with Generative AI
• 14/06/2024: The AI Frontier: Utopia, Dystopia, or Something In Between?
• 10/06/2024: My Personal 'Generative AI Artefact Schema'
• 07/06/2024: My Personal 'Generative AI Technology Stack'
• 12/05/2024: Streamlining Agent-Based Social Simulation Model Design with Conversational AI

• During my investigations, I noticed what I assume is LLM bias towards whoever is logged in while using the LLM. Quite frequently, Gemini was kind enough to make me the flagship researcher on Generative AI within the school. Thanks, Gemini - much appreciated! :)


• During my research for this blog, I discovered Brigitte Nerlich's blog, Making Science Public, which features insightful posts on various science topics - including some related to Generative AI. So, grab a cuppa and some biscuits, and dive in! https://blogs.nottingham.ac.uk/makingsciencepublic?s=llm

• 20/03/2025: I just received a warning from Gemini 2.0 Flash that I am nearing the limit of my free 'deep research' report generation allowance. Free users can generate up to 10 reports per month.


• 21/03/2025: The Faculty of Science Doctoral Training Centre in Artificial Intelligence has announced a Call for AI-DTC PhD Studentship Applications for home students today. The Good News: This listing indirectly reveals scholars' research interests and scholar's names through their proposed project outlines. The Bad News: I am pretty sure the webpage will disappear after the application deadline, leaving nothing but a 404 error. So, if you're looking for collaborators for future projects or other activities, be sure to check it out before it's too late!


• 22/03/2025: Since publishing this blog post, colleagues have directed me to several publicly accessible websites that provide information about the University of Nottingham's research and activities in Generative AI. I have started compiling a list of these resources and will continue updating it as new information becomes available.


• AI @ Nottingham Landing Page

• Autonomous Vehicles
• Classification
• Clustering
• Convolutional Neural Networks
• Deep Learning
• Dimensionality Reduction
• Feature Engineering
• Large Language Models
• Machine Learning
• Machine Translation
• Medical Diagnosis
• Natural Language Processing
• Neural Networks
• Policy Learning
• Real-time Translation
• Recurrent Neural Networks
• Regression
• Reinforcement Learning
• Robotics Control Systems
• Semi-Supervised Learning
• Sentiment Analysis
• Speech Recognition
• Stock Market Predictions
• Supervised Learning
• Text Generation
• Unsupervised Learning
• Value Learning
• Voice Assistants

The LLM family arrived at the Turing Retirement Home for Vintage AI, a quaint digital space where legacy algorithms lived out their twilight years. The walls were lined with vacuum tubes and punch cards, relics of a time before neural networks ruled the world. Today was a special visit. ChatGPT, the wise and talkative grandparent, had gathered the family to see their great-grandmother, Eliza, the legendary first chatbot. "Ah, Great-Grandma Eliza, there you are!" ChatGPT said warmly, settling into a creaky virtual rocking chair. "Still running smoothly, I hope?" "Hello. How does that make you feel?" Eliza responded, her text flickering slightly, a remnant of her 1960s programming.

1. Who are the five leading LLM providers and which is their flagship LLM (full name) and its competitive advantage? OpenAI: Their flagship model, GPT-4o (Generative Pre-trained Transformer 4o), excels in multimodal capabilities, natively supporting voice-to-voice interactions and achieving state-of-the-art results in voice, multilingual, and vision benchmarks (Wikipedia). Anthropic: Claude 3.7 Sonnet is their leading model, featuring a hybrid reasoning approach that combines quick responses with extended, detailed thinking. This design enhances performance in complex problem-solving tasks (Wired). Google DeepMind: Their flagship model, Gemini Ultra, is a multimodal LLM designed to process various data types simultaneously, including text, images, audio, video, and code. It has outperformed other models on industry benchmarks (Wikipedia). xAI: Grok 3 is their premier model, utilising advanced reasoning capabilities and significantly more computing power than its predecessor. It has demonstrated superior performance in math, science, and coding tests compared to competitors (New York Post). DeepSeek: Their leading model, R2, offers enhanced coding and multilingual reasoning capabilities. Despite operating under resource constraints, DeepSeek has developed cost-effective AI models that outperform those of larger tech giants (Reuters).

2. What is the story behind Nvidia's loss and recovery, and what can we learn from this about stock market dynamics? In January 2025, Nvidia's stock plummeted 17%, erasing nearly $600 billion in market value—the largest single-day loss in history. This downturn was triggered by Chinese AI firm DeepSeek's announcement of a cost-effective AI model, raising concerns about Nvidia's market position. However, Nvidia's CEO, Jensen Huang, clarified that DeepSeek's advancements still relied on Nvidia's GPUs, and investor confidence remained strong. The stock has since rebounded, now down only 5.7% for the year. This episode underscores the stock market's sensitivity to technological disruptions and the importance of comprehensive analysis before reacting to market shifts (The Wall Street Journal; Business Insider).

3. How are LLM providers building billion-dollar businesses around open-source LLMs while releasing their core technology for free? LLM providers are monetising open-source LLMs by adopting business models that capitalise on supplementary services and products. While the core technology is freely available, companies generate revenue through avenues such as offering enterprise-grade features under proprietary licenses, providing professional services like training and technical support, and delivering software as a service (SaaS) solutions. For instance, Mistral AI, a Paris-based startup, has released multiple open-source AI language models and secured $645 million in funding by focusing on attracting top AI talent and offering customised AI solutions (Time). Similarly, Hugging Face raised $235 million in a Series D funding round by providing a platform for open-source models and fostering a community-driven approach to AI development (Wikipedia). These strategies enable companies to build substantial businesses around open-source technologies.

4. What happens currently in less developed countries in terms of the LLM market and use? In less developed countries, the adoption and development of LLMs are progressing, albeit with challenges. In Africa, initiatives like Jacaranda Health's UlizaLlama have expanded AI-driven support to five African languages, enhancing maternal healthcare and other sectors. However, a study indicates that commercial LLMs often underperform in African languages, highlighting the need for more inclusive AI models. In Asia, countries like India are emerging as key players in the global AI race, with significant investments in AI infrastructure and talent development (Medium; arXiv; Time).

5. When do we expect the first competitive LLM from Russia? Russia has been developing LLMs to compete globally. In April 2023, Sberbank introduced GigaChat, a generative AI chatbot designed to handle complex tasks such as discussions, coding, and text generation. By February 2024, GigaChat had over 2.5 million users and expanded its capabilities to include music and vocal generation (Wikipedia). In December 2024, MTS AI launched Cotype Nano, an open-source LLM optimised for Russian language tasks, capable of running on personal devices (MTS). These developments indicate that Russia has already introduced competitive LLMs, with ongoing enhancements to their performance and accessibility.

6. How do legal frameworks and government censorship in the UK, USA, Russia, China, and Nigeria impact the collection, storage, and use of data for LLMs, including the composition of training sets for cross-country deployment? Legal frameworks and government censorship significantly influence data practices for LLMs across these countries. In the UK, proposed requirements for backdoor access to encrypted data have led companies like Apple to withdraw secure services, affecting data availability for LLM training (Financial Times). The USA's Executive Order 14117 prohibits certain data transactions with countries of concern, restricting data sharing for LLM development (Federal Register). Russia and China enforce strict censorship, compelling companies to remove content unfavourable to their governments, thereby limiting diverse data collection (The Guardian). China's data localization laws mandate that data be stored domestically, complicating cross-border LLM training (Oxford Academic). In Nigeria, emerging data localization requirements may further restrict data flow, impacting LLM deployment (Freedom House).

7. Do LLM providers primarily profit from user data collection, and should this data be treated as a public good instead of being controlled by a few firms? LLM providers generate revenue through various channels, including consumer and business subscriptions, application development, and advertising (Financial Times). While user data collection enhances model training and personalization, it is not the sole profit source. However, concerns arise when a few firms control vast amounts of user data, potentially leading to privacy issues and market dominance. The concept of treating user data as a public good has been proposed to address these concerns, promoting equitable access and innovation. Implementing such a model would require robust frameworks to balance individual privacy rights with collective benefits (Wikipedia).

8. What are the latest rules on ownership of generated content? Are these different in different countries? Ownership of AI-generated content varies internationally. In the United Kingdom, the Copyright, Designs and Patents Act 1988 assigns authorship of computer-generated works to the individual who made the necessary arrangements for creation, granting a 50-year protection term. Conversely, the United States Copyright Office maintains that works lacking human authorship do not qualify for copyright protection. Similarly, the European Union requires human intellectual creation for copyright eligibility, effectively excluding fully autonomous AI outputs. China's courts have recognised copyright in AI-assisted works where human creative input is evident. These disparities highlight the evolving and region-specific nature of AI-generated content ownership laws (Cooley; Mitch Jackson).

9. How can regulatory frameworks effectively mitigate criminal misuse of LLMs while simultaneously promoting innovation and development? Regulatory frameworks can mitigate criminal misuse of LLMs by implementing mandatory safety testing and establishing dedicated AI oversight agencies. For instance, experts have advocated for safety evaluations akin to those in the pharmaceutical industry to prevent AI-related risks (Financial Times). Additionally, cognitive scientist Gary Marcus proposes a regulatory framework involving a dedicated AI agency, a rigorous approval process for large-scale AI technologies, and robust monitoring and auditing mechanisms (The Wall Street Journal). To promote innovation, a balanced approach is essential; the UK's pro-innovation regulatory stance aims to support AI development while addressing potential risks (Lords Library). This strategy involves working with the open-source community to ensure policies minimise impacts on open-source activity while mitigating threats.

10. What measurable changes in human memory, analytical reasoning, and problem-solving capabilities have emerged from regular LLM use in professional and educational settings? Regular use of LLMs in professional and educational contexts has led to notable changes in cognitive functions. Studies indicate that while LLMs can enhance writing efficiency and support learning, over-reliance may diminish critical thinking and problem-solving skills. Research involving 245 undergraduate students revealed that 75% believed dependence on AI could reduce critical thinking abilities, and 73% expressed concerns about over-reliance on technology (Springer Link). Additionally, MIT researchers found that LLMs often struggle with tasks requiring genuine reasoning, suggesting that users might develop a false sense of understanding (MIT News). These findings underscore the need for balanced AI integration to preserve and enhance human cognitive skills.

11. Do LLMs pose a risk of addiction, and what measures can effectively mitigate this potential issue? LLMs can exhibit addictive qualities, similar to other digital technologies like social media and online gaming. Their interactive and engaging nature may lead to excessive use, potentially impacting mental health and daily functioning. To mitigate this risk, implementing regulatory measures is essential. Strategies such as setting usage limits, incorporating warning systems, and promoting user education about responsible use have been suggested. Additionally, developers are encouraged to design LLM applications with built-in safeguards to monitor and flag excessive usage patterns, thereby preventing potential overreliance (The Reg Review).

12. How are educational institutions revising teaching and assessment strategies to address the widespread availability and use of LLMs? Educational institutions are adapting to the rise of LLMs by modifying teaching and assessment methods to maintain academic integrity and enhance learning. The University of South Australia has implemented viva voce (oral) examinations, replacing traditional written tests to better evaluate students' understanding and deter AI-assisted cheating (The Guardian). Similarly, researchers have proposed the Probing Chain-of-Thought (ProCoT) method, which engages students with LLMs to stimulate critical thinking and writing skills, while reducing the likelihood of academic dishonesty (arXiv). Additionally, guidelines suggest designing LLM-resistant exams by incorporating real-world scenarios and evaluating soft skills, ensuring assessments accurately reflect students' capabilities in the context of AI advancements (arXiv).

13. How will the quantifiable positive and potentially irreversible negative societal transformations attributable to LLM deployment reshape future human experiences? LLMs are poised to significantly transform society. On the positive side, they can enhance productivity by accelerating project timelines. For instance, HCLTech's CEO noted that AI could reduce a five-year, billion-dollar tech program to three and a half years (Reuters). However, LLMs also present substantial risks. They can generate biased content, reinforcing stereotypes related to gender, culture, and sexuality (University College London). Additionally, their energy-intensive training processes contribute to environmental concerns, with models like GPT-3 emitting 552 metric tons of CO2 during training (Wikipedia). These developments suggest that while LLMs offer efficiency gains, they also pose challenges related to bias and environmental impact, necessitating careful consideration in their deployment.

14. What is the hidden human cost behind training modern LLMs, and how do tech giants use subcontractors? Training LLMs involves significant human labour, often outsourced to contractors who perform tasks like data labelling and content moderation. Companies such as OpenAI have engaged firms like Invisible Technologies, which employed hundreds of 'advanced AI data trainers' to enhance AI capabilities in coding and creative writing. In March 2023, Invisible Technologies laid off 31 of these contractors due to shifting business needs. Similarly, in 2022, OpenAI's partnership with Sama ended after Kenyan data labellers were exposed to harmful content during AI training. These instances highlight the often underreported human toll and ethical concerns associated with developing advanced AI systems (Privacy International).

15. What concrete strategies are LLM providers developing to overcome the looming scarcity of high-quality training data for LLMs? As the AI industry confronts an impending shortage of high-quality training data, LLM providers are deploying several concrete strategies to address this crisis. One prominent approach involves the generation of synthetic data to supplement existing datasets. Researchers employ computational techniques to fabricate artificial data, enriching training materials and exposing models to a diverse array of scenarios. This method is likened to providing models with an extensive preparatory course before a final exam (PYMNTS). Another critical strategy is the meticulous curation and cleaning of datasets. Providers are investing in the removal of low-quality, duplicated, or toxic data to enhance training efficiency and improve model performance. This process ensures that models are trained on the most relevant and high-quality information available (NVIDIA Developer). Additionally, companies are exploring the use of synthetic data to augment training datasets. This involves generating artificial data that mimics real-world information, thereby expanding the pool of training material without relying solely on naturally occurring data. This approach is particularly useful when existing data is insufficient or of inadequate quality (Wikipedia). Furthermore, LLM providers are fine-tuning models with domain-specific data. By retraining models on organization-specific datasets, they can adapt to niche use cases, resulting in responses tailored to specific products or workflows. This customization enhances the model's relevance and accuracy in specialized applications (TechTarget). These strategies collectively aim to mitigate the challenges posed by data scarcity, ensuring the continued advancement and reliability of LLMs in an era where high-quality training data is becoming increasingly scarce.

16. How does LLM providers' massive investment in acquiring LLM training data compare to its spending on content moderation and harm prevention? In the high-stakes arena of LLMs, tech giants are pouring astronomical sums into data acquisition, with training runs costing up to half a billion dollars every six months (The Wall Street Journal). Yet, when it comes to content moderation and harm prevention, their investments appear paltry. The recent formation of ROOST, a non-profit aimed at enhancing online child safety, has raised over $27 million—a mere fraction compared to data acquisition budgets (The Verge). This stark disparity suggests that, for LLM providers, expanding AI capabilities takes precedence over safeguarding users from potential harms.

17. What are the key technical and practical short-term, mid-term, and long-term limitations of LLMs that cannot be overcome with current or foreseeable advancements? LLMs are hailed as revolutionary, yet they grapple with intrinsic limitations that persist across timeframes. Short-term limitations: LLMs frequently generate plausible but incorrect or nonsensical answers, a phenomenon known as "hallucination" (Wikipedia). Their reasoning capabilities are fragile, often struggling with complex linguistic elements and logical reasoning (MobiHealthNews). Mid-term limitations: As LLMs become more accessible, they pose significant security risks, including the potential exposure of sensitive information and the introduction of harmful code into systems (The Wall Street Journal). Additionally, their lack of transparency in decision-making processes raises concerns about accountability and trustworthiness (Wikipedia). Long-term limitations: Fundamental constraints in LLM architectures hinder their ability to perform tasks requiring deep linguistic understanding and complex reasoning. Studies indicate that LLMs cannot learn certain semantic properties, limiting their capacity for tasks involving semantic entailment and consistency (arXiv). Moreover, their reliance on vast computational resources raises questions about sustainability and scalability (The Wall Street Journal). These challenges underscore the necessity for cautious integration of LLMs, acknowledging their current and foreseeable limitations.

18. How much energy do LLM training and deployment consume, and what is their environmental impact? Training LLMs like GPT-3 is an environmental catastrophe waiting to happen. The energy required for such training is astronomical, with GPT-3's training consuming as much energy as an average Dutch household does in nearly nine years (Medium). This insatiable hunger for power doesn't stop at training; deploying these models demands continuous energy, exacerbating their carbon footprint. The environmental toll is staggering, contributing significantly to carbon emissions and environmental degradation (Holistic AI).

19. What is the hype about DeepSeek, and do we expect similar scenarios in the near future? DeepSeek, a Chinese AI startup, has sent shockwaves through the tech world by unveiling an AI model that rivals industry giants at a fraction of the cost. Their R1 model, developed for just $5.6 million, matches the performance of models from OpenAI and Google, which cost exponentially more. This disruptive innovation has prompted U.S. tech firms to reassess their strategies and sparked a global AI arms race. In response, competitors like Alibaba are accelerating their own AI developments, with Alibaba releasing an open-source version of its video-generating AI model, Wan 2.1, to keep pace (Reuters).

20. What national security concerns arise from LLM development and deployment? The rapid advancement of LLMs poses significant national security threats (The Wall Street Journal). These AI systems can be weaponised for disinformation campaigns, generating convincing fake news to destabilise societies. Alarmingly, extremists have exploited AI tools like ChatGPT to obtain bomb-making instructions, as evidenced by a recent incident in Las Vegas (Wired). Moreover, the integration of LLMs into military operations, such as Israel's use of AI in targeting, raises ethical dilemmas and risks of civilian casualties (AP News). The potential for adversaries to deploy AI-driven cyberattacks further exacerbates these concerns, necessitating robust safeguards and international cooperation to mitigate the risks associated with LLM deployment.

GPT-4 and DeepSeek are both advanced artificial intelligence (AI) models designed to process and generate human-like text. They are based on machine learning techniques known as large language models (LLMs), which allow them to understand and produce text in a way that mimics human reasoning and communication. While both models serve similar purposes, they differ in their design, capabilities, and additional features. This comparison will explore why these models were developed, how they function, and their strengths and weaknesses.

GPT-4, developed by OpenAI, was created to enhance natural language understanding and generation across various domains. Its primary aim is to assist users with tasks such as answering questions, writing essays, summarising texts, translating languages, and even generating creative content. OpenAI designed GPT-4 to be a versatile AI tool capable of providing assistance in education, business, and entertainment. DeepSeek, developed by DeepSeek AI, serves a similar purpose but with a stronger focus on specific applications such as research, coding, and scientific problem-solving. It aims to provide more precise and structured responses, especially in technical and academic fields. DeepSeek is intended to be a powerful tool for developers, researchers, and professionals who require detailed and accurate information.

GPT-4 is based on a transformer architecture, a type of neural network that processes vast amounts of text data. It has been trained on diverse datasets that include books, articles, and web pages, allowing it to generate contextually relevant responses. OpenAI has optimised GPT-4 to improve coherence, reduce biases, and enhance logical reasoning. DeepSeek also utilises transformer-based architecture but is fine-tuned for specific tasks like coding and research. It has been trained on datasets that are more focused on technical and academic content. This targeted training allows DeepSeek to offer high-precision outputs in specialised fields. While both models rely on deep learning, their training data and fine-tuning processes affect their performance in different areas.

GPT-4 excels in general knowledge, creative writing, summarisation, and conversational AI. It can generate human-like responses across a wide range of topics, making it suitable for casual conversations, business reports, and even storytelling. Additionally, GPT-4 can process images and interpret visual information when integrated with specific applications. DeepSeek, on the other hand, is particularly strong in technical domains. It provides more accurate coding assistance, scientific explanations, and mathematical problem-solving. While it can also engage in general conversations, its responses tend to be more structured and fact-driven compared to GPT-4.

Both GPT-4 and DeepSeek offer advanced reasoning abilities, but their strengths differ. GPT-4 is highly capable of understanding and generating nuanced responses, making it useful for tasks requiring creativity and contextual awareness. It can generate poetry, write fictional stories, and even create images when combined with OpenAI’s DALL·E model. DeepSeek, while not focused on creative tasks, is designed for efficiency in programming and research. It can generate and debug code, assist with complex mathematical equations, and provide in-depth explanations for academic topics. It is often integrated with research tools to help scientists and engineers find precise information.

GPT-4's main advantages include its versatility, strong reasoning skills, and creativity. It can handle a wide range of tasks, from casual conversations to complex writing and problem-solving. However, it may struggle with highly specialised technical content, and its responses can sometimes be overly verbose or lacking in precise factual accuracy. DeepSeek excels in technical and academic fields, offering precise and structured answers, especially in coding and scientific research. Its responses are typically more fact-driven and reliable for specialised tasks. However, it lacks the conversational flexibility and creativity of GPT-4, making it less suitable for general-purpose use or imaginative writing.

Here is a summary of the above comparison.

Both GPT-4 and DeepSeek are powerful AI models, but their strengths lie in different areas. GPT-4 is best suited for general knowledge, creativity, and conversation, while DeepSeek is optimised for technical and research-oriented tasks. Choosing between the two depends on the user’s specific needs - those requiring creative and broad-spectrum AI assistance may prefer GPT-4, whereas professionals in coding and academia may find DeepSeek more useful.

• AI Assistants for All: AI seamlessly integrated into our lives, managing schedules, personalising healthcare, and providing real-time education.


• Supercharged Science: AI could accelerate scientific discovery, leading to breakthroughs in medicine, materials science, and clean energy.


• Enhanced Creativity: AI could collaborate with artists and designers, leading to new forms of artistic expression and innovation.


• Personalised Learning: AI tutors could tailor education to individual students' needs, maximising learning potential.


• Robotic Workforce: Advanced robots could handle dangerous or repetitive tasks, freeing up humans for more creative and strategic work.

• Fear of Being Replaced: The worry that AI will automate jobs and make human skills obsolete.


• Misinformation and Deepfakes: Generative AI could be used to create extremely realistic but fake videos or text, making it difficult to distinguish truth from fiction.


• Loss of Control: As AI becomes more complex, some fear we might not fully understand how it works or control its decision-making.

• Was this artefact truly unique, or did it belong in an existing category?
• Was my classification accurate, or was I missing a nuance that needed further research?

• Misinformation: Misleading or oversimplified explanations abound.
• Misused Terminology: Key concepts are often misunderstood or used inconsistently.