The Personal Site of Lalo Morales

I. Executive Summary: Vision, Challenges, and the Path to Innovation

The aspiration to develop a novel, top-tier, enterprise-grade, open-source, and free AI software, built entirely from first principles without reliance on external resources, represents an undertaking of unprecedented scale and complexity within the contemporary artificial intelligence landscape. This ambitious vision, further augmented by the potential integration of quantum computing and the aim for universal adoption, demands a meticulous and multi-faceted strategic blueprint. The current state of AI development, dominated by large corporations with immense computational and data resources, underscores the formidable nature of this endeavor.

The “no outside resources” constraint is perhaps the most defining and challenging aspect, necessitating independent acquisition of vast datasets, construction of formidable computational infrastructure, and cultivation of highly specialized expertise. Achieving “top-tier” status requires not merely replicating existing capabilities but innovating beyond current state-of-the-art models. The commitment to “open-source and free” distribution introduces intricate licensing considerations, demanding a robust strategy for community engagement and sustainable development. Furthermore, the inclusion of quantum computing, while visionary, positions the project at the very frontier of technological research, operating on a significantly longer timeline for practical, widespread application in general-purpose AI. This report provides a comprehensive gameplan, outlining the foundational research, strategic resource development, iterative model building, and long-term vision required to navigate these intricate phases and realize such a groundbreaking AI system.

II. The Grand Vision: Defining a “Top-Tier, Universally Used” AI

Defining a “top-tier, universally used” AI necessitates a deep understanding of the characteristics that distinguish leading foundational models today, coupled with a forward-looking perspective on critical unmet needs that future AI systems must address.

Characteristics of Foundational AI Models

Foundational models are distinguished by their extensive pre-training on vast amounts of unlabeled data, leveraging sophisticated deep learning techniques to learn broad, generalized representations.¹ This foundational training equips them with remarkable adaptability, enabling them to perform a wide array of general tasks, from language comprehension and generation to image analysis and natural language conversation.¹ Their utility is further enhanced by the ability to be fine-tuned on smaller, task-specific datasets, allowing them to specialize in particular applications while retaining their expansive general knowledge.¹

A hallmark of contemporary top-tier models, such as OpenAI’s GPT-4o and Google’s Gemini 2.5 Pro, is their advanced multimodal capability.⁴ These models can seamlessly process and generate information across diverse modalities, including text, audio, and images, often with real-time interaction.⁴ This multimodal approach is crucial for developing intuitive human-computer interactions that more closely mimic human communication, a key direction for future AI systems.⁶ The inherent versatility of these models allows them to execute a wide range of disparate tasks with high accuracy based on user prompts, encompassing natural language processing (NLP), visual comprehension, code generation, and human-centered engagement.²

Identifying Critical Unmet Needs in AI

For an AI system to truly achieve “top-tier” status and be “universally used,” it must address fundamental limitations that currently restrict the widespread applicability and trustworthiness of existing deep learning models. This requires a strategic focus on next-generation AI paradigms.

• Interpretability (Explainable AI – XAI): A significant and persistent challenge in AI is the “black-box” nature of many deep learning models.⁷ Their internal decision-making processes are often opaque, making it difficult to understand, debug, or trust their outputs. This lack of transparency is a major impediment to their adoption in high-stakes applications, such as financial services or safety-critical systems, where explainability is frequently a legal requirement or essential for accountability.⁹ While Explainable AI (XAI) research aims to provide transparency, current XAI systems often fall short of practical developer needs.⁸
• Data Efficiency: A core limitation of current large AI models is their immense appetite for training data, often requiring trillions of tokens to achieve state-of-the-art performance.¹ This necessitates significant investment in data acquisition, presents challenges in quality control, raises privacy concerns, and contributes to a substantial environmental footprint.⁷ There is a pressing need for AI models capable of learning effectively from limited examples, thereby reducing reliance on vast, potentially biased, or low-quality datasets.¹⁶
• Advanced Reasoning and Common Sense: Despite their prowess in pattern recognition, deep learning models often exhibit weaknesses in higher-level cognitive functions. They typically lack common sense, struggle with complex, multi-step logical inference, and demonstrate limited global generalization beyond their specific training data.¹⁶ This constrains their effectiveness for tasks requiring long-term planning, true creativity, or imagination.¹⁶ The emerging fields of Causal AI and Neuro-Symbolic AI directly address these shortcomings. Causal AI aims to move beyond mere correlation to understand and apply cause-and-effect relationships, enabling “what-if” analyses and robust root cause identification, which are critical for deriving actionable insights.²⁰ Neuro-Symbolic AI seeks to integrate the strengths of deep learning (perception, pattern recognition) with symbolic AI (logic, rules, explicit knowledge representation) to achieve more robust reasoning, interpretability, and common-sense understanding.¹⁹
• Trustworthiness and Safety: Beyond technical performance, ethical considerations are paramount for “universally used” AI. Fairness, accountability, openness, privacy, and the potential for bias are central ethical challenges in the development of large language models (LLMs).²⁷ Foundational models, due to their generative capabilities, can amplify existing risks and introduce new ones, such as producing biased, unreliable, or toxic content derived from their training data.³ Mitigating these risks requires careful data filtering and the explicit encoding of ethical norms into the models.³

Current Landscape of Foundational AI

The current AI landscape is characterized by powerful foundational models developed by major technology players. Proprietary leaders include OpenAI’s GPT-4o, Google’s Gemini 2.5 Pro, and Anthropic’s Claude 3.7 Sonnet, which are recognized for their advanced multimodal capabilities and sophisticated reasoning.⁴

However, open-source alternatives are rapidly advancing, frequently narrowing the performance gap and offering substantial advantages in terms of flexibility, customization, and avoiding vendor lock-in, making them increasingly attractive for enterprise adoption.⁵ Notable open-source families include:

• Meta AI (Llama Series): The Llama 3 series (2025) is highly optimized for speed, efficiency, and custom deployment, making it a preferred choice for startups and researchers seeking full control over their AI systems.⁴ While Llama 2 and 3 models generally allow commercial use, Meta’s “open source” designation has been debated due to acceptable use policies.¹¹ The deployment of “Space Llama” on the International Space Station, operating without internet access, exemplifies its adaptability for sensitive, offline environments.³²
• Google’s Gemma Family: Derived from the same research and technology as the Gemini models, Gemma is a family of lightweight, open models (e.g., Gemma 3, CodeGemma, PaliGemma 2) available in various parameter sizes (e.g., 3B, 7B, 12B, 27B) for diverse generative AI tasks, including text and image input, code completion, and visual data processing.³³ These models are open-weight, allowing for customization through fine-tuning using various AI frameworks.³⁴
• DeepSeek-R1/V3: Models from this Chinese startup are gaining significant attention for their advanced reasoning capabilities and cost-effective architecture.⁴ DeepSeek V3-0324 (2025) has demonstrated remarkable performance, outperforming some proprietary models in key benchmarks and showcasing the viability of open-source solutions for latency-sensitive applications.³⁵ DeepSeek-R1 is specifically noted as an open-source reasoning model that rivals OpenAI’s o1 in math, reasoning, and code tasks.³⁶ DeepSeek models frequently employ Mixture of Experts (MoE) transformer architectures for enhanced efficiency and scalability.³⁶
• Microsoft Phi Family: These Small Language Models (SLMs) offer cost-effective, high-performance AI solutions, particularly suited for memory/compute constrained environments and latency-bound scenarios.¹² Phi-4, a 14B parameter model, is trained on a blend of synthetic datasets, filtered public domain websites, and academic Q&A, designed for general-purpose AI systems requiring reasoning and logic.¹²
• Alibaba Cloud’s Qwen Family: The Qwen 3 model family, released in April 2025, was trained on an enormous 36 trillion tokens across 119 languages and dialects.¹³ It includes both dense and sparse models and supports reasoning capabilities.¹³ Qwen-VL models integrate text generation with visual understanding, capable of OCR and advanced summarization.³⁸
• EleutherAI GPT-Neo/J/NeoX/Pythia: A grassroots non-profit, EleutherAI focuses on open-source AI research, aiming to replicate and advance GPT-3-like models.³⁹ Their models, such as GPT-NeoX-20B, were among the largest open-source language models at their time of release and have significantly contributed to fostering new AI startups.³⁹ The Pythia model suite is particularly notable for its fully public training data and reproducible training order, facilitating critical research into model behavior and biases.³⁹

Architectural Paradigms: The predominant architecture for large language models remains the transformer, often implemented as a decoder-only model.² However, advancements like the Mixture of Experts (MoE) architecture, as seen in DeepSeek V3, are gaining traction for their ability to improve efficiency and scalability by selectively activating different subsets of parameters.³⁶ Multi-Token Prediction (MTP) is another technique improving inference efficiency by enabling parallel token prediction.³⁶

Table 1: Comparative Overview of Leading Foundational AI Models (2025)

Model Name	Developer	Open-Source Status	Key Capabilities	Noteworthy Features	Primary License Type (if open-source)	Typical Use Case
GPT-4o	OpenAI	Proprietary	Text, Audio, Image, Reasoning	Multimodal, real-time voice interaction, improved accuracy	N/A	Diverse, real-time applications (voice assistants, customer service) 4
Gemini 2.5 Pro	Google DeepMind	Proprietary	Text, Code, Multimodal Understanding, Reasoning	1M token context window, integrated with Google ecosystems	N/A	Advanced enterprise applications, developers, data teams 4
Claude 3.7 Sonnet	Anthropic	Proprietary	Text, Code, Visual Understanding, Reasoning	200,000 token context window, cost-effective	N/A	Legal tech, academic research, enterprise knowledge management 4
DeepSeek-R1/V3	DeepSeek AI	Open-Weight/Permissive	Reasoning, Text, Code, Multilingual	MoE architecture, cost-effective, high speed (20 tokens/sec on Mac Studio)	MIT License (R1), Custom (V3) 5	Custom applications, scientific research, financial modeling, Asian language markets 5
OpenAI o1	OpenAI	Proprietary	Deeper, Slower, Structured Reasoning	Chain-of-thought, enhanced safety alignment	N/A	Mathematics, programming, scientific research, secure enterprise 5
OpenAI o3-mini	OpenAI	Proprietary	Strong Reasoning, STEM domains	Compact, adaptable depth of reasoning	N/A	Education, engineering, mid-sized applications, cost-efficiency 5
Meta AI (Llama Series)	Meta	Open-Weight (with AUP)	Text, Multimodal (Llama 3.2), Generative	Optimized for speed/efficiency, custom deployment, fine-tunable	Custom (disputed as fully open) 5	AI startups, researchers, developers seeking full control, offline environments 5
Google Gemini	Google	Proprietary	Text, Multimodal, Long-context	Advancements in long-context, multimodal, efficiency	N/A	General purpose, formidable competitor to GPT-4/Claude 3 4
Microsoft Phi Family	Microsoft	Open-Weight (MaaS/Free)	Text, Vision, Audio, Reasoning, Logic	Lightweight, cost-effective, low latency, edge deployment	Free for real-time deployment (Azure AI/Hugging Face) 12	Memory/compute constrained environments, real-time guidance, autonomous systems 12
Alibaba Cloud (Qwen Family)	Alibaba Cloud	Permissive (Apache 2.0)	Text, Image, Video, Audio, Reasoning, Multilingual	36 trillion tokens (Qwen 3), 119 languages, MoE, VL models	Apache 2.0, Qwen Research License, Qwen License 13	Text creation, translation, dialogue, programming assistance, visual understanding 38
EleutherAI (GPT-NeoX)	EleutherAI	Permissive (Apache 2.0)	Text Generation, NLP, Reasoning	Largest open-source GPT-3-like model at release, fully public training data (Pythia)	Apache 2.0 39	Research on language models, general-purpose text generation 39

The current landscape reveals that while proprietary models often lead in raw performance and multimodal integration, open-source alternatives are rapidly closing the gap, particularly in efficiency and specific reasoning capabilities. This suggests that a new, top-tier model could indeed emerge from an open-source foundation.

A significant observation in the current AI landscape is that the term “open-source” for foundational models often refers specifically to the availability of model weights, rather than full transparency regarding training data or methodologies.¹¹ For instance, while Meta’s Llama models are widely used, their license has been debated for its “acceptable use policy”.¹¹ Similarly, some Qwen models released only weights, not the dataset or training method.¹³ This prevalent practice creates a substantial challenge for a project aiming for “no outside resources.” If existing “open-source” foundational models do not provide fully open training data or detailed replication instructions, the developer cannot simply leverage these foundations for their data or training process. This implies that the project must undertake the immense burden of entirely new data collection and curation, significantly expanding its scope and resource requirements from day one.

Furthermore, the pursuit of “top-tier, universally used” AI, especially when built from scratch, inherently pushes the boundaries toward addressing current AI limitations such as interpretability, data efficiency, and advanced reasoning beyond mere correlation. For an AI to truly achieve universal utility, it must transcend the fundamental shortcomings that currently limit the widespread applicability and trustworthiness of existing deep learning models. While deep learning excels at pattern recognition, it often falls short in areas requiring explicit logical inference, common sense, and transparent decision-making. The emergence of Neuro-Symbolic AI and Causal AI paradigms directly addresses these fundamental cognitive and operational limitations, indicating that they represent the next frontier in AI development. Therefore, to genuinely achieve “top-tier” and “universally needed” status, the new model cannot merely be a larger or slightly more optimized version of existing deep learning models. Instead, it must strategically innovate in areas where current models are demonstrably weak, suggesting a strategic imperative to integrate principles from these next-generation AI paradigms from the very architectural design phase. This approach aims for a system that is not only powerful but also robust, explainable, and capable of more general intelligence.

III. The Genesis: Training a Foundational Model from Scratch (“No Outside Resources”)

The constraint of “no outside resources” when developing a top-tier foundational AI model fundamentally transforms the project from a conventional development effort into an undertaking comparable to those of major technology corporations. This section details the immense requirements for data, computational infrastructure, and the inherent limitations and ethical considerations that accompany such an endeavor.

The Scale of Data: Acquisition, Curation, and Preprocessing for Trillions of Tokens

Training a foundational model from scratch demands an unprecedented volume of diverse data, encompassing text, images, audio, and video.¹ To contextualize this, GPT-3, an earlier foundational model, utilized a vocabulary of over 50,000 tokens.¹⁰ More recent models like Llama 1 and 2 were trained on 1.4 trillion and 2 trillion tokens, respectively, while Alibaba’s Qwen 3 utilized an astonishing 36 trillion tokens, and Microsoft’s Phi-4 processed 9.8 trillion tokens.¹¹

The process of data collection itself is a monumental undertaking. It requires gathering comprehensive datasets that span a wide range of topics, contexts, and linguistic nuances to ensure the model’s broad understanding and applicability.¹ For example, EleutherAI’s “The Pile” dataset, an 825 GB English language corpus, was meticulously constructed from 22 distinct high-quality data subsets, including professional and academic sources.⁴⁰ This highlights the scale and effort involved in creating a truly foundational dataset.

Crucially, data preprocessing is indispensable for optimal model performance. This involves rigorous cleaning to eliminate low-quality, duplicated, or toxic data, correcting errors, and ensuring consistent formatting.¹ Tokenization, the process of converting raw text into numerical tokens, is a critical step, though its efficiency can vary significantly across languages, with some non-English languages requiring a suboptimal amount of tokens per word.¹⁰ In the future, data cleaning methodologies may also need to evolve to filter out content generated by other LLMs to maintain data quality and prevent model degradation.¹⁰

Computational Infrastructure: Hardware Requirements, Energy Consumption, and Financial Implications

Training foundational models necessitates vast computational resources, typically involving thousands of high-performance Graphics Processing Units (GPUs) or Tensor Processing Units (TPUs).¹ OpenAI’s GPT-3, for instance, was trained on a supercomputer utilizing thousands of GPUs over several weeks.¹ IBM Research developed Vela, an AI supercomputer equipped with thousands of Nvidia A100 GPUs specifically for such tasks.⁴³ More recently, Microsoft’s Phi-4, a 14B parameter model, required 1920 H100-80G GPUs for its training, which spanned 21 days.¹²

The training process itself is protracted, often extending over weeks or even months.¹ The financial cost associated with training a foundational model from scratch is substantial, ranging from tens of thousands to millions of dollars, directly correlating with the size of the dataset and the power of the compute resources utilized.⁴³ More powerful resources, while accelerating training, inherently incur higher costs.⁴³ For example, training GPT-3 (175 billion parameters) was estimated to cost around $4.6 million and would take 355 years on a single GPU.⁴⁶

This immense computational demand translates into significant energy consumption and a substantial carbon footprint.¹ A study by Stanford University indicated that training a large AI model can generate as much carbon dioxide as five cars over their lifetimes.¹

Beyond initial training, the deployment of large language models (LLMs) is also compute-intensive.⁴³ The overall AI infrastructure—the “backbone” supporting all machine learning operations—must encompass specialized hardware, robust data pipelines, efficient container orchestration (e.g., Kubernetes), comprehensive DevOps tools, and continuous monitoring solutions.⁴⁴ Key considerations for this infrastructure include scalability to handle growing complexity and data volumes, cost efficiency through optimal resource utilization, reliability to safeguard vital processes, and speed to market for rapid iteration.⁴⁴

Table 2: Estimated Resource Requirements for Training a Foundational AI Model (from scratch)

Model Size (Parameters)	Training Dataset Size (Tokens)	Estimated GPU/TPU Count (e.g., A100/H100 equivalents)	Estimated Training Duration	Approximate Training Cost (USD)	Estimated Carbon Footprint (CO2e, e.g., car lifetimes)
175 Billion (GPT-3)	400 Billion	Thousands of GPUs	Several Weeks	$4.6 Million	Equivalent to 5 cars over their lifetimes 1
14 Billion (Phi-4)	9.8 Trillion	1920 H100-80G GPUs	21 Days	High (Millions, inferred)	Significant (inferred) 12
36 Trillion (Qwen 3)	36 Trillion	Thousands of GPUs (inferred)	Weeks/Months (inferred)	Very High (Tens of Millions, inferred)	Very Significant (inferred) 13

The requirement to develop a “top-tier” foundational AI model with “no outside resources” places the project squarely in the domain of undertakings typically reserved for major technology corporations. This means facing the same astronomical costs, infrastructure challenges, and ethical responsibilities that these large entities grapple with. The sheer scale of data acquisition and curation, requiring trillions of tokens, coupled with the need for thousands of high-performance GPUs and months of dedicated training, necessitates an investment in the millions of dollars. This is not merely a technical hurdle; it represents a fundamental shift in the project’s nature from a personal or small-scale initiative to a large-scale enterprise. The developer must prepare for the organizational and financial implications of establishing a dedicated data acquisition and curation pipeline, building or acquiring a supercomputing cluster, and assembling a highly specialized and expensive team of AI/ML engineers and researchers. This also means embracing the ethical responsibilities, such as managing massive energy consumption, mitigating bias in self-collected data, and ensuring data privacy, that come with operating at such a scale.

Core Model Architecture Design: Deep Dive into Transformer-Based Models and Emerging Architectures

The foundational models that currently dominate the AI landscape are primarily built upon complex neural networks, with the transformer architecture being the cornerstone.² This architecture, characterized by its self-attention mechanisms, is central to the design of widely recognized models such as OpenAI’s GPT series and Meta’s Llama series.² Specific architectural enhancements frequently employed include the SwiGLU activation function, rotary positional embeddings (RoPE) for handling sequence order, and RMSNorm for normalization, all of which contribute to the efficiency and performance of these large models.¹¹

Beyond the standard transformer, emerging architectures like the Mixture of Experts (MoE) are gaining prominence. Models such as DeepSeek V3 utilize MoE to achieve greater efficiency and scalability by selectively activating different subsets of parameters for different inputs.³⁶ Another efficiency-boosting technique is Multi-Token Prediction (MTP), which allows the model to predict multiple tokens in parallel, significantly speeding up inference.³⁶ The overall performance of a large language model (LLM) after pretraining is fundamentally influenced by three key factors: the total computational resources expended during pretraining, the sheer size of the artificial neural network (measured by its number of parameters), and the volume of its pretraining dataset (i.e., the number of tokens in the corpus).¹⁰

Inherent Limitations and Ethical Considerations

The development and deployment of large-scale AI models, particularly from scratch, come with significant inherent limitations and ethical considerations that must be proactively addressed.

• Environmental Impact: The training and operation of these models consume substantial energy, directly contributing to a significant carbon footprint and considerable water usage.¹ This raises critical concerns about the long-term environmental sustainability of current AI development practices.
• Bias: A pervasive challenge is the propensity of AI models to inherit and amplify biases present in their training datasets. This can manifest as the over- or under-representation of certain demographic groups, the erasure of representation, or the reinforcement of demeaning and negative stereotypes.³ To mitigate this, developers must meticulously filter training data and explicitly encode ethical norms into their models.³
• Unreliability and Hallucination: Despite extensive training, foundational models can produce unreliable, nonsensical, inaccurate, or even toxic and inappropriate outputs.³ These “hallucinations” often stem from insufficient training data, incorrect assumptions, or inherent biases within the data.⁶ The risk of such errors is a significant concern for enterprises, leading to discussions about novel concepts like “AI hallucination insurance” to mitigate financial and reputational risks.⁶
• Lack of Comprehension/Interpretability: While these models can generate grammatically correct and factually plausible answers, they often struggle with true contextual comprehension and lack social or psychological awareness.³ This “black-box” nature makes it exceedingly difficult to understand how the model arrives at its predictions, hindering debugging and error identification.⁷
• Data Privacy and Security: The reliance on immense datasets for training large AI models inherently raises significant concerns regarding data privacy and security. The potential for misuse of sensitive data by malicious actors can lead to severe consequences, including identity theft, financial loss, and privacy invasion.⁷

The inherent limitations of current deep learning architectures, such as their “black-box” nature, data inefficiency, and lack of common sense reasoning, prevent them from achieving true “general intelligence” and universal trustworthiness. This drives the research into and development of hybrid AI paradigms like Neuro-Symbolic and Causal AI. This becomes a critical strategic path for the developer to achieve their “top-tier, universally used” goal, by creating a system that is not only powerful but also more intelligent, transparent, and reliable. If the aim is to create a groundbreaking and “top-tier” AI that achieves universal utility, simply scaling up a traditional deep learning model might not be sufficient. Instead, it is strategically advantageous to consider integrating principles and architectures from Neuro-Symbolic or Causal AI from the initial design phase. This approach would enable the system to exhibit more robust reasoning, greater interpretability, and a higher degree of general intelligence, distinguishing it from existing models and addressing critical unmet needs in the field.

IV. The Open-Source Imperative: Licensing, Community, and Enterprise Adoption

The commitment to developing an “open-source and free” AI software is a powerful statement, aligning with principles of accessibility and collaborative innovation. However, it requires a nuanced understanding of licensing frameworks, a proactive approach to community building, and a clear strategy for enterprise adoption.

Navigating Open-Source AI Licenses

Open-source software (OSS) is fundamentally defined by its source code being freely available for anyone to use, modify, and distribute, with the overarching goals of unrestricted access and collaboration.⁴¹ However, within the realm of large language models (LLMs), the term “open source” is not uniformly applied. Licenses come with varying levels of freedom and restrictions, particularly concerning commercial use, model weights, and the underlying training data.³⁰

• Proprietary Licenses: These licenses strictly limit access and usage, with the developing organization retaining full control over distribution. Commercial use is typically prohibited without explicit permission, as seen with OpenAI’s GPT models.³⁰
• Open Source Licenses (Key Categories): Licenses for AI model software generally fall into three categories:

• Permissive Licenses (e.g., Apache 2.0, MIT, BSD): These are the most flexible, allowing users to copy, modify, and distribute the source code with minimal obligations, usually limited to including copyright notices, attribution, and a disclaimer of liability.³⁰ They are generally considered the least risky for integration into commercial applications.⁴¹ Examples include Stable Diffusion (MIT License), DALL-E Mini (Apache-2.0), InvokeAI (Apache 2.0), HiDream-I1 (MIT License), DeepSeek-R1 (MIT License), Qwen 3 (Apache 2.0), and EleutherAI’s GPT-Neo models (Apache 2.0).¹³ FLUX.1 [schnell] is also released under Apache 2.0.⁴⁹
• Reciprocal (or Weak Copyleft) Licenses: These licenses require that any modifications to the original code inherit the same license terms. However, they typically allow for linking with proprietary code without “tainting” the entire proprietary codebase.⁴¹ The risk is reduced if no modifications are made to the licensed source code.⁴¹
• Restrictive (or Copyleft) Licenses (e.g., GPL 2.0, GPL 3.0, Affero GPL): These are the most stringent. They mandate that any modified code, and often any proprietary code that incorporates or links with the OSS code, must also be released under the same license terms. This can effectively “taint” proprietary code and necessitate its mandatory source code distribution.⁴¹ Examples include Fooocus and Unstable Diffusion, which use GPL-3.0.⁴⁸

• Hybrid Licenses: Some models operate under licenses that combine elements of both open source and proprietary terms. These might permit usage and modification but impose specific conditions, such as non-commercial use or a requirement to attribute the original creator.³⁰ Meta’s Llama series, while generally allowing commercial use for Llama 2 and 3, includes an acceptable use policy that has led to disputes regarding its classification as truly “open source” by the Open Source Initiative.¹¹ Similarly, Mistral models are released under a research license for non-commercial use, with a separate commercial license required for business applications.³⁰ FLUX.1 [pro] and [dev] also require direct contact with Black Forest Labs for commercial usage.⁴⁹
• Training Data Licenses: It is crucial to note that licenses for the training data can be separate from the model code and may impose additional restrictions.⁴¹ These can include attribution (e.g., CC BY, ODC-By), share-alike, no derivative works, or non-commercial licenses (e.g., CC BY-NC).⁴¹ For commercial applications, it is paramount to avoid licenses for training data or trained models that prohibit commercial use, or include share-alike/no derivative works restrictions if modifications are anticipated.⁴¹

Given this complexity, seeking assistance from a qualified lawyer to analyze all applicable licenses early in the development process is strongly advised to mitigate potential legal risks.⁴¹

Benefits of an Open-Source Ecosystem

Open-source AI models offer significant advantages, including greater flexibility in customization and deployment, particularly when deployed on-premises or within a private cloud, which allows organizations to maintain complete control over their data.³⁰ Beyond potential cost savings from avoiding licensing fees, open-source models provide greater control over data and systems, preventing vendor lock-in.⁴⁷

The open-source paradigm inherently fosters technical innovation by enabling diverse actors—from startups and universities to large tech giants—to contribute, leading to an additive effect that accelerates robust innovation across the field.⁵⁰ Platforms like Hugging Face serve as vital central hubs, hosting over 1.5 million public models and actively advocating for open-source development.⁵⁰ Open models, infrastructure, and scientific practices form the foundational bedrock of AI innovation, facilitating a vibrant ecosystem where researchers, companies, and developers can build upon shared knowledge.⁵⁰ This collaborative environment supports a dynamic commercial ecosystem, as companies can adapt and fine-tune open models to specific market needs, leading to tangible economic benefits.⁵⁰ Indeed, many organizations report a positive Return on Investment (ROI) from utilizing open-source AI tools.⁵⁰ For individual developers, experience with open-source AI tools is highly valued in the field, and working with such tools significantly contributes to job satisfaction.³¹ Ultimately, making AI openly available has the power to drive discoveries, enhance lives, and spur economic growth by democratizing access to immense opportunities.⁵² Meta, for instance, explicitly states that open-sourcing its AI models benefits the company by establishing its AI as a global standard.⁵²

Challenges and Risks of Open-Source Models

Despite the numerous benefits, open-source models present distinct challenges and risks:

• Security: Cybersecurity is a primary concern for organizations leveraging open-source AI, with 62% of respondents in one survey citing it as a challenge.³¹ A significant risk arises from the inability of original developers to control how open models are used or altered once released, making it challenging to track their dissemination and potential misuse, particularly by malicious actors.⁵³
• Intellectual Property (IP) Infringement: This is a substantial concern, cited by 50% of respondents ³¹, largely due to the variability and ambiguity in open-source licensing terms and the potential for “tainting” proprietary code with restrictive licenses.³⁰ The use of publicly available data, such as YouTube subtitles by EleutherAI’s “The Pile” dataset, has already drawn criticism and accusations of intellectual property theft.³⁹
• Regulatory Compliance: Adherence to evolving AI regulations is another significant concern, cited by 54% of respondents.³¹ The open-source nature of models like DeepSeek V3-0324 raises important ethical questions, and there are potential security risks if these powerful models are adapted for malicious purposes.³⁵
• Uncertainty and Ambiguity: There remains considerable uncertainty regarding the full spectrum of benefits and risks associated with open models, as well as how different actors will ultimately use them.⁵³ This ambiguity complicates risk assessment and mitigation.
• Quality of Service: While open-source models are rapidly improving, some, particularly smaller ones, may still exhibit limitations in areas such as multilingual support or performance in highly specialized domains compared to their proprietary counterparts.¹²

Strategies for Community Building and Sustainable Development

To ensure the long-term viability and impact of an open-source, foundational AI project, robust strategies for community building and sustainable development are essential. Fostering strong collaborations, both within academia and between academia and industry, is crucial for advancing open-source AI research and development.⁵⁴ Lowering the barrier to entry into the field through the development of user-friendly programming languages and tools can significantly increase participation and collaboration.⁵⁴

Platforms like Hugging Face serve as vital central hubs for the open-source machine learning community, offering model hubs, datasets, and “Spaces” for live demos, thereby fostering a highly collaborative environment.⁵⁰ These platforms actively organize hackathons, model competitions, and research collaborations to drive innovation.⁵¹ Establishing clear and transparent guidelines for data usage, including standardized protocols for anonymization, consent management, and usage tracking, is essential for ethical and legal compliance.⁵⁰ Public-private partnerships can play a key role in creating specialized data trusts for high-value domains, ensuring responsible data access while enabling innovation.⁵⁰ Beyond these, continuous support through robust documentation, responsive community forums, and regular updates will be critical for maintaining engagement and attracting a diverse pool of contributors. Developing a clear governance model that balances centralized direction with decentralized contributions will be key to managing the project’s evolution and ensuring its long-term health.

V. Quantum Computing Integration: The Frontier of AI

The potential integration of quantum computing into a top-tier AI system represents a visionary leap, aligning with the ambition to create a universally impactful technology. However, it is crucial to temper this vision with a realistic assessment of the current state and projected timeline of quantum AI.

Current State of Quantum Computing in AI

Quantum-enhanced machine learning (QML) involves executing machine learning algorithms for the analysis of classical data on a quantum computer.⁵⁵ By leveraging qubits and quantum operations, QML aims to improve computational speed and data storage capabilities for algorithms.⁵⁵ Quantum computers, with their ability to process vast amounts of data in parallel, can significantly accelerate machine learning algorithms, particularly in solving complex problems that classical computers struggle with, such as optimization, data analysis, and pattern recognition.⁵⁷

Applications where quantum-enhanced AI is already being explored include:

• Optimization Problems: This is one of the most promising areas, with potential to revolutionize supply chains, financial portfolios, and drug discovery, where classical computers are limited by complexity.⁵⁷ Quantum Approximate Optimization Algorithm (QAOA) and Quantum Annealing are specifically designed for such combinatorial optimization problems.⁵⁹
• Data Analysis and Pattern Recognition: Quantum computers can process large datasets and identify patterns that classical methods might miss.⁵⁷
• Natural Language Processing (NLP): Quantum algorithms could analyze vast amounts of text and context at higher speeds, improving voice assistants and translation services.⁵⁷
• Healthcare and Drug Discovery: Quantum computing can help AI analyze molecular data more efficiently, enabling faster drug discovery and personalized medicine by simulating complex molecular behavior.⁵⁷
• Finance and Risk Analysis: Quantum-enhanced AI offers new ways to analyze financial data, optimize investment portfolios, and assess risk, leading to smarter decisions.⁵⁷

Beyond enhancing classical AI, a symbiotic relationship exists where AI is also playing a key role in the development of quantum computing.⁵⁷ AI is being used to optimize quantum algorithms, help quantum systems solve problems more efficiently, improve how quantum computers detect and fix errors (leading to more reliable performance), and even assist in optimizing quantum chip design.⁵⁷

Quantum Machine Learning (QML) Algorithms and Architectures

Research in QML focuses on developing algorithms and architectures that harness quantum principles.

• Quantum Associative Memories and Pattern Recognition: These systems can recognize stored content based on similarity, even with incomplete or corrupted patterns. Unlike classical associative memories, quantum versions are free from cross-talk and have superior storage capacity.⁵⁵
• Quantum Neural Networks (QNNs): These computational models are based on quantum mechanics principles, aiming to combine classical artificial neural networks with quantum information advantages for more efficient algorithms.⁶¹ However, challenges exist, such as integrating non-linear activation functions within the linear mathematical structure of quantum theory and addressing the no-cloning theorem for communication between qubit layers.⁶¹ Theoretical studies, such as those by Google Quantum AI researchers, explore how quantum algorithms could learn certain neural networks more efficiently than classical methods, particularly for “periodic neurons”.⁶²
• Quantum-Enhanced Support Vector Machines (QSVMs): These benefit from quantum systems’ computational power to process high-dimensional datasets and find patterns.⁵⁸
• Hybrid Quantum-Classical Workflows: Given the current limitations of quantum hardware, platforms like PennyLane and Paddle Quantum integrate quantum computing with classical machine learning software (e.g., PyTorch, TensorFlow) to create hybrid workflows.⁵⁶ These allow developers to experiment on various quantum hardware providers through the cloud.⁵⁶

Practical Quantum Advantage and Timeline

“Quantum advantage” is defined as a quantum computer outperforming a classical computer on a commercially viable task.⁵⁹ As of today, quantum computing is still in the proof-of-concept phase and has not yet reached widespread practical quantum advantage.⁶⁰

The timeline for achieving significant breakthroughs and widespread commercial viability for quantum computing and its integration with AI is generally projected for the end of this decade and the beginning of the next.⁵⁹ Industry-scale implementations are potentially 5-10 years away from widespread availability.⁵⁹

• Historical Milestones: Foundations trace back to the 1960s-70s (Feynman, Wiesner), with David Deutsch proposing a universal quantum computer in the 1980s.⁶³ The 1990s saw breakthroughs in algorithms like Peter Shor’s for factoring and Lov Grover’s for unstructured search.⁶³ Early hardware demonstrations by IBM and D-Wave emerged in the 2000s.⁶³
• Recent Progress: In 2019, Google claimed “quantum supremacy” by demonstrating its processor could solve a specific problem faster than classical supercomputers, marking a pivotal proof-of-concept for quantum outperformance.⁶³ Widespread cloud-based access to quantum hardware (IBM Q, Azure Quantum, Amazon Braket) is also a significant recent development.⁶³
• Future Projections: Practical quantum computers with sufficient qubits and error correction for complex applications are likely 3-7 years away.⁵⁹ The Cloud Security Alliance strongly recommends enterprises achieve full “quantum-readiness” by April 14th, 2030.⁶³ NIST’s roadmap suggests deprecation of classical algorithms vulnerable to quantum attacks by 2030-2035.⁶⁴ IBM has an ambitious roadmap envisioning a quantum system with 200 qubits capable of running 100 million gates by 2029.⁶⁴

Challenges and Future Prospects

Despite immense potential, integrating quantum computing with AI faces significant challenges. Quantum computing is still in its infancy, and quantum hardware is far from universally accessible.⁵⁷ Quantum algorithms need further refinement to run effectively on real-world problems.⁵⁷ Technical hurdles for QNNs include overcoming “barren plateaus” in optimization, accurately discretizing real-valued data for quantum processing, and preparing the required quantum example states.⁶² There is a continuous need for better quantum hardware, including more stable qubits, improved coherence times, and larger qubit counts.⁶⁰

In the near future, AI may run on a combination of quantum and classical computers, maximizing their respective strengths through hybrid quantum-classical computing.⁶⁰ Significant research is still required to develop scalable QML algorithms that provide real-world advantages over classical AI.⁶⁰

As an alternative or complementary approach, neuromorphic computing is gaining attention. This field applies principles of neuroscience to computing systems to mimic the brain’s function and structure.⁵⁴ Neuromorphic chips have the potential to outpace traditional computers in energy and space efficiency, as well as performance, offering a promising solution to the untenable scaling of power-hungry AI systems.⁴⁶ Applications include scientific computing, AI, augmented/virtual reality, and robotics, all at a fraction of the energy consumed by general-purpose AI platforms.⁵⁴ Breakthroughs like the NeuRRAM chip demonstrate the potential for dynamic, versatile, and energy-efficient AI applications running directly in memory.⁵⁴

Quantum computing, while highly promising for AI, is still in its nascent stages for practical, widespread application, particularly for complex, general-purpose AI tasks. Its integration into a “top-tier, universally used” AI is therefore a long-term, speculative endeavor. The current limitations in hardware maturity, qubit stability, error correction, and the need for significant algorithm refinement mean that quantum advantage for broad AI applications is still several years away.

The “no outside resources” constraint profoundly impacts quantum integration. Developing quantum hardware from scratch is an undertaking of extraordinary complexity and cost, far exceeding even the immense resources required for classical foundational AI. Even accessing sufficient cloud-based quantum resources for foundational model training remains prohibitively expensive and specialized. This pushes the developer towards a hybrid or research-focused approach in the near term, rather than full-scale quantum-native AI development. The most realistic path involves exploring how quantum algorithms can enhance specific, computationally intensive components of a classical or hybrid AI model, or engaging in collaborative research with existing quantum computing initiatives, rather than building a quantum computer or a fully quantum-native AI from the ground up.

VI. Gameplan: From Vision to Execution

Realizing the vision of a new, top-tier, enterprise-grade, open-source, and free AI software built from scratch, with potential quantum integration, requires a meticulously phased gameplan. This blueprint acknowledges the immense challenges and positions the project for long-term success and impact.

Phase 1: Foundational Research and Strategic Planning (Years 0-1)

This initial phase is critical for laying a robust theoretical and strategic groundwork, particularly given the “no outside resources” constraint.

• Deep Dive into Unmet Needs: The first step involves a comprehensive analysis to prioritize specific, high-impact unmet needs in AI where the new software can offer a unique value proposition. This means focusing on areas where current deep learning models fall short, such as interpretability, data efficiency, or advanced reasoning capabilities. Identifying these gaps will guide the model’s core design and ensure its relevance for universal adoption.
• Architectural Blueprint: Based on the identified unmet needs, a detailed architectural blueprint must be designed. This should consider hybrid approaches, such as integrating Neuro-Symbolic or Causal AI principles from the outset. Such a design aims to transcend the limitations of purely sub-symbolic deep learning models, enabling more robust common sense, logical inference, and transparency. The architecture should be modular and extensible to allow for future quantum integration.
• Data Strategy: A comprehensive plan for ethical, high-quality data acquisition and curation from first principles is paramount. This acknowledges the immense scale required for a “top-tier” foundational model and the “no outside resources” constraint, meaning existing “open-source” datasets with opaque origins or restrictive licenses may not be suitable. The plan must define diverse data sources (text, image, audio, video), outline meticulous collection methodologies, and establish robust preprocessing pipelines for cleaning, tokenization, and quality control. Particular attention must be paid to mitigating biases in the collected data.
• Computational Infrastructure Assessment: A detailed analysis of hardware needs is essential. This involves identifying the specific GPU/TPU requirements, assessing power consumption, and exploring strategies for acquiring or building a supercomputing cluster given the “no outside resources” constraint. This phase must realistically evaluate the financial and logistical feasibility of such an undertaking, exploring energy-efficient hardware solutions and potentially innovative approaches to resource acquisition that align with the project’s ethos.
• Legal and Ethical Framework: Establish clear guidelines for data privacy, bias mitigation, and open-source licensing from day one. This includes defining data governance protocols, ensuring compliance with relevant regulations (e.g., GDPR), and proactively addressing potential ethical pitfalls. Engaging qualified legal counsel early in this phase is strongly recommended to navigate the complexities of open-source licenses, particularly concerning commercial use and the provenance of training data.

Phase 2: Core Model Development and Iteration (Years 1-5)

This phase focuses on the iterative development and scaling of the core AI model, emphasizing open-source principles and community engagement.

• Minimum Viable Product (MVP) Development: Begin by developing a smaller-scale model to validate architectural choices, data pipelines, and core functionalities. This MVP should focus on demonstrating proficiency in addressing a specific, high-impact unmet need identified in Phase 1. This iterative approach allows for early validation and refinement without committing to the full scale of a foundational model immediately.
• Iterative Training and Scaling: Gradually increase the model size and the volume of the curated dataset. Throughout this process, continuously monitor performance, identify and mitigate biases, and optimize resource consumption. Implementing robust MLOps (Machine Learning Operations) practices will be crucial for managing the lifecycle of the model, ensuring reproducibility, and facilitating efficient iteration.
• Open-Source Release Strategy: A carefully considered open-source release strategy is vital. This involves selecting a truly permissive license (e.g., Apache 2.0, MIT) for the model code to encourage widespread adoption and contribution, while ensuring that the licenses for the self-collected training data align with commercial use and avoid any “tainting” clauses. A clear contribution guideline and code of conduct should be established to foster a healthy and productive open-source community.
• Community Engagement: Actively foster a vibrant developer community through transparent communication, providing accessible tools and documentation, and leveraging collaborative platforms like Hugging Face. Organizing hackathons, workshops, and providing responsive support will encourage external contributions and accelerate development. The goal is to build a self-sustaining ecosystem around the model.
• Early Enterprise Partnerships: Seek out early partnerships with enterprises interested in co-developing or fine-tuning the model for specific use cases. These partnerships can provide valuable real-world validation, generate feedback, and potentially offer funding or computational resources that align with the “no outside resources” spirit, such as dedicated compute for specific research collaborations.

Phase 3: Advanced Capabilities and Quantum Horizon (Years 5+)

This long-term phase focuses on pushing the boundaries of AI, integrating advanced reasoning, and exploring the frontier of quantum computing.

• Refinement of Advanced Reasoning: Continuously enhance the Neuro-Symbolic or Causal AI components of the model. This involves ongoing research and development to improve common sense reasoning, explicit logical inference, and the model’s ability to learn from limited examples and generalize across domains. The aim is to move closer to artificial general intelligence (AGI) capabilities.
• Quantum AI Research Integration: Initiate dedicated research and development into quantum machine learning (QML) algorithms relevant to the model’s core tasks. This could involve exploring quantum algorithms for optimization, complex pattern recognition, or data analysis that could provide a “quantum advantage” over classical methods. A realistic approach would be to focus on hybrid quantum-classical computing, where quantum processors accelerate specific, computationally intensive sub-tasks of the larger AI model.
• Hardware Co-design and Exploration: Given the nascent stage of quantum hardware, engage in collaborations with quantum hardware developers. Explore the potential of specialized neuromorphic chips as a complementary or alternative approach for energy-efficient, brain-inspired computation, which could offer significant advantages in power consumption and performance for certain AI applications.
• Universal Adoption Strategy: Beyond technical performance, focus on strategies for truly universal adoption. This includes ensuring ease of deployment across various environments (cloud, on-premise, edge), providing robust documentation, developing user-friendly interfaces, and continuously improving the model’s performance, reliability, and safety.
• Long-Term Sustainability: To ensure the project’s longevity as a free and open-source entity, explore diverse funding models. This could include applying for research grants, securing enterprise support for specific features or integrations, or developing premium services (e.g., specialized fine-tuning, dedicated support, managed deployments) built on top of the free core model, without compromising its open-source nature.

VII. Conclusions and Recommendations

The ambition to create a new, top-tier, enterprise-grade, open-source, and free AI software from first principles, without external resources, and potentially integrating quantum computing, represents an extraordinary challenge. The analysis unequivocally indicates that this is an undertaking on par with the most resource-intensive projects conducted by major technology giants. The “no outside resources” constraint, coupled with the aspiration for “top-tier” performance, necessitates independent acquisition of trillions of tokens of data, the deployment of thousands of high-performance GPUs, and a multi-million dollar investment over several years. This demands a fundamental re-conceptualization of the project from a personal endeavor to a large-scale, multi-disciplinary enterprise from its inception.

To truly achieve “top-tier” and “universally used” status, the strategic imperative is to move beyond merely scaling up existing deep learning paradigms. The inherent limitations of current deep learning architectures—such as their “black-box” nature, data inefficiency, and lack of common sense reasoning—present critical unmet needs for general intelligence and trustworthiness. This necessitates a strategic focus on, and integration of, next-generation AI paradigms like Neuro-Symbolic AI and Causal AI. By designing an architecture that combines the strengths of deep learning with explicit reasoning and knowledge representation, the resultant system can be more intelligent, transparent, and reliable, thereby addressing fundamental cognitive shortcomings that limit current AI’s universal applicability.

The commitment to “open-source and free” is a powerful enabler for universal adoption and innovation. However, it requires meticulous navigation of the nuanced open-source licensing landscape, particularly distinguishing between “open-weights” and truly open training data. A permissive license for the core model code is crucial, but careful attention must be paid to the provenance and licensing of the vast training datasets to avoid legal pitfalls and ensure commercial viability. Robust community building, fostered through transparent communication, accessible tools, and collaborative platforms, will be essential for long-term sustainability and widespread contribution. Proactive ethical governance, addressing bias, data privacy, and potential misuse, must be embedded throughout the development lifecycle.

Regarding quantum computing integration, the analysis suggests it remains a long-term, high-potential, but currently speculative endeavor for general-purpose AI. While quantum-enhanced machine learning shows promise for specific optimization and pattern recognition tasks, practical quantum advantage for complex AI models is still several years away, with widespread industry adoption potentially 5-10 years out. The prohibitive cost and specialized nature of quantum hardware mean that a near-term strategy should focus on hybrid quantum-classical approaches or dedicated research collaborations, rather than attempting to build a quantum computer or a fully quantum-native AI from scratch. Neuromorphic computing, with its focus on energy efficiency and brain-inspired architectures, presents a compelling complementary or alternative avenue for future hardware innovation.

Recommendations:

1. Re-evaluate “No Outside Resources” for Infrastructure: Given the astronomical costs and technical expertise required for training a foundational model from scratch, consider strategic partnerships for computational resources (e.g., cloud credits, academic collaborations with supercomputing centers) that align with the “free and open” spirit without compromising ownership of the intellectual property. Building a dedicated supercomputing cluster from scratch is an undertaking of immense capital and operational expenditure.
2. Prioritize Next-Generation AI Paradigms: To achieve “top-tier” and “universally used” status, the core architectural design should integrate principles from Neuro-Symbolic AI and/or Causal AI from day one. This will enable the model to address critical unmet needs in reasoning, interpretability, and data efficiency, differentiating it from existing purely deep learning models.
3. Meticulous Data Strategy and Governance: Develop an exceptionally rigorous data acquisition and curation pipeline. Focus on high-quality, ethically sourced, and permissively licensed data. Implement advanced data cleaning techniques, including filtering out LLM-generated content. Establish clear internal governance policies for data usage, privacy, and bias detection from the outset.
4. Strategic Open-Source Licensing: Select a truly permissive open-source license (e.g., Apache 2.0) for the model’s code. Critically, ensure that all training data acquired or generated aligns with the commercial use and “free” nature of the project. Seek qualified legal counsel to navigate the complexities of data and model licensing to avoid future legal challenges.
5. Phased Quantum Integration: Approach quantum computing as a long-term research and development initiative. In the near term, focus on exploring hybrid quantum-classical algorithms for specific, computationally intensive sub-components of the AI model. Foster collaborations with leading quantum research institutions and consider the potential of neuromorphic computing as an energy-efficient alternative for future hardware.
6. Build a Core Team and Community: Recognize that this project requires a highly specialized, multi-disciplinary team encompassing AI/ML researchers, data engineers, software architects, and legal/ethical experts. Simultaneously, invest heavily in community building through transparent communication, accessible tools, and active engagement on platforms like Hugging Face to foster a vibrant and self-sustaining open-source ecosystem.

Works cited

1. Understanding Foundation AI Models: Training, Applications, and Impact – Ziton, accessed June 1, 2025, https://ziton.ca/understanding-foundation-ai-models-training-applications-and-impact/
2. The Guide to Foundation Models in Modern AI Development – Viso Suite, accessed June 1, 2025, https://viso.ai/deep-learning/foundation-models/
3. What are Foundation Models? – Generative AI – AWS, accessed June 1, 2025, https://aws.amazon.com/what-is/foundation-models/
4. Top 10 Most Popular AI Models in 2025 – SoluLab, accessed June 1, 2025, https://www.solulab.com/top-ai-models/
5. Top most popular AI models in 2025 – Varmeta, accessed June 1, 2025, https://www.var-meta.com/blog/top-most-popular-ai-models-in-2025
6. The Future of Artificial Intelligence | IBM, accessed June 1, 2025, https://www.ibm.com/think/insights/artificial-intelligence-future
7. Advantages and Disadvantages of Deep Learning | GeeksforGeeks, accessed June 1, 2025, https://www.geeksforgeeks.org/advantages-and-disadvantages-of-deep-learning/
8. Explainability for Embedding AI: Aspirations and Actuality – arXiv, accessed June 1, 2025, https://arxiv.org/html/2504.14631v1
9. The Urgency of Interpretability – Dario Amodei, accessed June 1, 2025, https://www.darioamodei.com/post/the-urgency-of-interpretability
10. Large language model – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Large_language_model
11. Llama (language model) – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Llama_(language_model)
12. microsoft/phi-4 – Hugging Face, accessed June 1, 2025, https://huggingface.co/microsoft/phi-4
13. Qwen – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Qwen
14. en.wikipedia.org, accessed June 1, 2025, https://en.wikipedia.org/wiki/Environmental_impact_of_artificial_intelligence#:~:text=The%20environmental%20impact%20of%20artificial,provide%20solutions%20to%20environmental%20problems.
15. Environmental impact of artificial intelligence – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Environmental_impact_of_artificial_intelligence
16. 7 Limitations of Deep Learning Algorithms of AI – Dr. Amit Ray, accessed June 1, 2025, https://amitray.com/7-limitations-of-deep-learning-algorithms-of-ai/
17. How to Use AI to Find Unmet Needs from Customer Interviews – Insight7, accessed June 1, 2025, https://insight7.io/how-to-use-ai-to-find-unmet-needs-from-customer-interviews/
18. Unmet Needs of Artificial Intelligence in Small Bowel Capsule Endoscopy – MDPI, accessed June 1, 2025, https://www.mdpi.com/2075-4418/15/9/1092
19. Neuro-Symbolic Artificial Intelligence: Integrating Learning and Reasoning – Alphanome.AI, accessed June 1, 2025, https://www.alphanome.ai/post/neuro-symbolic-artificial-intelligence-integrating-learning-and-reasoning
20. Causal AI: How cause and effect will change artificial intelligence – S&P Global, accessed June 1, 2025, https://www.spglobal.com/en/research-insights/special-reports/causal-ai-how-cause-and-effect-will-change-artificial-intelligence
21. Causal AI: Use Cases, Need, Benefits, Challenges and Strategies – LeewayHertz, accessed June 1, 2025, https://www.leewayhertz.com/causal-ai/
22. Example of Causal AI: Applications & Insights 2025 – BytePlus, accessed June 1, 2025, https://www.byteplus.com/en/topic/537612
23. Causal AI: Use Cases, Benefits, Challenges and Implementation – A3Logics, accessed June 1, 2025, https://www.a3logics.com/blog/causal-ai-use-cases/
24. Neuro-Symbolic AI in 2024: A Systematic Review – arXiv, accessed June 1, 2025, https://arxiv.org/pdf/2501.05435
25. Neuro-Symbolic AI in 2024: A Systematic Review – arXiv, accessed June 1, 2025, https://arxiv.org/html/2501.05435v1
26. Neuro-symbolic AI – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Neuro-symbolic_AI
27. gaper.io, accessed June 1, 2025, https://gaper.io/ethical-considerations-llm-development/#:~:text=What%20are%20the%20ethical%20considerations,ethical%20issues%20in%20LLM%20development.
28. Ethical Considerations in LLM Development – Gaper.io, accessed June 1, 2025, https://gaper.io/ethical-considerations-llm-development/
29. Expanding on ethical considerations of foundation models – IBM, accessed June 1, 2025, https://www.ibm.com/think/insights/expanding-on-ethical-considerations-of-foundation-models
30. Open Source LLMs – Nebuly, accessed June 1, 2025, https://www.nebuly.com/blog/open-source-llms
31. Open source technology in the age of AI – McKinsey, accessed June 1, 2025, https://www.mckinsey.com/capabilities/quantumblack/our-insights/open-source-technology-in-the-age-of-ai
32. Space Llama: Meta’s Open Source AI Model is Heading Into Orbit, accessed June 1, 2025, https://about.fb.com/news/2025/04/space-llama-metas-open-source-ai-model-heading-into-orbit/
33. Gemma models overview | Google AI for Developers – Gemini API, accessed June 1, 2025, https://ai.google.dev/gemma/docs
34. Use Gemma open models | Generative AI on Vertex AI – Google Cloud, accessed June 1, 2025, https://cloud.google.com/vertex-ai/generative-ai/docs/open-models/use-gemma
35. DeepSeek V3-0324: The First Open-Source AI to Beat Non-Reasoning Models! – Hyperight, accessed June 1, 2025, https://hyperight.com/deepseek-v3-0324-the-first-open-source-ai-to-beat-non-reasoning-models/
36. DeepSeek AI: Open-Source Models Revolutionizing Language, Reasoning, and Multimodal AI – Encord, accessed June 1, 2025, https://encord.com/blog/deepseek-ai/
37. Phi Open Models – Small Language Models | Microsoft Azure, accessed June 1, 2025, https://azure.microsoft.com/en-us/products/phi
38. Qwen LLMs – – Alibaba Cloud Documentation Center, accessed June 1, 2025, https://www.alibabacloud.com/help/en/model-studio/what-is-qwen-llm
39. EleutherAI – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/EleutherAI
40. Eleuther AI NLP Models – Originality.ai, accessed June 1, 2025, https://originality.ai/blog/eleutherai-nlp-models
41. Open Source Licensing Considerations for Artificial Intelligence Application – Knobbe Martens, accessed June 1, 2025, https://www.knobbe.com/wp-content/uploads/2025/04/Legaltech-News-Open-Source-Licensing.pdf
42. New “”Open-Source”” Video generation model – LocalLLaMA – Reddit, accessed June 1, 2025, https://www.reddit.com/r/LocalLLaMA/comments/1kgrjor/new_opensource_video_generation_model/
43. What it Takes to Train a Foundation Model – AI with Armand, accessed June 1, 2025, https://www.armand.so/what-it-takes-to-train-a-foundation-model/
44. Build AI Infrastructure: A Practical Guide – Mirantis, accessed June 1, 2025, https://www.mirantis.com/blog/build-ai-infrastructure-your-definitive-guide-to-getting-ai-right/
45. AI infrastructure: 5 key components, challenges and best practices – Spot.io, accessed June 1, 2025, https://spot.io/resources/ai-infrastructure/ai-infrastructure-5-key-components-challenges-and-best-practices/
46. What is the future of artificial intelligence beyond deep learning? – Quora, accessed June 1, 2025, https://www.quora.com/What-is-the-future-of-artificial-intelligence-beyond-deep-learning
47. Part 1 – Open Source AI Models: How Open Are They Really?, Legaltech News, accessed June 1, 2025, https://www.hunton.com/insights/publications/part-1-open-source-ai-models-how-open-are-they-really
48. Best Open Source AI Image Generator Models in 2025, accessed June 1, 2025, https://www.aiarty.com/ai-image-generator/open-source-ai-image-generator.htm
49. A Guide to Open-Source Image Generation Models – BentoML, accessed June 1, 2025, https://www.bentoml.com/blog/a-guide-to-open-source-image-generation-models
50. Hugging Face calls for open-source focus in the AI Action Plan – AI News, accessed June 1, 2025, https://www.artificialintelligence-news.com/news/hugging-face-open-source-focus-ai-action-plan/
51. Hugging Face: The Open-Source Hub for Machine Learning Innovation – AI Expert Network, accessed June 1, 2025, https://aiexpert.network/hugging-face/
52. Open Source AI | Meta, accessed June 1, 2025, https://ai.meta.com/temp/open-source-ai/
53. Open Foundation Models: Implications of Contemporary Artificial Intelligence | Center for Security and Emerging Technology, accessed June 1, 2025, https://cset.georgetown.edu/article/open-foundation-models-implications-of-contemporary-artificial-intelligence/
54. Scaling up Neuromorphic Computing for More Efficient and Effective AI Everywhere and Anytime, accessed June 1, 2025, https://today.ucsd.edu/story/scaling-up-neuromorphic-computing-for-more-efficient-and-effective-ai-everywhere-and-anytime
55. Quantum machine learning – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Quantum_machine_learning
56. Quantum Machine Learning: What It Is, How It Works, and More | Coursera, accessed June 1, 2025, https://www.coursera.org/articles/quantum-machine-learning
57. Quantum Computing and AI: The Future of Innovation – SpinQ, accessed June 1, 2025, https://www.spinquanta.com/news-detail/quantum-computing-and-artificial-intelligence
58. Redefining Intelligence: The 7 Best Quantum Machine Learning Tools – BlueQubit, accessed June 1, 2025, https://www.bluequbit.io/quantum-machine-learning-tools
59. Quantum Computing: The Future Of AI Employee Scheduling – myshyft.com, accessed June 1, 2025, https://www.myshyft.com/blog/quantum-computing-applications/
60. Your Quick Guide to Quantum and AI: The Future of Computing or Just Hype?, accessed June 1, 2025, https://meetiqm.com/blog/quantum-ai-the-future-of-computing-or-just-hype/
61. Quantum neural network – Wikipedia, accessed June 1, 2025, https://en.wikipedia.org/wiki/Quantum_neural_network
62. Google Researchers Say Quantum Theory Suggests a Shortcut for Learning Certain Neural Networks, accessed June 1, 2025, https://thequantuminsider.com/2025/03/31/google-researchers-say-quantum-theory-suggests-a-shortcut-for-learning-certain-neural-networks/
63. Quantum computing timeline & when it will be available – Sectigo, accessed June 1, 2025, https://www.sectigo.com/resource-library/quantum-computing-timeline-things-to-know
64. Understanding the timeline of quantum computing: when will it become reality?, accessed June 1, 2025, https://securityboulevard.com/2025/05/understanding-the-timeline-of-quantum-computing-when-will-it-become-reality/
65. today.ucsd.edu, accessed June 1, 2025, https://today.ucsd.edu/story/scaling-up-neuromorphic-computing-for-more-efficient-and-effective-ai-everywhere-and-anytime#:~:text=Applications%20for%20neuromorphic%20computing%20include,efficiency%2C%20as%20well%20as%20performance.
66. What Is Neuromorphic Computing? – IBM, accessed June 1, 2025, https://www.ibm.com/think/topics/neuromorphic-computing