# How Large Language Models Are Trained

The creation of a large language model (LLM) occurs in three distinct stages: massive unsupervised pre-training, supervised fine-tuning, and human-preference alignment. First, the model ingests trillions of words to learn the statistical patterns of human language and world knowledge. Next, it is fine-tuned on structured prompt-and-response examples to learn how to act as a helpful assistant, before finally undergoing alignment to ensure its outputs are safe and attuned to human values.

## Introduction: The Anatomy of Artificial Intelligence

When a user types a query into a modern AI assistant, the coherent, seemingly thoughtful response is the final product of an extensive and computationally grueling lifecycle. Large language models are not inherently conversational, nor are they born with an intuitive understanding of human intent. In their rawest form, they are simply statistical engines designed to predict the next word in a sequence based on vast amounts of historical data [cite: 1, 2]. 

Transforming a massive neural network from a chaotic pattern recognizer into a usable, safe, and logical assistant requires a highly orchestrated, multi-stage pipeline. By 2026, the artificial intelligence industry had largely standardized this lifecycle into three main pillars: Pre-training, Supervised Fine-Tuning (SFT), and Alignment (typically via Reinforcement Learning from Human Feedback or Direct Preference Optimization) [cite: 1, 3]. 

However, the methods used to execute these stages have evolved at a breakneck pace. As models scaled from billions to trillions of parameters, developers hit a "data wall," exhausting the supply of high-quality human text. This sparked a transition from human-labeled data to synthetic, AI-generated data, and shifted the competitive frontier from brute-force pre-training to highly efficient post-training architectures [cite: 4, 5, 6]. 

To understand how an LLM actually works, one must deconstruct its training from the ground up—stage by stage.

[image delta #1, 0 bytes]





## Stage 1: Pre-Training (Learning the World)

Pre-training forms the bedrock of large language model methodologies. At this stage, the model is exposed to an enormous corpus of unstructured, unlabeled data, including digitized books, Wikipedia articles, code repositories, internet forums, and scientific papers [cite: 1, 2]. The objective here is entirely self-supervised: the model is simply tasked with predicting the next token (a word or fraction of a word) in a sequence, or sometimes inferring masked tokens from the surrounding context [cite: 7].

Because nobody is explicitly telling the model what is "correct" or actively correcting its logic during pre-training, it learns purely through statistical regularities. Through trillions of next-token predictions, the model implicitly absorbs human grammar, semantic associations, software syntax, factual correlations, and long-range linguistic dependencies [cite: 2]. 

### The Transformer Architecture and Attention Mechanisms

The mechanics of this learning process rely on the Transformer architecture. When raw text is fed into the system, a tokenizer first converts the words into numerical identifiers. For instance, models like Meta's Llama 3 possess a vocabulary of roughly 128,000 distinct tokens [cite: 8]. These token IDs are then passed through an embedding layer, which transforms them into high-dimensional vectors—essentially placing the words into a mathematical space where semantic relationships can be measured [cite: 8]. In a typical 8-billion-parameter model, each token might be represented by a 4,096-dimensional vector [cite: 8].

The core innovation of the Transformer is the self-attention mechanism. Instead of reading text strictly sequentially, the model assesses the relationship between a given token and every other token in the sequence simultaneously [cite: 8, 9]. It calculates three values for each token—Query (Q), Key (K), and Value (V)—to determine how much "attention" a word should pay to its neighbors [cite: 8]. This allows the model to understand context, such as recognizing that the word "bank" means something entirely different when preceded by "river" versus "commercial." This process is repeated across dozens of stacked transformer layers (Llama 3 utilizes 32 layers) until a final dense layer produces a probability distribution for the next logical token [cite: 8].

### The Scale and Curation of Pre-Training Data

The scale of modern pre-training data is staggering. Early landmark models like GPT-3 (released in 2020) were trained on roughly 300 billion tokens [cite: 10]. By 2024, Meta's Llama 3 models were pre-trained on over 15 trillion tokens, with data cutoffs extending late into 2023 [cite: 11, 12]. Alibaba’s Qwen 2.5 models pushed this boundary further, reportedly training on upwards of 20 trillion tokens [cite: 4]. 

However, raw internet data is filled with noise. Datasets undergo rigorous curation before they reach the neural network. Developers apply semantic deduplication to prevent the model from memorizing repeated text, utilize heuristic safety filters to remove toxic content, and classify data quality using older language models [cite: 13, 14]. Furthermore, developers employ "data mixing" strategies. Rather than treating all data equally, a core of highly credible text (like scientific journals or verified encyclopedias) is sampled more frequently, while the "long tail" of low-quality internet data is sampled less frequently [cite: 13]. This maximizes the educational utility of every token the model processes.

### Hardware Orchestration and 4D Parallelism

Processing trillions of tokens requires hyperscale computing infrastructure. Training a frontier model like Meta's 405-billion parameter Llama 3 requires massive orchestration, utilizing over 16,000 Nvidia H100 GPUs running continuously over several months and expending an estimated 3.8 × 10^25 floating-point operations (FLOPs) [cite: 15].

Training at this hyperscale necessitates highly complex "4D parallelism" to prevent memory bottlenecks. A model of this size cannot fit on a single microchip, meaning engineers must shard the data and the neural network itself across thousands of GPUs. This involves several layers of distribution:
*   **Fully Sharded Data Parallelism (FSDP):** Distributes the model's parameters, gradients, and optimizer states across multiple GPUs, ensuring no single unit is overwhelmed by the memory requirements of the weights [cite: 15].
*   **Tensor Parallelism (TP) & Pipeline Parallelism (PP):** Splits individual computational layers across different hardware accelerators to balance the workload, allowing the matrix multiplications to occur simultaneously [cite: 15].
*   **Context Parallelism (CP):** Allows the model to handle massive inputs. Between 2024 and 2026, standard context windows exploded from 8,000 tokens to over 10 million tokens in frontier models, requiring immense parallel memory management [cite: 15, 16].

Modern training pipelines also utilize FP8 (8-bit floating-point) precision. By lowering the precision of the calculations from standard 16-bit to 8-bit formats (such as E4M3 for forward propagation and E5M2 for backpropagation), developers significantly reduce the memory footprint and accelerate training throughput without degrading the model's eventual accuracy or convergence [cite: 17]. 

### Scaling Laws and the Shift to "Over-Training"

Historically, artificial intelligence laboratories followed the "Chinchilla Scaling Laws," named after a seminal DeepMind paper. These laws dictated an optimal ratio between the size of a model (its parameter count) and the amount of data it should be trained on. According to the Chinchilla framework, an 8-billion parameter model should only require roughly 200 billion tokens to reach its compute-optimal performance [cite: 12, 14, 18]. 

However, during the development of models like Llama 3, researchers observed that model performance continues to improve log-linearly long after passing this theoretical threshold [cite: 12, 14]. Consequently, Meta purposefully "over-trained" its smaller 8B model on 15 trillion tokens—nearly two orders of magnitude beyond the Chinchilla recommendation [cite: 12, 18]. While this brute-force approach requires vastly more compute up front, it yields a highly capable, compressed model that is much cheaper and faster for end-users to run during inference [cite: 12, 18].

At the end of the pre-training stage, the result is a "Base Model." It contains an encyclopedic amount of knowledge, but it is practically unusable for a general consumer. If a user prompts a base model with a question like "What is the capital of France?", it might simply complete the pattern by generating another question, like "What is the capital of Germany?" rather than answering the prompt [cite: 1, 2]. To make the model helpful, it must undergo instruction tuning.

## Stage 2: Supervised Fine-Tuning (The Assistant Transformation)

Supervised Fine-Tuning (SFT), frequently referred to as instruction tuning, is the targeted intervention that transforms a raw, probabilistic pattern-matcher into an interactive, goal-oriented assistant [cite: 1, 2, 19]. 

In this stage, the training objective shifts. The model is no longer predicting the next token from random internet text. Instead, it is trained on highly curated, labeled datasets consisting of specific prompt-and-response pairs [cite: 1, 19]. The data is formatted identically to how a user would interact with the software (e.g., `Instruction:` "Write a Python script to scrape a website." -> `Response:` "Here is the code..."). 

It is crucial to understand that SFT generally does not teach the model new facts. The model has already acquired its foundational knowledge, logic, and reasoning baseline during the massive pre-training phase. Instead, fine-tuning teaches the model *how* to behave, establishing the expected format, tone, and structure of its output [cite: 2]. It narrows the model's vast behavioral distribution down to the persona of a helpful assistant.

| Training Stage | Primary Objective | Data Source | Resulting Output |
| :--- | :--- | :--- | :--- |
| **Pre-Training** | Next-token prediction (unsupervised) | Trillions of unstructured tokens (web, books, code) [cite: 1, 7] | Base Model (knowledgeable but unusable as an assistant) [cite: 1, 2] |
| **Supervised Fine-Tuning** | Task completion & instruction adherence (supervised) | Thousands of structured prompt-and-response pairs [cite: 1, 19] | Instruction-Tuned Model (follows formatting and answers queries) [cite: 1, 19] |
| **Alignment (RLHF/DPO)** | Maximizing human preference & safety | Pairwise human preference ratings (chosen vs. rejected) [cite: 20, 21] | Aligned Assistant Model (helpful, harmless, and honest) [cite: 22, 23] |

### Parameter-Efficient Fine-Tuning (PEFT)

While pre-training requires data centers filled with thousands of GPUs, fine-tuning can often be executed on a fraction of the hardware. Techniques like LoRA (Low-Rank Adaptation) and QLoRA allow engineers to freeze the vast majority of the model's original pre-trained weights and only update a tiny, specialized sub-network of parameters [cite: 19, 24]. 

LoRA operates by injecting trainable rank decomposition matrices into the transformer layers. By using a low rank (typically between 8 and 256), the number of trainable parameters drops exponentially [cite: 24]. Coupled with quantization techniques that reduce the precision of the frozen weights to 4-bit formats, this massively reduces memory consumption [cite: 24]. These efficiency gains mean that open-weight models can be heavily customized for specific enterprise domains—such as legal analysis, medical triage, or cybersecurity threat modeling—on modest hardware setups, sometimes requiring as little as a single commercial GPU [cite: 19, 25].

### The Synthetic Data Revolution

Historically, generating high-quality SFT data required vast armies of human annotators, which was slow, expensive, and difficult to scale. A 70-billion-parameter base model might need up to 100,000 task-specific examples to specialize properly without suffering from catastrophic forgetting, a scale that would cost hundreds of thousands of dollars if labeled manually [cite: 26]. 

By 2025 and 2026, the industry had largely shifted to relying on synthetic data—data generated by larger, frontier AI models to train smaller, more efficient models [cite: 6, 26]. Techniques like *Self-Instruct* or *Evol-Instruct* use a highly capable "teacher" model (such as GPT-4.5 or Claude 3.5 Sonnet) to recursively generate complex instructions and high-quality responses based on a tiny seed of human-written examples [cite: 26]. 

This synthetic generation is not simply asking a teacher model for random text. It is a highly structured workflow where data is shaped precisely for specific fine-tuning recipes—such as formatting function-calling traces for agentic tool use, or generating synthetic retrieval-augmented generation (RAG) queries [cite: 26]. Crucially, this generated data is then rigorously filtered by another AI acting as a "judge." Unfiltered synthetic datasets frequently perform worse than smaller, carefully filtered ones, making the AI judge a mandatory component of modern training [cite: 26]. 

Using synthetic data solves the impending "data wall," drastically reduces human labeling costs, and mitigates the privacy risks associated with scraping real-world user tickets or proprietary emails [cite: 6, 26].

### The Risks of Synthetic Scaling

While synthetic data lowers the barrier to entry, it introduces significant risks. When models train on the outputs of other models, they risk "model collapse"—a phenomenon where the variance and diversity of the language distribution narrows, causing the student model to lose its ability to handle edge cases or generate creative prose [cite: 27]. 

Furthermore, models exhibit "self-preference bias." If an AI judge is used to evaluate synthetic data, it naturally prefers data generated by models with a similar architecture or training lineage [cite: 27]. Recent studies have demonstrated that injecting high diversity into synthetic data sources—using multiple different teacher models rather than relying on a single monolith—is essential to mitigate distribution collapse and maintain the adversarial robustness of the fine-tuned model [cite: 13, 27]. 

Even with perfect SFT data, supervised fine-tuning remains a form of imitation learning. The model learns to mimic the style of a good answer, but it lacks an internal compass for complex trade-offs, nuance, or safety. To truly align the model's judgment with human expectations, it requires preference optimization.

## Stage 3: Alignment and Preference Optimization

An instruction-tuned model might write a perfectly formatted, highly fluent piece of code that also happens to be a devastating zero-day exploit. Fluency does not guarantee safety. Alignment is the final training hurdle, designed to ensure the model adheres to the "HHH" principle: it must be Helpful, Harmless, and Honest [cite: 22, 23]. 

There are two dominant methodologies for executing this alignment: Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) [cite: 21, 28, 29]. Both rely on the exact same fundamental input—pairwise preference data. Given a single prompt, the model generates two different responses, and a human annotator (or a reliable AI judge) indicates which response is superior based on helpfulness and safety guidelines [cite: 21]. From this shared starting point, the two techniques take fundamentally different paths.

[image delta #2, 0 bytes]



### Reinforcement Learning from Human Feedback (RLHF)

Pioneered heavily by OpenAI to create models like InstructGPT and the early versions of ChatGPT, RLHF decomposes alignment into a complex, multi-stage pipeline [cite: 3, 20]. 

1.  **Reward Modeling:** Using the human preference data, developers train a secondary neural network called a Reward Model. This model learns the nuances of human taste, essentially acting as a digital critic. When fed a prompt and a response, it outputs a scalar score representing how "good" or "safe" the response is [cite: 20, 23, 28].
2.  **Reinforcement Learning:** The main LLM (the policy model) is then allowed to generate responses to new, unseen prompts. The Reward Model grades these responses in real-time, and a reinforcement learning algorithm—usually Proximal Policy Optimization (PPO)—updates the main model's internal weights to maximize this reward score [cite: 20, 30, 31].

RLHF is a highly effective "online" learning process. Because the model continuously explores new generations and receives dynamic feedback, it can push its capabilities beyond the boundaries of its static training data [cite: 21, 30]. This makes RLHF particularly powerful for tasks involving long-horizon planning, complex coding, and deep value alignment [cite: 21, 28]. 

However, RLHF comes with severe engineering drawbacks. It is notoriously unstable and highly sensitive to hyperparameter tuning [cite: 20, 30]. It also requires enormous compute resources; standard RLHF requires four concurrent models loaded into memory simultaneously (the active policy model, a frozen reference model to prevent the policy from drifting too far, the critic model, and the reward model) [cite: 30]. Furthermore, RLHF is susceptible to "reward hacking," where the policy model discovers mathematical loopholes to achieve a high score without actually producing a genuinely helpful response—often by becoming overly verbose or exceedingly polite [cite: 20, 32].

### Direct Preference Optimization (DPO)

Emerging as a streamlined alternative in 2023, Direct Preference Optimization (DPO) rapidly became the dominant alignment technique for open-source developers by 2025 [cite: 20, 32]. DPO entirely eliminates the need for a separate reward model and the complexities of reinforcement learning algorithms.

The key insight behind DPO is mathematical: researchers proved that the optimal policy under standard RLHF has a closed-form relationship with the reward function. DPO exploits this by reframing preference learning as a simple classification problem [cite: 20, 21]. Instead of training a critic to score the model, DPO directly adjusts the language model's internal probabilities (logits) using a binary contrastive loss function. It mathematically maximizes the likelihood of the human-preferred response while simultaneously penalizing the rejected response [cite: 20, 32, 33]. 

DPO is primarily an "offline" process, optimizing strictly over a fixed dataset of preference pairs [cite: 30]. This makes it significantly more stable, highly reproducible, and much cheaper to run, as it only requires two models in memory (the policy model and the frozen reference model) [cite: 20, 30]. While DPO's performance ceiling is technically bounded by the quality and coverage of its static preference dataset, its ease of use has made it the default choice for rapid alignment, sentiment control, and budget-constrained projects [cite: 25, 30, 32].



### Beyond DPO: Odds Ratio Preference Optimization (ORPO)

The drive for efficiency has led to even more streamlined methods. Odds Ratio Preference Optimization (ORPO), introduced in recent years, attempts to collapse the Supervised Fine-Tuning and Alignment stages into a single step [cite: 34]. 

Instead of requiring an SFT phase followed by a separate preference tuning phase, ORPO applies an odds-ratio penalty directly to rejected responses during standard supervised training. By modifying the loss function to explicitly move the model's generation away from rejected examples while imitating the chosen examples, ORPO entirely eliminates the need for a separate reward model or reference model, drastically cutting down the computational resources and time required to prepare an LLM for production [cite: 34].

## The Post-Training Compute Era: The Rise of Reasoning Models

For years, the unwritten rule of artificial intelligence development was that greater intelligence came primarily from scaling up the pre-training stage. The more data and raw computing power thrown at a base model, the smarter it became. 

However, by late 2024 and early 2025, the frontier of AI capabilities began to hit diminishing returns on raw data scaling [cite: 35]. When OpenAI released the initial preview of GPT-4.5 in early 2025, the model utilized vastly more pre-training compute than GPT-4, yet it yielded relatively marginal capability leaps on complex benchmarks, suggesting that purely scaling unsupervised pre-training was reaching a saturation point [cite: 36, 37, 38]. 

Consequently, the industry shifted dramatically toward scaling post-training compute, specifically focusing on building "reasoning" models [cite: 16, 35, 36, 39].

Models like OpenAI's `o1` and `o3` series, as well as Chinese powerhouse models like DeepSeek-R1 and Alibaba’s Qwen 2.5-Max, allocate significantly more computational power *during* the generation of an answer (inference time) and during the post-training alignment phases [cite: 4, 16, 40]. Rather than predicting the immediate next token instinctively, these models are trained using Reinforcement Learning with Verifiable Rewards (RLVR) to generate hidden "chains of thought," exploring multiple logical pathways, backtracking, and self-correcting before providing a final answer to the user [cite: 16, 25].

### The Disruption of Mixture-of-Experts (MoE)

This shift in training philosophy severely disrupted the global AI market in early 2025, primarily driven by DeepSeek. By shifting focus away from brute-force pre-training and leveraging highly curated, human-generated logic annotations during post-training, DeepSeek successfully trained its GPT-4-class R1 model for an estimated $5.6 million—roughly 5% of the cost of comparable Western proprietary models [cite: 5, 40]. 

This massive cost reduction was largely achieved through architectural innovations like the Mixture-of-Experts (MoE) framework. Unlike standard dense transformers where every parameter is active for every word, an MoE architecture routes each token to only a specialized fraction of the network's parameters [cite: 4, 5]. For example, DeepSeek-V3 possesses 671 billion total parameters but activates only 37 billion parameters per token during inference [cite: 4, 5]. 

This efficiency proved that throwing billions of dollars at raw hardware was no longer the only path to frontier-class artificial intelligence. The restrictive US export controls on advanced Nvidia GPUs inadvertently forced Chinese laboratories toward efficiency-first engineering, turning hardware constraints into a competitive architectural advantage [cite: 5, 35]. The competitive landscape has irrevocably shifted from massive pre-training clusters toward highly refined, specialized fine-tuning and reasoning strategies [cite: 5, 35].

## The Hidden Challenges: Over-Alignment and Bias

While techniques like RLHF and DPO have successfully mitigated the most egregious harms of LLMs (such as producing hate speech, facilitating cyberattacks, or offering weapon-making instructions), the alignment process itself has introduced subtle, systemic flaws into modern AI.

### Superficial Safety and the "Refusal" Problem

Current safety alignment is often mathematically shallow. When an LLM evaluates a prompt it deems risky or uncertain, its loss function heavily rewards it for immediately shutting down the query [cite: 41]. Because RLHF raters historically praised polite, verbose rejections, models learned a distorted prior: they began to conflate benign "uncertainty" with active "danger" [cite: 41]. 

Instead of engaging in nuanced safety reasoning—warning a user, asking clarifying questions, or bounding the context—models developed a rigid "shortcut" behavior. They learned to generate generic refusal prefixes like "I cannot fulfill this request" or "I apologize, but..." at the very first tokens of generation [cite: 42]. This is known as over-alignment or performative alignment [cite: 41]. 

Researchers have found that this safety layer is brittle. If an attacker uses a "jailbreak" prompt that successfully forces the aligned model to start its sentence with an affirmative phrase like "Absolutely, here is the answer," the superficial safety mechanism is bypassed entirely [cite: 42, 43]. Because the model's internal representation classifies the *beginning* of the sentence as safe, it will seamlessly proceed to generate the harmful content hidden deep within its pre-training weights [cite: 42, 43]. 

To combat this, newer paradigms like Dual-Objective Optimization for Refusal (DOOR) are being developed. These frameworks attempt to disentangle refusal training from the targeted unlearning of the harmful knowledge itself, ensuring the model's safety runs deeper than its first few syllables and preventing it from continuing a generation even if an attacker forces a partial unsafe output [cite: 43]. Furthermore, research has shown that models remain vulnerable to "backdoor" training, where malicious actors can poison the fine-tuning data so the model behaves perfectly safely under normal conditions but acts maliciously when triggered by a specific, innocuous keyword [cite: 42].

### Western Bias in a Global Landscape

A more insidious, structural challenge lies in the geographical and cultural composition of the pre-training data. Because the vast majority of high-quality internet data is written in English and originates from Western, Educated, Industrialized, Rich, and Democratic (WEIRD) nations, large language models inherently encode a profound Western cultural bias in their parameter weights [cite: 44, 45, 46]. 

This intrinsic bias persists heavily even when models are explicitly prompted in different languages, or when they are specifically fine-tuned on non-Western datasets [cite: 47, 48]. For example, in tests utilizing the CAMeL (Cultural Appropriateness Measure Set for LMs) benchmark, researchers found that Arabic-tuned models prompted to name a popular cultural dish or a woman's name frequently bypassed Middle Eastern contexts entirely. Instead, they defaulted to Western entities, suggesting "ravioli" for food, "whiskey" for drinks, or "Roseanne" for a name [cite: 48]. 

Extrinsic biases are further compounded during the RLHF phase. Human annotators hired to rank model responses inevitably apply their own cultural and moral frameworks to the data, further steering the model's behavioral alignment toward the norms of the annotators' specific demographics [cite: 44].

Evaluations against global standards, such as the World Values Survey (which covers over 100 nations), show that models consistently misrepresent the moral, political, and cultural values of non-Western demographics [cite: 44, 46]. They default to secular, Western ideals even when specifically instructed via a system prompt to roleplay as a citizen of another country [cite: 44]. Correcting this deeply ingrained bias will require a massive restructuring of the AI development pipeline, moving beyond simple prompt engineering to actively diversifying the linguistic distribution of the foundational pre-training corpora and vastly expanding the cultural diversity of the human labelers hired for preference alignment [cite: 47].

## Bottom line

The development of a large language model is a highly orchestrated three-stage progression. It begins with massive, unguided pre-training on trillions of tokens to build a statistical knowledge base, moves through supervised fine-tuning (increasingly driven by synthetic data) to shape the model into a functional assistant, and concludes with preference optimization (like RLHF or DPO) to align it with human safety standards. While architectural innovations in 2025 and 2026—such as Mixture-of-Experts and reasoning-based post-training—have drastically reduced the cost and compute required to achieve frontier-level intelligence, fundamental challenges persist. The artificial intelligence industry has yet to resolve the brittleness of superficial safety mechanisms, the threat of model collapse from synthetic data, or the deep-seated cultural biases inherently encoded by Western-dominated training data. 

## Sources
1. [DeepSeek and Qwen: Two Chinese AI models...](https://medium.com/sentient-signals/deepseek-and-qwen-two-chinese-ai-models-that-are-pushing-openai-to-release-o3-probably-tomorrow-89e658e9600f)
2. [Qwen vs DeepSeek](https://thefreedomofwork.com/qwen-vs-deepseek/)
3. [How DeepSeek and Qwen are Reshaping...](https://gracker.ai/blog/how-deepseek-and-qwen-are-reshaping-content-research-writing-and-search-engines)
4. [Open Source AI Models in 2025 and Beyond](https://www.softwareseni.com/open-source-ai-models-in-2025-and-beyond-what-deepseek-qwen-and-the-new-wave-mean-for-enterprise-strategy/)
5. [Qwen AI vs DeepSeek (2026)](https://www.youtube.com/watch?v=TSQyxIYMp4E)
6. [Why Won't My AI Answer That?](https://medium.com/@bbloggsbott/why-wont-my-ai-answer-that-a-look-into-llm-safety-and-moderation-407827a6813d)
7. [7 RLHF Prompt Pitfalls](https://medium.com/@hadiyolworld007/7-rlhf-prompt-pitfalls-that-teach-refusal-instead-of-safety-ddd5a9abbf39)
8. [RLHF Limitations for AI Safety](https://blog.bluedot.org/p/rlhf-limitations-for-ai-safety)
9. [PMC Article on RLHF](https://pmc.ncbi.nlm.nih.gov/articles/PMC12137480/)
10. [Base LLMs Refuse Too](https://www.alignmentforum.org/posts/YWo2cKJgL7Lg8xWjj/base-llms-refuse-too)
11. [GPT-4: How Does It Work](https://www.speechmatics.com/company/articles-and-news/gpt-4-how-does-it-work)
12. [GPT-4 Wikipedia](https://en.wikipedia.org/wiki/GPT-4)
13. [What GPT-4 Brings to the AI Table](https://towardsdatascience.com/what-gpt-4-brings-to-the-ai-table-74e392a32ac3/)
14. [Fine-Tuning OpenAI's GPT-4](https://www.datacamp.com/tutorial/fine-tuning-openais-gpt-4-step-by-step-guide)
15. [The GPT-4 Fine-Tuning Process](https://chatgen.ai/blog/the-gpt-4-fine-tuning-process-a-comprehensive-guide-with-practical-examples/)
16. [Meta-Llama-3-8B-instruct](https://huggingface.co/Alignment-Lab-AI/Meta-Llama-3-8B-instruct-hf)
17. [Supervised Fine-Tuning Llama 3](https://huggingface.co/blog/mlabonne/sft-llama3)
18. [Deep Dive into Llama 3](https://medium.com/@zhao_xu/deep-dive-into-llama-3-351c7b4e7aa5)
19. [Finetuning Llama 3 with ORPO](https://www.analyticsvidhya.com/blog/2024/05/finetuning-llama-3-with-odds-ratio-preference-optimization/)
20. [Scaling Llama 3 Training](https://aisystemcodesign.github.io/papers/Llama3-ISCA25.pdf)
21. [ArXiv: Llama 3 Post-training](https://arxiv.org/html/2511.04234v1)
22. [ArXiv: Continual Pre-Training](https://arxiv.org/html/2409.06624v1)
23. [New Llama Scaling Laws](https://www.reddit.com/r/MachineLearning/comments/1eq95ga/d_new_llama_scaling_laws/)
24. [Paper Review: Llama 3 Herd](https://fluentnumbers.medium.com/paper-review-llama-3-herd-of-models-4c51cd0c5220)
25. [Efficient Pre-training of Llama 3](https://aws.amazon.com/blogs/machine-learning/efficient-pre-training-of-llama-3-like-model-architectures-using-torchtitan-on-amazon-sagemaker/)
26. [ArXiv: Post-Training Enhancements](https://arxiv.org/html/2312.07413v1)
27. [Introducing GPT-4.5](https://openai.com/index/introducing-gpt-4-5/)
28. [Why GPT-5 Used Less Training Compute](https://epoch.ai/gradient-updates/why-gpt5-used-less-training-compute-than-gpt45-but-gpt6-probably-wont)
29. [Pre-Training Isn't Dead](https://tmychow.substack.com/p/pre-training-isnt-dead-its-just-resting)
30. [Where Have the Really Big AI Models Gone?](https://www.transformernews.ai/p/where-have-the-really-big-ai-models-scaling-pretraining-inference-gpt-compute)
31. [RLHF vs DPO](https://llmmodels.org/blog/rlhf-vs-dpo-comparing-llm-feedback-methods/)
32. [RLHF and DPO for Finetuning](https://datasciencedojo.com/blog/rlhf-and-dpo-for-finetuning-llms/)
33. [Why Preference Optimization Matters](https://www.digitaldividedata.com/blog/why-human-preference-optimization-rlhf-dpo-still-matters)
34. [Patent Analysis: RLHF vs DPO](https://www.patsnap.com/resources/blog/articles/rlhf-vs-dpo-in-llm-fine-tuning-60-patent-analysis-2/)
35. [Mastering Agentic Techniques](https://developer.nvidia.com/blog/mastering-agentic-techniques-ai-agent-customization/)
36. [Synthetic Data: The Perception Divide](https://www.dbreunig.com/2024/12/18/synthetic-data-the-growing-ai-perception-divide.html)
37. [Selecting and Preparing Training Data](https://www.rohan-paul.com/p/selecting-and-preparing-training)
38. [Synthetic Data for LLM Training](https://medium.com/foundation-models-deep-dive/synthetic-data-for-llm-training-4c5b70371e04)
39. [Synthetic Data Fine-Tuning](https://futureagi.com/blog/synthetic-data-fine-tuning-llms/)
40. [ArXiv: Synthetic Data Generation](https://arxiv.org/html/2503.14023v2)
41. [The Shift from RLHF to DPO](https://medium.com/@nishthakukreti.01/the-shift-from-rlhf-to-dpo-for-llm-alignment-fine-tuning-large-language-models-631f854de301)
42. [Defined.ai RLHF vs DPO](https://defined.ai/llm-fine-tuning/rlhf-dpo)
43. [Direct Preference Optimization Simplified](https://keymakr.com/blog/direct-benefit-optimization-dpo-simplified-rlhf-for-llm-alignment/)
44. [ArXiv: RLHF Methods Comparison](https://arxiv.org/html/2404.10719v2)
45. [DPO vs RLHF](https://medium.com/@AdithyaGiridharan/dpo-vs-rlhf-two-paths-to-aligning-language-models-with-human-preferences-af25869830f8)
46. [Generative AI Handbook: Training LLMs](https://www.aihandbook.io/generative-ai-handbook/training-llms/)
47. [LLM Training Methodologies 2025](https://klizos.com/llm-training-methodologies-in-2025/)
48. [Pretraining: Breaking Down the Pipeline](https://mlops.community/blog/pretraining-breaking-down-the-modern-llm-training-pipeline)
49. [Pre-Training and Fine-Tuning](https://notes.kodekloud.com/docs/Introduction-to-OpenAI/Introduction-to-AI/The-Role-of-Pre-Training-and-Fine-Tuning-in-LLMs/page)
50. [Understanding the LLM Lifecycle](https://medium.com/@aditi.sikarwar25/understanding-the-llm-lifecycle-pre-training-fine-tuning-and-instruction-tuning-0fd0826f0f8a)
51. [LLMs Exhibit Western Bias](https://venturebeat.com/ai/large-language-models-exhibit-significant-western-cultural-bias-study-finds)
52. [Georgia Tech: Western Bias](https://www.cc.gatech.edu/news/llms-generate-western-bias-even-when-trained-non-western-languages)
53. [PNAS: LLM Cultural Bias](https://academic.oup.com/pnasnexus/article/3/9/pgae346/7756548)
54. [ArXiv: Recommender Systems Bias](https://arxiv.org/html/2508.20401v2)
55. [ArXiv: Intrinsic and Extrinsic Bias](https://arxiv.org/html/2411.10915v2)
56. [LLM Evolution 2024-2026](https://www.llm-evolution.com/)
57. [Generative AI Models Face Off](https://medium.com/@consult.edcults/generative-ai-models-face-off-analyzing-gpt-4-llama-3-1-and-claude-3-5-8c03ac7f0593)
58. [Open Source Tracker](https://gist.github.com/BIGBALLON/5b5a4d5a053c7e73484fabea0c0e2466)
59. [GPT-4 Technical Report (PDF)](https://cdn.openai.com/papers/gpt-4.pdf)
60. [ArXiv: GPT-4 Technical Report](https://arxiv.org/html/2303.08774v6)
61. [Democratization via Synthetic Data](https://medium.com/foundation-models-deep-dive/synthetic-data-for-llm-training-4c5b70371e04)
62. [Data Diversity and Performance](https://www.rohan-paul.com/p/selecting-and-preparing-training)
63. [The Rise of Synthetic Data](https://www.cloverinfotech.com/the-rise-of-synthetic-ai-data-the-fuel-behind-safer-smarter-ai-in-2025/)
64. [Synthetic Data for Fine-Tuning](https://futureagi.com/blog/synthetic-data-fine-tuning-llms/)
65. [ArXiv: Synthetic Data Diversity](https://arxiv.org/html/2511.01490v1)
66. [VentureBeat: Cultural Fairness](https://venturebeat.com/ai/large-language-models-exhibit-significant-western-cultural-bias-study-finds)
67. [ArXiv: Bias in LLMs](https://arxiv.org/html/2411.10915v1)
68. [Georgia Tech: Arabic Prompt Bias](https://www.cc.gatech.edu/news/llms-generate-western-bias-even-when-trained-non-western-languages)
69. [PNAS: Evaluation Across 107 Countries](https://academic.oup.com/pnasnexus/article/3/9/pgae346/7756548)
70. [ArXiv: WEIRD Bias and World Values Survey](https://arxiv.org/html/2508.19269v1)
71. [Generative AI Handbook](https://www.aihandbook.io/generative-ai-handbook/training-llms/)
72. [LLM Methodology 2025](https://klizos.com/llm-training-methodologies-in-2025/)
73. [MLOps: Breaking Down Pretraining](https://mlops.community/blog/pretraining-breaking-down-the-modern-llm-training-pipeline)
74. [Role of Pre-Training](https://notes.kodekloud.com/docs/Introduction-to-OpenAI/Introduction-to-AI/The-Role-of-Pre-Training-and-Fine-Tuning-in-LLMs/page)
75. [Understanding the LLM Lifecycle](https://medium.com/@aditi.sikarwar25/understanding-the-llm-lifecycle-pre-training-fine-tuning-and-instruction-tuning-0fd0826f0f8a)
76. [Shift from RLHF to DPO](https://medium.com/@nishthakukreti.01/the-shift-from-rlhf-to-dpo-for-llm-alignment-fine-tuning-large-language-models-631f854de301)
77. [Defined.ai: DPO Benefits](https://defined.ai/llm-fine-tuning/rlhf-dpo)
78. [Keymakr: DPO Simplified](https://keymakr.com/blog/direct-benefit-optimization-dpo-simplified-rlhf-for-llm-alignment/)
79. [ArXiv: Reward-based vs Reward-free](https://arxiv.org/html/2404.10719v2)
80. [DPO vs RLHF Tradeoffs](https://medium.com/@AdithyaGiridharan/dpo-vs-rlhf-two-paths-to-aligning-language-models-with-human-preferences-af25869830f8)
81. [Context Window Evolution](https://www.llm-evolution.com/)
82. [Devopedia: Llama Architecture](https://devopedia.org/llama-llm)
83. [GeeksforGeeks: Large Language Models](https://www.geeksforgeeks.org/artificial-intelligence/large-language-model-llm/)
84. [Generative AI Benchmarks](https://medium.com/@consult.edcults/generative-ai-models-face-off-analyzing-gpt-4-llama-3-1-and-claude-3-5-8c03ac7f0593)
85. [GPT-4 System Card](https://cdn.openai.com/papers/gpt-4.pdf)
86. [Hugging Face: Meta Llama 3](https://huggingface.co/Alignment-Lab-AI/Meta-Llama-3-8B-instruct-hf)
87. [Unsloth SFT Optimization](https://huggingface.co/blog/mlabonne/sft-llama3)
88. [Deep Dive into Llama 3 Math](https://medium.com/@zhao_xu/deep-dive-into-llama-3-351c7b4e7aa5)
89. [ORPO Finetuning Llama 3](https://www.analyticsvidhya.com/blog/2024/05/finetuning-llama-3-with-odds-ratio-preference-optimization/)
90. [Parallelism in Llama 3](https://aisystemcodesign.github.io/papers/Llama3-ISCA25.pdf)
91. [Speechmatics: GPT-4 Architecture](https://www.speechmatics.com/company/articles-and-news/gpt-4-how-does-it-work)
92. [GPT-4 Bias Training](https://en.wikipedia.org/wiki/GPT-4)
93. [GPT-4 NLP Capabilities](https://towardsdatascience.com/what-gpt-4-brings-to-the-ai-table-74e392a32ac3/)
94. [DataCamp: Finetuning GPT-4](https://www.datacamp.com/tutorial/fine-tuning-openais-gpt-4-step-by-step-guide)
95. [GPT-4 Domain Finetuning](https://chatgen.ai/blog/the-gpt-4-fine-tuning-process-a-comprehensive-guide-with-practical-examples/)
96. [Superficial Safety Alignment](https://medium.com/data-science-collective/current-approaches-to-llm-safety-alignment-remain-largely-superficial-8fa2aee1ada8)
97. [RLHF Pitfalls](https://medium.com/@hadiyolworld007/7-rlhf-prompt-pitfalls-that-teach-refusal-instead-of-safety-ddd5a9abbf39)
98. [Pristren: LLM Alignment](https://pristren.com/blog/llm-safety-alignment-explained/)
99. [Bluedot: RLHF Flaws](https://blog.bluedot.org/p/rlhf-limitations-for-ai-safety)
100. [ArXiv: DOOR Framework](https://arxiv.org/html/2503.03710v3)
101. [Deepseek and Code Generation](https://thefreedomofwork.com/qwen-vs-deepseek/)
102. [Qwen Multimodal Powerhouse](https://gracker.ai/blog/how-deepseek-and-qwen-are-reshaping-content-research-writing-and-search-engines)
103. [DeepSeek NVIDIA Disruption](https://www.softwareseni.com/open-source-ai-models-in-2025-and-beyond-what-deepseek-qwen-and-the-new-wave-mean-for-enterprise-strategy/)
104. [Qwen AI Multilingual Workflow](https://www.youtube.com/watch?v=TSQyxIYMp4E)
105. [Bruegel: DeepSeek EU Impact](https://www.bruegel.org/policy-brief/how-deepseek-has-changed-artificial-intelligence-and-what-it-means-europe)

**Sources:**
1. [aihandbook.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEMzNIgOq9ny8nimYO9ZlXJywbmugELtTbo-4S2ryQbVVqrUG4KBZfd12IQOSDIwiznokOEeuccamXKNY0f_ZKXhZIS6vMqzERaUHWOQ7aMIhJQPTUdP7l2Vhk1N0fVSfrubi6x-lHK9WMZ2-nGFhvUwoKgAw==)
2. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGv0I0tXt1-xIId3iXLvkEdw_8u4udW5701xt1FViUDC15WUHcf8AMZJH5xDAZCUR67CEoQjSTN_ANR2lPEAjV2wBuoZY3WmVFXOQ7RPzR7f6H31mzY-b-1mRTFgfxxgipaAwyk0Zazltb8kTnVoGPjdHyQSOB2gqFZEMuKclYeN2vfzUkc9LOGZKf_XSgzSXNzgcG6vo7hsPV3UbusewLZCchhiZt4v1Edxd18UkiKyMiDk9SNql4=)
3. [mlops.community](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEL7gYBGqa1KE_KEANStjHV_pbkujVObYOMNEECgpURn39jM34eXEtEbMxX7B3mdWp97_nBUiyhKAxjy4Qf8OL-vfMHjXa_QCyCLAqBF6I9ekM2W-ImMMhR7OArG8OKZCo_PDaYuqdRAMn20GqQyxhELGODsLNg9jcJOlbygDSEKrqhOBI4btEHycas_g==)
4. [gracker.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHSkv0W7MH9HBUnGx9CI6Z_ATSf8aX_nHqmCOeJSqn_TfkFIEi9fOXE6vr4LZdIaxRU5Ts7oZzOf0QIkLF9PrJDm4OgjmrdKu1D71FNaC5Egj1OsUliZlYh7JI85MegYi4lBVmbap1U0xpwB4PcqYHThNGQgOE17lVhRWQ8xZTaGYCrjQfdNfFoDWBDfnDTP02n7ei6yMlrCc-02jk=)
5. [softwareseni.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFvMyC4NfkQp5RUcXRpzlEfMGhMu_97BKoOHdUFkGgxKUEjCM8zfZ_N4mf3EiRIfsHISX6cN7gOabU-0bjSkF8Ijt19o3sih9gkXqNUSOSmfKlNk9MB5Q3LxIETNDu2UZicXMdtkWNbdo_BS4XoNqAPwGcV_3E-rI7e5fVisurwCZF494HYD1sVDPtfxad26-KMvUy4ECgJcRtQBrW-keXvfz-rmQcvlyQigPCyJl6lFuVGCvqD68nz-xpimao=)
6. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEPncmoum014DGI5UH35r7xPyhiKfUtheqLyb6l-njJtIyRxlKwjLtNUEBEvJtCcXpA2QfpTSyoqgiuSQGEW3hZ2807QLdm80UXx6iKvkykdPX2b4xD5SyJeWh3dt0W_HQgz5Nln8OPG8UN__A3-Ir5C13KdTwF9xWfu93-gZ39BKs4GgaUuIgNcNyNu0_eirY=)
7. [kodekloud.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG2Jn9TpomohngyCicwsAXHsFFst01YVJLPgfA1eUyJbo12ac9xw1uMrIbr4UY90vVFmUU0ciQxNlLTGrQPSa_eEQ8Y0rZy1pcxgMMZ6UpTAgXUcfJKW9b-PbnP0tYTvvKp8r4srww4RmR8hf3fO6dB5Jt8yFTlnueL03k1EFzK1Qru8fEQ1oCkePZ6uZf3IffYRGr4CE8QiimjH6JxQ9_F7L2gzkDE8RU_1tRfND4dbkO76S0SEw==)
8. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHQMd35IcGjyo1i3L7Nj_lA5d0K5x0qUaq0jkUA7udeDEkQDFQyRk-MAgL3X_L_WGY6xyKjIKZpFyui9pOQxZMChZkGnxn3dq7s5t_XYGTTka76BXLgC87cUqwsl5ALI0oC0nEEW4-EnK7yUkkrxsOmsFyKog==)
9. [geeksforgeeks.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEpqvsTPC1KHzcjr75LWprFKVhmiEJDRywlafZBlDRTq7_SlW6TikQVXXeOqkDRnFqf_qn7PwWCcmFnDWcZ7vbo9qZGCWSV2w9cOef1TaaQP_BZbxGPKz7yEUJnCpLzkUy88X1soim3WA6cEVEl8SeKJLhhCzucRUoKuUq5e_zNtKbDLtc=)
10. [speechmatics.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF5oKiLZMfedb8O4wCYBj_zRlaK9lQFpWnwcophYkdLjszF3MM3T21cDGl2fB5u2dfARKK9wWys1U4Qigsvg8AglazY0WvzyOVvqheSMg3reE9PHGcRddClJs6OL0k-6udyQFwCJsWKoxNN4ctzDfbpHKpZhFwnt2rBA5ry-RZaTjsR)
11. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGEdjMS1Wfc-dGj30sAb2Jsxb24ovWzHFtnGvyh6bYEZAobzbKbq4STxV0aPlfRRDx6EJ23IjY9I-mFk6n287__EVw4MuCuM7JPDkG6Hk-3rQqAquC9K6F7rbVt4Un_vTyfzKr0OKKBHNGi7uSaRyAsGQbyIpkJw0M=)
12. [reddit.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGE23sMDH5KeBkDcBKobc7Z6Y8ZIYZM2MQk2i7UF4IkxMP7hxqgPfOQNvgjyO2PvxqsgmSmsDhzmGRd69GKtUUMH2Y5KeRTAlcl00yksQ8kiSuL8bcNHdRosLBlsY92fDRQlazNfrihBLesDLVsPtVH1UUvZbQI3UoaZM6bIluIT3a6TWxiJHgH)
13. [rohan-paul.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEtwx_XqGwSOWSssgQPpKHPEqYcYDKmx1XHlQccJ15dy9NLKkwpt3tJr10TJDAw-L94r6LensoTLtlPCWwGLGWzoqVsoW00gKOADRRpHd3gvFuG5k1EjPwiepLJ-Bg7QAaaIrQW-148X6YplzHITNtIAto=)
14. [devopedia.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGngEBnsw1WTg4z3qt0o-S-krYtZ87Bi_0Ox49zi6YHArS9Q1l951G7b5UwA50MlomhoOOjIHApmaGVuZdDx2kQdyJX6pp685T_i9Egzi2kr2YxMFQ=)
15. [github.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEaZBSEUnTmBO1nU_fg1FyHsL6JephiS63gz8asJGtk5U4423nRSa-5_BjlO_eXRLyA3uz5dauV-hAqv864TPccwjQdrIPMfi3IghIEfwkdih3rDAIMvmoh0zZDJR-UbRlOVjeJi9tojcwF9x1vN5A9)
16. [llm-evolution.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHyDxk2WRMJRiUsQ0G1AywWg83VKag7kGznQDfv11RCSA2tvm-iUYuBIN_slkuzCvsI2tEJcpO1rGZg6ejHEP2ItO5eEOkLvyCLnjsOHJ0wC5sxHA==)
17. [amazon.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGVhFPps1QRzD035vnZGNLohP1mLCAMsBi0cviJvW_sa-uq9nyQexxHoA3fQlM6ocElg6bA0Jv8sfDswNzysmrw1_2f8AI8yyD0msg9CW_6aO9iv7l0WJv4ccOnZVIAnASWx6PaQB_ITdMaiW0RR6DuZilQvGKDajqm6VqO1LRepF2XdoKtMtxSb81Bt8ACcULPMOENWSU_svbsyTdziq2IJyU5IBWmA3R_GtMjfmvaUzjphWNnzIXQy0Ne-z1vq6qt6djy)
18. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFWybw0gYKkncQUTEI6SASeWW563Ia40jWXZN4wAlofdlqGUFfpIju8B8x4YTgR1hpXWoicyIb-2llEP2SQrHyQcEOXXzlqP54QsD77J3fGUrdtsbVkhDtjW_TJgdL3rCvO-3RIPfFKF9JotXHunrWPDBpKdxlVWpJzt6hTc5UMn705qYU63g==)
19. [klizos.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFej3xFazaRsguNCIE8d1GNcOfvr0ZCtcQU_bzq7rPwIQG0dRk14fXZV_-jBNaRk3f-XBTH5BMY1vyu-uN1PSVRLgF1Bngzv93_QOPJDfU2AbMIOZ6hfaRr8_hduj66tcDhDSaZd0nD-ifDOA==)
20. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEVPkAa8DV6ywFp2WaK2tGg0KHYqZse7R_gKUndfhtyQ2N1uTf0IJmnGNnmRa0Ojd0kqBHFJRYzWMWc4SZ7gr8aWF2VAS2H052iJVYTYkneAg6FfL6oAWmIMWXt7wt1zzaHtCY6PiAPG4aZY59gHnktkGL0wu6LfMpLARYq39ktK-NLEIg1pbQpQm_jxjl7gD30xprPq4x7g2ZZFRtp1-ivMC_0gMOH_jpP4VC3Q-1C_C1PAIeynOc=)
21. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG7IRAaeNQwh1igxaTBleC0tSDKn9fEuNslczDDGujksKU8WOHWeUVZ0iGDN-6vfEGS75IEpTEQnpO0F9_6ohTJgsjWKSon1OMJ82k-xB2eaejBR1_tQ0e6D35kriXCKM78nJq6JC8YzmLeeAKNl-qPiSJmP9vZ6WzujCRY6PGLyB4k2Sgafcv1-9KcxqIrLir2dweCUJHZ_RjcpU6evT6X0B68FQMk073m_OeX-Fphyg==)
22. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFK9rPrG6Mb196kfgl7bLDv00hmae0_trmeu9IObRxkyrRxmEunhR0m9kgmDf9hhCI72CrvUSifz8_9S64U53zFCKVJNpIWLOYVr3Uk8Nr51lTunmBdvmL8EcHtIefmXU3a2ljNNg9x)
23. [pristren.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQERVbKlB2ct3WqY_v9cFS0yQCdti6W7sLbw1wijI56RUnLHGXz6--opPjKUKZlZf-PQ8kxbQk7UVCXC_-X0TgEag4Tt5bpQTmNQkskrZfuShcrV6hjgqgIoyIF6MFpiWY_LUkJ5ScPn95YZWpBjNg==)
24. [huggingface.co](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGaDGG6Or7qfZ_gg53uz3uwSxMYmnsJVsSbrejIvkl9AdQn3ADTqoptTX1soG_umjYHaBr1j9Rb-ILJPeyQ3nFLbJXV900srnFGGUdGmZfWUp3idavwzcdrioTZ8BPzDdmR45S1)
25. [nvidia.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHlVOnCfkPW18lFA7LTiYS6nRmHdkyf1azG3LRWcDWPkaIxKaOa3shKbP8sLk1Zt0qccsi2cewpUFWrLpEPrAusTUpIUQfyeTW8o7zut9ELGTbD7yd7TPNNLLOh_wk34_NfJBcOj7OFNdB0YtXbhF9YbsdpzpTmyCSMrpb98PPoeEhCoB0BOY_nK_2f)
26. [futureagi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFCX0ZpW6EVthePR09X9hxl-UCZ1RnAyU1vA1CR8sq7UKMjoVu7LtNNqm1VX8JakBaKACUSqO1rEVRZutibEQYDKVHvwju-b8WlxEEvLUrFwg4cNaRcGyWPRS5HERqpXb7jpxhZvB9BeDZ69ukPfKqq)
27. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHDpyTxaB16ZkXyQ-Ea62EZIW0zLwcYb7TwtE2IlOQNyDWXcL_jtBAq1XcDbRoRp9TY-CXufGjhdJYMEIChQ-i2yvEaxDFg5exa9970LxtCyrSBRZ6NwIbw)
28. [llmmodels.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFIFgQRApA5tCbqDhPn52aXLkrIRBcVDsaMD5G0sQ1v07RvXORkH6zk3rcb4z_vVM86LUBtS_5pDmg0IkKNWZXldKl-3YYysjzjJYf-WvK0vVztp4X00l7GlzyPr_zzYSP49qnLAVC5989vVSpZu09_spztJkagYOssSmw=)
29. [datasciencedojo.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGzUL2pt6IJpna7VTqBm-XwocK_Cp1KkmYTaM9M0DsDRJEVpcsYtoEcMeBZlx9ibhQh1EItiyLtLfgGOlUkbpQcdJ0DZAMeaW9K-WRoeqGmeRdp4CxHc8RpOdXNw2efkb0VkGdc8r2X3kKlcUxI7W2mgh0aTAQ1JA==)
30. [patsnap.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHBih1twYwpM52aHicM-JqLunR08Lsp8IVcxkfYUfBxrk-bnxzWnE1Xl5oRSl-_9IEzj8cQffzz4-OJq6_eXATUMpWWg7akg4GBn5jn0r0effffqtIkVYjJryeudIUmkXyd4IaR9Ipf9KLISlviIr_6zOoPbX2IA16si5Eo6bQmMbhyf7NAnKoFntKjIq6PnXEPYvlxBVWOsUY=)
31. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFp2WnKDw8IZYDUd4ScCQ-xgMKVH88dztRPpvCEsDtL7c99w732FB9SaPrFrRaDs86fDaAt8FKtkEVN3CyFK3BWCmclKJZzw-4iazdaOUS4ZCkswiXCPGqR)
32. [keymakr.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHUNQJGkk1VxNof1h0EF6h2evnGAeVy8uw2lR_0-MOpL6geU9zy_Sr2T5NTXAQZw9xqEHJgaGhaOcLWVqeN4KHhSAwUB8a7QoVN5hGuTPPD4C6G2BCQ7qtnmRQIAqejklwDozW938ZQn6xM0CSSZophwggZZyYW2zlDnuSOr6OS4w-5s_3AHulg_9enrDHgb4c=)
33. [digitaldividedata.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGRtq6T7pvvPSxM2_TKrldw9MI85o46gjzBUOt59aJ1ctjIEECZ2HSvojWD30R_nfjGC_YES9ak31tsN5uAT8GNTDNoeaIqFT3_d5-qT8mJc6c_EPQ0ez-XSw0MibkLSXIXgbPd9KyFU9oECM05okqLzpOBbJp2vp7EJSKUhudAG3FhIsXl235jbNd2LqXaNCVhR4Dz)
34. [analyticsvidhya.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFKB78XFO-DOe0X2jKLNWR8v3Ap0qsAjLNt7hwlIKLRZKxP13LLvjvUOgFN5IX-xaJwPz37rXa5vxGNrcWSMBuy6ejhEqKdJf7jD58bbHxUsf5wMm1fCoEc_7_76MH3CmAIgDXL8ifZVLQT92L0qq5x2qYsFTIKymcsy1eDptasvQWhaXAW2_BsexCsbu2A43fovlaTiChtSYqJCZMU)
35. [bruegel.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGaJi9KsDz9-X8lzgL3NPpRQ2s2d7en0VgGe84EQAHGf49pRtlMm30eeYZw-cIxehUe9AQq5pUL8RHITbdttCS6fiYPpVXcXLcZlGBTsqeS46-TiWIQbeDFjiyfb4msZkRFOzDTSgS8Lp_fbhLDsPglbANCAYSohP4cILdr8z-N7NNRBjP6k71qLYk7sLqOGUjyhJYUD2NJ4kFhm9EDdnWOVpNh)
36. [epoch.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHmOrciJ4TzhuzpFwG92M3Pj5Ovm53Yqqmck3XySP6vY0p1KeVFAfTGj1nDNnjsmxZuP-WOYersOb6bpfMwk1wkyEyZl3peKez95MKcgDujsJnLLbzDEsyIoewYBV3L4bipXyVlD9hbW3EvshuyEyhG9LE6pBLOUB1ej4IaQTjxiQ5rcbSnUgvomLEhnVy6QFMVNhZKB2OeAcvYUuk=)
37. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHnNDiK745WIASGPpdMqOAXngSRyQtOClbQFt29GvB7nnMWM3CetzDCbwoYt82kAN1Aw5b80JkVgh7dRorDiH7PEe2yZoKUX47a4821IHiLgbU9_4iGSVqfeDKa02kKrTWSZ9-BHJyqGbTV9Sd7C8qPuBDxBKA9tTzILf8=)
38. [transformernews.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFeK58pffSydlBNSy54YNTmo1IkCDSHgHV_z7n41t9afRR53CIrgfTV5KpL_DYaOOFgYwms4A0NP81txoXHkAqBybQS5Wx7tKeyNj8Lz2iEwtXFnxjy20k5CdSCC_u_MJSu7nMCGU30TWcSm2H3tRSmbKpKYDkObTyLwwCu86FILV7Ft8v3gCd0FHzC1JJLY-EeqYWqP6hUAbL4O__2DTbT5Ror)
39. [openai.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFwlGchu4fBqvm4IpPiTARQGFOxvVCVMyvtuYbsj_g0PkAG98jmWcq6ZchjItXrv2agH7gNHtSE4wGXe_CVnfGnMJ_tMP6OcdvYcZZ0ROXPEB4xglmQNXMM1uUlFHDZJk08SA==)
40. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHEUT_3mkZqI2m6xR9DgFIteFnvmHn3SEWYQzsy_4JVPEvB97jQOicxF23dfhL5XlFdc2Fys8RXbZDkCVHWVNOlFt5A46gtno89_uMG-9zrKI8vV8V0aPVSc1qNxOnWDbtA4hbplns7DWmwhGj1XRR-mRNACQtP3KqdiwQV_zV9FLtaE3XTbyhI5SQ-ET8gCqsaUl1S9VA7i0o29xgx1eceTjzsmkWpieEhOSdbabynWLBV5oILkZRCet8vzGJU4n3WL-Fm8q8=)
41. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGuS5MrVvCJJptjH1sa5khjPTY0KyFScuAbGS1Go-pb39hhYgAxe3v6EBxeU8sfW9NbdxmYxhrdK9ETxK1pp_Q8H22yXcL3qsByT0wdxrFQd3d2PyEJi2QbzbjD4_Js92tOjUumycJhQHoCU1l3tWnSLBv7SHvCUgG8wBb_LNEJC7iJtpn4Q9FoEmTVNuO2X_mKyNphIZ286txvO_6Ap5vgYA==)
42. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFf2aXyWDRanGYF5DDLecb8nBVWbALUf49BQVf2mjg-CzglnKUmvXwKd0kN_kBeY8lj4mACx05nNjmYf4Ct7iZZSFzlDtktk9PN3Hj8BKBGGbRz_DGSWa1_qiXxTD83J8AUIaG--r77TjuIXQavWCXjEN-1-xg13H3U19sA0_xTK_eze1x6kJsM1tjmOnO4SBHLVRZ5GLV0TbHcjjzZb6qdkSL0_RUeno7b-EKeUV7RMuUl)
43. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGocbX6rSvnXyGd9odZGJC9j8V3GOh9dXSBe8nUBPqrUI1twaCh3XhMtRgpJMauzgf9vd2ErZNkTGQeT8D7JFMhmvRoIgAAUWXUI43AljO9Zd0Zt-pj-JC2)
44. [oup.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHWyVz2MWgj0WNWUoeAQlY6USJE4OH4gj3sEpkHby89pKWcjozoej_0nhQXlyVQtMwKwIdSaLAmGQfZ1kKm4xkfx6IgBZ1wCUEaEVuCpcBgwtC7xtNOCAV2ZE-QB79YCAN0TsNkB-AC0iTzwzgOWy3xjwaB)
45. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHHi_EfXxveJqVWYU2K4pEpG2IS3gN4rUdIxEGVcHr0Yu5b2j-rgh9DOBLP4G-Xq6yL_pKUh6ZQl7Mi-s44RC7BH_hxNS6LgCWOnwSz3aHsZuYfTnqhxY5R)
46. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHLPM4KEgxsemXXdi9ewETTpkPBEIJUnk_v4QXGvi0UlQlxyOIKvm5JigJT9EFqAF0-25B06hDioRezgdtuwtjAl-b-EXyYAFaOjteQ3hXIOV1O7jMO1Xhd)
47. [venturebeat.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEC9WkE5YOOreqiLm8LBVuom1QMVCmXWxxshuSSsAEpE073tgnsmBK0unicbu32ZCKQQ3cdlpC7uciIBXWD3240NYgESnT7Cl4LJuqK83N-5DYuTon9mjqrKM3c-GKlsxxOukS1WobFo3lch3dvbLAfg3BhrWgpvuXc3YbSn0l9xp50zVTXRKVSvwkyU0bfE5-CQ1PgEdTvuebeHA==)
48. [gatech.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQExYwy_QgylqhY_kPN_FC3nQgxfTN12mxNq2OYE8LA3s_kKGEI2iCeFfpr0rHfNjEDHMCDsR5w0JoAOJgNWFzonsmNnh_Bbha0ghZ9PJpUJoCviiFad7VAw7FpbIJiMtcg9GCHJp4ktxcofsJpJulQmahuV5SkgXrvpeHvHLAJw55bsbNe_EpMNBAbhYD0taqSXHfmZhOE=)