From Transformers to DeepSeek‑R1: The 2017‑2025 Evolution of Large Language Models

1. What Is a Language Model?

A language model (LM) is an AI system that learns patterns from massive text corpora to understand and generate human‑like language, enabling applications such as translation, summarisation, chatbots, and content creation.

1.1 Large Language Models (LLMs)

LLMs are a subset of LMs distinguished by their massive scale—often billions of parameters (e.g., GPT‑3 with 175 B). The term gained prominence after the 2018‑2019 emergence of Transformer‑based models like BERT and GPT‑1, and exploded in usage after GPT‑3’s 2020 release.

1.2 Autoregressive Language Models

Most LLMs operate autoregressively, predicting the next token based on preceding context, which enables coherent text generation.

1.3 Generation Capability

Through iterative token prediction, LLMs can produce complete sentences, paragraphs, or longer passages, supporting creative writing, dialogue agents, and automated support systems.

2. The Transformer Revolution (2017)

Vaswani et al. introduced the Transformer architecture in the seminal paper “Attention Is All You Need,” overcoming the sequential processing limits of RNNs/LSTMs and enabling parallel computation.

2.1 Key Innovations of the Transformer

Self‑Attention: Computes relevance of each token to all others, allowing parallel processing and global context awareness.

Multi‑Head Attention: Multiple attention heads capture diverse aspects of the input.

Feed‑Forward Networks, Layer Normalisation, and Residual Connections: Stabilise training and support deep stacks.

Positional Encoding: Injects order information without sacrificing parallelism.

These innovations made large‑scale language modelling feasible and set the foundation for modern LLMs.

3. The Pre‑Training Transformer Era (2018‑2020)

3.1 BERT – Bidirectional Context (2018)

Google’s BERT (Bidirectional Encoder Representations from Transformers) introduced masked language modelling (MLM) and next‑sentence prediction (NSP), achieving state‑of‑the‑art results on GLUE, SQuAD, and many downstream tasks.

3.2 GPT – Generative Pre‑Training (2018‑2020)

OpenAI’s GPT series leveraged the Transformer decoder for autoregressive generation. GPT‑2 (2019) demonstrated strong zero‑shot abilities, while GPT‑3 (2020) scaled to 175 B parameters, delivering few‑shot and zero‑shot performance across a wide range of tasks.

4. Post‑Training Alignment (2021‑2022)

4.1 Supervised Fine‑Tuning (SFT)

SFT, also known as instruction tuning, trains models on high‑quality input‑output pairs to follow user instructions.

4.2 Reinforcement Learning from Human Feedback (RLHF)

RLHF improves on SFT by training a reward model on human‑ranked outputs and then fine‑tuning the LLM with Proximal Policy Optimisation (PPO). This yields responses that are more helpful, truthful, and harmless.

4.3 ChatGPT (2022)

ChatGPT, built on GPT‑3.5 and InstructGPT, was fine‑tuned on massive dialogue data and RLHF, delivering engaging multi‑turn conversations and sparking the “ChatGPT moment.”

5. Multimodal Models (2023‑2024)

5.1 GPT‑4V – Vision‑Language Integration

GPT‑4V combines GPT‑4’s language abilities with advanced computer vision, enabling image captioning, visual question answering, and cross‑modal reasoning.

5.2 GPT‑4o – Full‑Modality Frontiers

GPT‑4o adds audio and video inputs, supporting transcription, video description, and text‑to‑audio synthesis, expanding AI capabilities in entertainment and design.

6. Open‑Source and Open‑Weight Models (2023‑2024)

Open‑weight LLMs (e.g., Meta’s LLaMA, Mistral 7B) provide publicly available model weights, while open‑source projects (e.g., OPT, BERT) release full code and architecture, fostering community‑driven innovation and efficient fine‑tuning tools such as LoRA and PEFT.

7. Reasoning Models: From System 1 to System 2 (2024‑2025)

7.1 OpenAI o1 – Long‑Chain‑of‑Thought Reasoning

Released in September 2024, o1‑preview introduces internal “long chain‑of‑thought” (CoT) reasoning, allowing the model to decompose problems, critique solutions, and explore alternatives before emitting a concise answer. It excels on math, coding, and scientific benchmarks, often matching expert performance.

7.2 OpenAI o3 – Next‑Generation Reasoning (2025)

OpenAI’s o3 series builds on o1’s architecture, delivering groundbreaking results on ARC‑AGI, Codeforces, and FrontierMath, with performance far surpassing standard LLMs.

8. Cost‑Efficient Inference Models: DeepSeek‑R1 (2025)

8.1 DeepSeek‑V3 (Dec 2024)

DeepSeek‑V3 is a MoE‑based LLM with up to 671 B parameters (370 B active) that reduces memory via Multi‑Head Latent Attention, DeepSeekMoE routing, and Multi‑Token Prediction. It offers comparable quality to top‑tier closed‑source models at roughly 1/30 of the cost.

8.2 DeepSeek‑R1‑Zero and DeepSeek‑R1 (Jan 2025)

DeepSeek‑R1‑Zero removes the SFT stage and uses rule‑based RL (Group Relative Policy Optimisation) to align directly from the DeepSeek‑V3‑Base checkpoint. DeepSeek‑R1 adds a small curated dataset and additional RL phases to improve readability and alignment, achieving competitive scores on mathematics, coding, common‑sense, and writing benchmarks while costing 20‑50× less than comparable U.S. models.

Conclusion

The evolution from the 2017 Transformer to the 2025 DeepSeek‑R1 illustrates four pivotal milestones: the Transformer foundation, the scaling breakthrough of GPT‑3, the democratising impact of ChatGPT, and the cost‑efficient, open‑weight era ushered in by DeepSeek‑R1. Together, these advances have transformed LLMs into versatile, multimodal reasoning systems that are increasingly accessible to both researchers and industry practitioners.