Paper: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020)

Let’s stop guessing. Let’s search.

LLMs hallucinate. They don't know your private data.
In this episode, we dive into RAG, the architecture that changed how modern AI systems handle knowledge.

Instead of relying solely on parametric memory (weights), models can now retrieve external documents to produce factual answers.

We break down the 2020 seminal paper and explain:
🔹 Why Large Language Models hallucinate.
🔹 The "Retriever-Generator" architecture.
🔹 Why retrieval has become the backbone of Enterprise AI.

Clear intuition, real examples, and practical insight.

🎧 Listen now to master the tech behind ChatPDF and Search.

As we close the year, tell us what you think in the Q&A!

In this episode of AI Papers Explained, we explore Mamba: Linear-Time Sequence Modeling with Selective State Spaces, a 2023 paper by Albert Gu and Tri Dao that rethinks how AI handles long sequences.Unlike Transformers, which compare every token to every other, Mamba processes information linearly and selectively, remembering only what matters.

This marks a shift toward faster, more efficient architectures, a possible glimpse into the post-Transformer era.

When AI Learned to See:

In this fourth episode of AI Papers Explained, we explore Learning Transferable Visual Models From Natural Language Supervision — the 2021 OpenAI paper that introduced CLIP.After Transformers, BERT, and GPT-3 reshaped how AI understands language, CLIP marked the moment when AI began to see through words.By training on 400 million image-text pairs, CLIP learned to connect vision and language without manual labels.
This breakthrough opened the multimodal era-leading to DALL·E, GPT-4V, and Gemini.

Discover how contrastive learning turned internet captions into visual intelligence.

In this third episode of AI Papers Explained, we explore GPT-3: Language Models are Few-Shot Learners, the landmark paper by OpenAI (2020).
Discover how scaling up model size and training data led to new emergent capabilities and marked the beginning of the large language model era.
We connect this milestone to the foundations laid by Attention Is All You Need and BERT, showing how GPT-3 transformed research into the age of general-purpose AI.

🎙️ Source: Brown et al., OpenAI, 2020 — Language Models are Few-Shot Learners (arXiv:2005.14165)

In this second episode of AI Papers Explained, we explore BERT, the model that taught Transformers to truly understand human language.

Building upon the foundation laid by Attention Is All You Need, BERT introduced two key innovations:

• Bidirectional Attention, allowing context comprehension from both directions.

• Masked Language Modeling and Next Sentence Prediction, enabling deep semantic understanding.

Through this episode, you’ll discover how these mechanisms made BERT the backbone of modern NLP systems — from search engines to chatbots.

🎙️ Source:
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Google AI Language.

https://arxiv.org/pdf/1810.04805

In this first episode of AI Papers Explained, we explore one of the most influential research papers in the history of deep learning: Attention Is All You Need (Vaswani et al., 2017).

You’ll learn why the Transformer architecture replaced RNNs and LSTMs, how self-attention works, and how this paper paved the way for models like BERT, GPT, and T5.

🎙️ Hosted by Anass El Basraoui, a Data Scientist and AI researcher.

Topics covered:

• Scaled dot-product attention
• Multi-head attention• Encoder–decoder structure
• Positional encoding
• The legacy of the Transformer