Discover AI Courses

Good LLMs Need BAD Data: The Shocking Truth

Discover the counterintuitive finding that including "bad data" in LLM training can lead to more controllable AI systems, as Harvard researchers demonstrate how this approach enables better post-training behavior mitigation.

• YouTube
• 35 minutes
• Self-Paced
• Free Video

The 6Dim Key: Unlock In-Context Learning - How AI Systems Learn from Few-Shot Examples

Discover how in-context learning works in AI systems, quantum computing, and biological compounds through the 6-dimensional key framework.

• YouTube
• 31 minutes
• Self-Paced
• Free Video

NEW AI Thought Machine - Artificial Time (No Transformer)

Explore a new neuronal alternative to transformer architecture: Continuous Thought Machine (CTM) with artificial dimensions for dynamic neuronal synchronization to improve AI reasoning.

• YouTube
• 19 minutes
• Self-Paced
• Free Video

Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures

Explore DeepSeek's latest research paper detailing their next model architecture with innovations in Multi-head Latent Attention, Mixture of Experts, FP8 training, and Multi-Plane Network Topology for enhanced AI infrastructure.

• YouTube
• 23 minutes
• Self-Paced
• Free Video

AI Video Generation Just Got Smart: GPDiT - Generative Pre-trained Autoregressive Diffusion Transformer

Discover how GPDiT (Generative Pre-trained Autoregressive Diffusion Transformer) is advancing AI video generation technology through this detailed explanation of the research paper and its implications.

• YouTube
• 27 minutes
• Self-Paced
• Free Video

0.6B to 235B: How to Build Qwen3's Dual Mode AI

Explore the pre-training phases and post-training elements of Qwen3's dual mode AI, including the Strong-to-weak Distillation process for smaller models.

• YouTube
• 19 minutes
• Self-Paced
• Free Video

RLHF's Missing Piece: Qwen's World Model Aligns AI with Human Values - GRPO

Explore Qwen's new WorldPM model that encodes human preferences at scale, solving key RLHF challenges by creating a world model that better aligns AI with human values.

• YouTube
• 21 minutes
• Self-Paced
• Free Video

MCP Security Threats: Protect Your AI Agents

Discover essential security measures for AI agents, focusing on threats to Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication, with countermeasures to protect privacy and confidential data.

• YouTube
• 22 minutes
• Self-Paced
• Free Video

Neural Scaling for Small LMs and AI Agents - How Superposition Yields Robust Neural Scaling

Explore how AI models efficiently represent information through superposition, revealing why larger foundation models improve following power-law decay patterns.

• YouTube
• 28 minutes
• Self-Paced
• Free Video

How to Make Smaller LLMs R1-Smart

Discover how to enhance smaller language models with R1-Smart techniques from UC Berkeley researchers, exploring reasoning capabilities and limitations after SFT.

• YouTube
• 31 minutes
• Self-Paced
• Free Video

Risk-Aware RAG for Smarter AI Reasoning (ARISE)

Discover ARISE, a novel framework using risk-adaptive Monte Carlo Tree Search to guide knowledge-augmented reasoning in LLMs, balancing exploration and exploitation for more robust and efficient reasoning capabilities.

• YouTube
• 18 minutes
• Self-Paced
• Free Video

Massive CoT Problems: Sonnet 3.7 Reasoning - Chain-of-Thought Reliability in AI Models

Explore the reliability of Chain-of-Thought reasoning in AI models like Claude 3.7 Sonnet, examining how these reasoning processes impact AI safety research and potential issues with trusting what models say in their thought processes.

• YouTube
• 21 minutes
• Self-Paced
• Free Video

AI Reasoning on a Möbius Strip

Explore why LLMs benefit from generating extra reasoning tokens and which aspects of task complexity determine optimal reasoning length in this Harvard research presentation.

• YouTube
• 40 minutes
• Self-Paced
• Free Video

Self-Principled Critique Tuning with DeepSeek-GRM-27B

Discover the innovative Self-Principled Critique Tuning (SPCT) method and DeepSeek-GRM-27B model, exploring how this new approach could revolutionize reasoning capabilities in AI systems.

• YouTube
• 19 minutes
• Self-Paced
• Free Video

Llama 4 Maverick 400B: First Real-World Reasoning Test

Explore a real-world reasoning test comparing Llama 4 Maverick 400B with Claude 3.7 Sonnet, evaluating Meta's claims about the model's performance on logic and causal reasoning tasks.

• YouTube
• 17 minutes
• Self-Paced
• Free Video