• Santulan - The Space Between Logic and Emotion

    There are moments in life when decisions do not feel like decisions at all. They feel like a quiet tug-of-war. One part of you is calm and almost clinical. It lays out the facts, measures the consequences, reminds you of what is practical, responsible, and safe. The other part does...

  • Why "Be Desirable" Is the Best Career and Life Advice You’ll Ever Get

    I remember a time when the atmosphere at work felt heavy. Rumors of layoffs floated through every hallway and even casual conversations carried a quiet tension. Out of curiosity and maybe a bit of nervousness, I asked my manager what he thought about everything that was happening. He didn’t answer...

  • Walking With Time Instead of Chasing It

    The Science of Time Perception Our brains do not count seconds the way a clock does. Instead, they build a picture of time from our senses and emotions. That is why time can feel different from one situation to another. When we are in danger or in pain, the brain...

  • Stop Quoting Papers, Start Building Models

    We are in an era where AI breakthroughs happen faster than most people can keep up. A new architecture drops on arXiv every week. You open LinkedIn, and someone’s already posting hot takes about the latest fine-tuning technique or debating whether LoRA is still relevant. Everyone’s got an opinion. Everyone...

  • Paper Summaries

    LLaMA-Omni: Seamless Speech Interaction with Large Language Models Introduction In-Depth Look at the LLaMA-Omni Model Architecture Constructing the Training Data: InstructS2S-200K Evaluating LLaMA-Omni’s Capabilities The Perfect Blend: Redefining RLHF with Mixture of Judges Introduction Architecture and Implementation Experimental Results Key contributions of the paper LLaMA-Omni: Seamless Speech Interaction with Large...

  • Diffusion Models

    The Diffusion Process: Reversing the Diffusion: Denoising Synthesizing New Data: Image Generation from Pure Noise Harnessing Text: Conditional Image Generation The Advantages of Diffusion Models: Quality, Diversity, and Stability Diving into the Mechanics of Diffusion Models Forward Diffusion: Reverse Diffusion: The Loss Function: Guiding the Denoising Symphony Time Embeddings: Providing...

  • Building your own Autonomous LLM Agent

    Plan & Tasks Task Object Plan Object Actions and Observations Memory LLM Agent Autonomous LLM Agent References In this blog, we are going to talk about LLM-based autonomous agents. Unlike the typical LLMs we’re accustomed to, which primarily focus on generating textual content, an autonomous LLM agent transcends this by...

  • Aligning LLMs using Human Edits

    Understanding SALT Loss Function Likelihood Loss ((L_p)) Unlikelihood Loss ((L_r)) Additionally Example Catastrophic Forgetting Issue The Concept of Imitation Edits Solution Leveraging RSALT for Effective Training Strategic Training Methodology SALT vs DPO/RLHF Conclusion There has been some cool progress in Aligning Large Language Models to human preferences, and guess what?...