InAI AdvancesbyFabio Yáñez Romero·Jul 27The Illusion of Deep Learning: AI’s Best Optimizer Is Deaf and AmnesicHow Muon and M3 fix AdamW’s two limitations: one inside every step, the other across the entire training run.A response icon1A response icon1
InAI AdvancesbyFabio Yáñez Romero·Jun 30The Illusion of Deep Learning: Why AI Needs Brainwaves to RememberStatic MLPs are holding back Large Language Models. Discover how the Continuum Memory System (CMS) escapes the “Frequency Zero” trap and…A response icon1A response icon1
InAI AdvancesbyFabio Yáñez Romero·Jun 24The Illusion of Deep Learning: How HOPE Gives LLMs NeuroplasticityBeyond static weights: Discover how HOPE, Living Gates, and Delta Gradient Descent are building the first truly self-evolving neural…A response icon7A response icon7
InTowards AIbyFabio Yáñez Romero·Jun 23Your Second Brain Doesn’t Need RAG. It Needs a MapKarpathy’s markdown system for retrieving information with agents appears to be the best way to avoid losing context compared with RAG and…A response icon1A response icon1
InTowards AIbyFabio Yáñez Romero·May 7Harness: The Era Enterprises Were Built ForPrompt engineering rewarded teams unburdened by process. The harness rewards the opposite — and enterprises have been building it…A response icon1A response icon1
InAI AdvancesbyFabio Yáñez Romero·Apr 26MIRAS: The Blueprint Behind Transformers, Mamba, and TitansGoogle’s new framework reveals every modern sequence model is solving the same four-choice optimisation problemA response icon2A response icon2
InTowards AIbyFabio Yáñez Romero·Apr 1021 Models in One Pipeline: What Actually Drives Knowledge Graph QualityGemma3 needed 27B parameters to extract a usable graph from legal text. With Gemma4’s MoE architecture, a model a fraction of that size…A response icon2A response icon2
InTowards AIbyFabio Yáñez Romero·Mar 30From Text to Knowledge Graph in One Command: Building a Modular LLM-Backed FrameworkHow abstraction, registries, and a clean CLI turn chaotic text into structured, queryable knowledge — without locking you into a single…A response icon6A response icon6
InAI AdvancesbyFabio Yáñez Romero·Feb 19Google’s Titans: 4 Ways to Wire Long-Term Memory Into a TransformerMAC, MAG, MAL, and Standalone LMM — how the same memory module creates different architectures with different trade-offsA response icon1A response icon1
InAI AdvancesbyFabio Yáñez Romero·Feb 7Titans: How Google Taught AI to Surprise, Forget, and RememberGoogle’s new memory system for language models — and why attention alone was never enoughA response icon5A response icon5