Pero a mayor capacidad del modelo más riesgo de memorización.
LLMS
-
Microsoft Launches Tool to Convert Files into Markdown
By
–
Breaking: Microsoft just shipped a tool that converts almost any file format into clean Markdown automatically. PDFs, Word docs, Excel sheets, PowerPoint decks, audio files, YouTube URLs. All of it. Clean output every time. Markdown is the format large language models process best and getting files into it has always been messier than it should be. No clean parser, broken layouts, scrambled text. This removes all of that. One install. Clean structured markup. Ready for any model to consume. (Link in the comments) [Translated from EN to English]
→ View original post on X — @aihighlight, 2026-04-09 12:31 UTC
-
Benchmarking AI Models: Balancing Accuracy with Cost Efficiency
By
–
Aquí siempre en el equipo de Noam, ojalá más y más benchmarks se reportarán cruzando accuracy con coste!
-
Mythos Model Evaluation: Why Single Benchmark Reporting Matters
By
–
Respecto a Mythos me han preguntado por qué en el vídeo de Youtube no he hecho mención a esta gráfica que todos estas comentando, y hay un par de motivos por el que descarté hablar de ello tras leer la Model Card. 1) Reportar la eficiencia de un modelo sobre un único benchmark
-
LCMs: A New Cognitive Layer for AI Beyond LLMs
By
–
LLMs predict tokens.
LCMs reason in concepts. That's the shift. Language-agnostic meaning → better abstraction, longer context, less repetition. If it works, this isn't an upgrade.
It's a new cognitive layer for AI. [Translated from EN to English]→ View original post on X — @ingliguori, 2026-04-09 12:17 UTC
-
Google’s MedGemma 1.5: Specialized 4B Medical Model Outperforms Larger Models
By
–
Really exciting: Google's MedGemma 1.5 packs 3D radiology, whole-slide pathology, longitudinal X-ray analysis, and clinical document understanding into a single open-weight 4B model (!), with a massive +47% F1 jump in pathology and +11% in MRI classification over v1. The specialized 4B model outperforms Gemini 3.0 Flash on out-of-distribution CT analysis, proving that targeted medical post-training beats raw scale. <3 Samuel Schmidgall (@SRSchmidgall) The MedGemma 1.5 technical report is out 👇 arxiv.org/pdf/2604.05081v1 — https://nitter.net/SRSchmidgall/status/2041973798589903260#m
→ View original post on X — @kimmonismus, 2026-04-09 12:08 UTC
-
50 Steps to Master Agentic AI in 2025-26
By
–
50 Steps to Master #AgenticAI in 2025-26
by @ingliguori #LLM #GenerativeAI #ArtificialIntelligence #MachineLearning -
GPT Model Thinking Feature Default Configuration Change Impact
By
–
I changed a default in the last release where thinking was off for GPT models unless configured; I bet a lot of folks ran into that and it caused bad performance.
-
RPG-Style World Built for AI Agents
By
–
ESTE TÍO CONSTRUYÓ UN MUNDO RPG PARA SUS AGENTES IA PORQUE SE CANSÓ DE DASHBOARDS Y TERMINALES
— Nico (@nicos_ai) 9 avril 2026
5 agentes, cada uno con su personaje y su puesto de trabajo, moviéndose por el mapa como en un juego de rol
cuando se acumulan suficientes problemas sin resolver, los agentes caminan… pic.twitter.com/7kaLVKj00mTHIS GUY BUILT AN RPG WORLD FOR HIS AI AGENTS BECAUSE HE GOT TIRED OF DASHBOARDS AND TERMINALS 5 agents, each with their own character and job role, moving around the map like in a role-playing game when enough unresolved issues pile up, the agents walk to a meeting point and
-
Opus loses warmth and personality in tone generation
By
–
Tal cual. El tono es lo más triste porque la calidez de Opus se pierde por completo, e incluso cuando fuerzas a que lo intente suena todo el rato a esto