OpenSource Podcast Summaries
OpenSource on Yedapo: 14 summarized podcast and YouTube episodes. Each includes key takeaways, core concepts and notable quotes with timestamps.
Kimi K3 Just Revealed The Worlds Most Powerful AI (Beats Fable 5 and GPT-5.6)
TheAIGRID
Jul 17, 2026
Moonshot’s new Kimmy K3 model, featuring 2.8 trillion parameters, effectively closes the performance gap between open-source and top-tier closed-source models like Claude Fable 5. By utilizing an efficient mixture-of-experts architecture, it delivers frontier-level coding and reasoning capabilities at a fraction of the cost, signaling a major shift in the global AI competitive landscape.
Key insight: Kimmy K3 is the first open-weights model to reach 2.8 trillion parameters, achieving top-tier performance on coding and reasoning benchmarks while costing significantly less than its closed-source counterparts.
Nicht-Steam-Spiele unter Linux ✓ | Faugus, Heroic & Lutris
c't 3003
Jul 3, 2026
Linux-Gaming ist längst kein Nischenprojekt mehr, doch abseits von Steam erfordert die Installation von Launchern wie Epic, GOG oder Battle.net oft manuelle Konfiguration. Tools wie Heroic, Lutris und der aufstrebende Faugus-Launcher automatisieren diesen Prozess, wobei die Wahl des Launchers entscheidend für den Komfort und die Kompatibilität ist.
Key insight: GE-Proton (GloriousEggroll) ist der entscheidende Community-Patch, der Spiele außerhalb von Steam erst spielbar macht, da er notwendige Codecs und Fixes enthält, die im Standard-Proton fehlen.
AI Just Entered A New Era
Two Minute Papers
Jul 1, 2026
The emergence of GLM 5.2 proves that open-weights AI is rapidly approaching the capabilities of proprietary systems like Claude. By utilizing advanced training techniques like multi-token prediction and individual step-grading, these models offer a future where individuals can own and control powerful, uncensored intelligence rather than relying on restrictive, government-regulated corporate platforms.
Key insight: GLM 5.2 employs an ingenious 'anti-hacking' measure that lets an AI attempt to cheat on benchmarks, only to provide it with false data so it gains no advantage, effectively neutralizing benchmark manipulation.
GPT 5.6 banned, Fable banned… it’s actually over.
David Ondrej
Jun 26, 2026
The recent government-led blocking of frontier AI models like GPT-5.6 signals a shift toward a 'permanent underclass' where only elites access superintelligence. To avoid digital enslavement, individuals must pivot to self-hosting open-source models and decentralizing data contribution to break the duopoly of closed-source labs and state control.
Key insight: If you don't have access to superintelligent AI in five years, you will be effectively crippled and unable to compete, making self-hosting your own models as essential as having your own electricity or water supply.
Eine Open-Source KI soll's mit Claude Fable aufnehmen?
The Morpheus Tutorials
Jun 19, 2026
GLM 5.2 aus China schließt die Lücke zu führenden US-Modellen wie Claude Fable erstaunlich schnell. Während Minimax und Kimi K2.7 bei der App-Entwicklung mit Design-Schwächen kämpfen, zeigt sich GLM 5.2 als hochfunktionaler Code-Generator, der die Betriebskosten massiv senkt und die Lücke zu westlichen Modellen auf nur noch wenige Monate verkürzt.
Key insight: Das GLM 5.2 Modell konnte in unabhängigen Elo-Rankings in der Code-Kategorie bereits Claude Fable schlagen, obwohl es nur einen Bruchteil der API-Kosten der US-Konkurrenz verursacht.
Los Modelos de IA Chinos son Salvajes
midudev
Jun 18, 2026
La industria tecnológica china está desafiando el control de las empresas estadounidenses sobre la IA. Con lanzamientos como GLM 5.2 y Kimi k2.7, ofrecen modelos de alto rendimiento y bajo costo que apuestan por el código abierto, cuestionando la sostenibilidad de los modelos cerrados frente a la apertura global.
Key insight: El modelo Kimi k2.7 ofrece una eficiencia de costos cinco veces menor que competidores de gama alta como Opus 4.8, manteniendo un rendimiento competitivo que pone en duda la superioridad tecnológica de los modelos cerrados occidentales.
OpenCode Desktop: Crea Proyectos con Múltiples Agentes IA al Mismo Tiempo
Fazt Code
Jun 18, 2026
OpenCode Desktop es una interfaz agnóstica que permite gestionar múltiples suscripciones y modelos de IA en paralelo, evitando la dependencia de un único proveedor. A diferencia de otras herramientas, ofrece un entorno flexible para ejecutar proyectos, integrar MCPs y automatizar tareas mediante 'skills' sin estar limitado al ecosistema de GPT o Claude.
Key insight: A diferencia de la mayoría de las herramientas comerciales que te obligan a usar su propia suscripción, OpenCode permite autenticar tus cuentas existentes (incluyendo GPT) y alternar entre ellas en una misma sesión de trabajo.
“Curing All Disease by next century is too conservative" - Mark Zuckerberg
No Priors: AI, Machine Learning, Tech, & Startups
Jun 10, 2026
Mark Zuckerberg, Priscilla Chan, and Alex Reeves are pivoting their philanthropic efforts toward a 'virtual biology' initiative that integrates frontier AI with wet-lab research. By building hierarchical models of cells and proteins, they aim to move science from discovery-based to engineering-based, providing open-source tools to accelerate the global effort to cure, prevent, and manage all diseases.
Key insight: The team successfully folded over 1.1 billion proteins using a general-purpose language model, discovering that the ability to design new proteins and antibodies emerged as a property of the model simply learning to 'understand' protein language, rather than being explicitly programmed for those tasks.
FFmpeg: The Incredible Technology Behind Video on the Internet | Lex Fridman Podcast #496
Lex Fridman
May 6, 2026
FFmpeg and VLC form a critical, volunteer-driven binary star system that underpins nearly all modern digital media. Despite their massive impact on platforms like Netflix and YouTube, these projects rely on a small, dedicated community of engineers who prioritize technical excellence and open-source ethics over corporate profit, often rejecting millions to maintain their independence.
Key insight: FFmpeg and VLC are so ubiquitous that they process video for billions of devices, yet the core development teams for these global pillars consist of fewer than 15 people.
¿Cuál es la mejor IA para programar?
HolaMundo
Apr 27, 2026
La ejecución local de modelos de lenguaje para programación ya es viable en equipos con 32GB de RAM, incluso con tarjetas gráficas integradas. Si bien la velocidad varía según el tamaño del modelo y la cuantización, modelos como Qwen 3 ofrecen un equilibrio sorprendente entre capacidad y rendimiento, permitiendo automatizar tareas de codificación sin depender de la nube.
Key insight: El modelo Qwen 3.6 de 27 mil millones de parámetros logra superar a Opus 4.5 en benchmarks, demostrando que los modelos open source de menor tamaño son actualmente altamente competitivos frente a alternativas propietarias.
Gemma 4 E2B + Hermes Agent + vLLM: Multimodal AI Stack Locally for Free
Fahd Mirza
Apr 3, 2026
Fahad Mirza demonstrates how Google's Gemma E2B integrates with Hermis agent to create a fully local, multimodal powerhouse. This 2-billion parameter model shatters the myth that high-tier vision and audio capabilities require massive server farms. By leveraging VLLM, Mirza proves that edge devices can now execute complex autonomous agency with minimal hardware.
Key insight: Despite its tiny footprint, Gemma E2B successfully handles audio transcription in over 100 languages and performs vision-based OCR while consuming less than 8GB of VRAM.
MolmoWeb - Fully Open Multimodal Web Agents - Control Browser Locally
Fahd Mirza
Mar 25, 2026
Fad Miraza reveals how AI2’s Molmo Web bypasses HTML entirely, controlling browsers through raw visual perception. By mimicking human sight, this open-source 8B model challenges the dominance of massive, closed-source agents in complex navigation tasks.
Key insight: Molmo Web outperforms agents built on much larger closed models by ignoring structured page data and relying exclusively on screenshots to execute actions.
Chinesische LLMs im Test
The Morpheus Tutorials
Feb 19, 2026
Chinesische KI-Modelle wie Minimax 2.5, Kimi und GLM-5 haben technologisch massiv aufgeholt und konkurrieren nun direkt mit Modellen wie Claude Opus oder GPT-5. Während sie bei Coding-Aufgaben und Effizienz beeindrucken, zeigen sie bei politisch sensiblen Themen weiterhin die erwartete Zensur und Bias-Problematik.
Key insight: Minimax 2.5 bietet ein Flaggschiff-Modell-Erlebnis für nur 50 Euro im Monat bei 1000 Anfragen alle 5 Stunden, was die Kosten-Nutzen-Rechnung gegenüber US-Anbietern wie Anthropic radikal verändert.
L'attaque informatique la plus sophistiquée de l'Histoire
Micode
Nov 5, 2024
Un développeur mystérieux, "Jia Tan", a infiltré pendant trois ans le projet XZ pour y introduire une porte dérobée sophistiquée. Cette attaque "Supply Chain" visait à compromettre les serveurs SSH mondiaux, mais a été déjouée in extremis par un ingénieur Microsoft détectant une latence de 500 millisecondes. L'incident révèle la vulnérabilité critique des logiciels open source maintenus par des bénévoles isolés.
Key insight: La porte dérobée n'était pas présente dans le code source public sur GitHub, mais uniquement dans les versions packagées (tarballs) envoyées aux distributions Linux, une ruse technique conçue pour contourner les audits de code automatisés.