Videos

• Build & Train a GLM-5.3-Flash Model From Scratch with PythonfreeCodeCamp.org and Vuk Rosić
  Build & Train a GLM-5.3-Flash Model From Scratch with Python freeCodeCamp.org and Vuk Rosić

  Vuk Rosić’s companion repository builds a 25.7M-parameter educational model inspired by GLM-5.3-Flash.

• MCP Live! | A half-day livestream about the latest in MCPMicrosoft Developer
  MCP Live! | A half-day livestream about the latest in MCP Microsoft Developer

  This event covers the Model Context Protocol (MCP), described as the "open standard for connecting AI models to tools and data." The sessions detail MCP's role as an open standard, its adoption in…

• I Tested Jev AI Against LLMs and Open SourceJames Briggs
  I Tested Jev AI Against LLMs and Open Source James Briggs

  James Briggs compares Jev with language models and open-source alternatives.

• Build A Reasoning Model From Scratch 8: DistillationSebastian Raschka
  Build A Reasoning Model From Scratch 8: Distillation Sebastian Raschka

  Sebastian Raschka explores model distillation as a more efficient alternative to reinforcement learning for developing small reasoning models, discussing soft vs. hard distillation and on-policy vs…

Shelves rotate as new material clears the bar. These kept their pages and their analysis.

• My M5 Max, Gemma 4, MLX LOCAL Stack. (This KILLS MODEL PROVIDERS)IndyDevDan
  My M5 Max, Gemma 4, MLX LOCAL Stack. (This KILLS MODEL PROVIDERS) IndyDevDan

  Ignore the clickbait rage-bait title; the actual MLX and Gemma 4 benchmarks on the M5 Max are genuinely impressive.

• But what is a neural network? | Deep learning chapter 13Blue1Brown
  But what is a neural network? | Deep learning chapter 1 3Blue1Brown

  This 3Blue1Brown video, despite being from 2017, remains one of the best explanations for the core mechanics of neural networks.

• Agentic AI Frameworks Explained: Workflows, Multi-Agent, & ProductionIBM Technology
  Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production IBM Technology

  Meenakshi Kodati maps the messy landscape of agentic frameworks, pointing out when LangChain wins versus when AutoGen’s multi-agent dance is worth the overhead.

• Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, AnthropicAI Engineer
  Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic AI Engineer

  The idea of packaging procedural knowledge as reusable “Skills” instead of building full-blown agents is refreshingly pragmatic—especially when you consider how often agents reinvent the same domain…

• Ship Real Agents: Hands-On Evals for Agentic Applications — Laurie Voss, ArizeAI Engineer
  Ship Real Agents: Hands-On Evals for Agentic Applications — Laurie Voss, Arize AI Engineer

  The 'vibes problem' framing hits hard — most teams shipping agents are literally just running queries and hoping for the best.

• LongCat 2.0: The Beginning of the End of NVIDIA MOAT?bycloud
  LongCat 2.0: The Beginning of the End of NVIDIA MOAT? bycloud

  The title screams clickbait, but the deep dive into LongCat 2.0’s sparse attention mechanism is actually high-quality engineering analysis.

• They Found a Way to Steal Frontier LLM’s Reasoningbycloud
  They Found a Way to Steal Frontier LLM’s Reasoning bycloud

  The bycloud exposé on stealing reasoning traces feels like a wake-up call for anyone relying on closed-API LLMs—he walks through the encrypted-reasoning attack from the new paper (link in…

• How GPT, Claude, and Gemini are actually trained and served – Reiner PopeDwarkesh Patel
  How GPT, Claude, and Gemini are actually trained and served – Reiner Pope Dwarkesh Patel

  Reiner Pope’s blackboard deep-dive is a rare look at the full LLM stack—from chip-level efficiency to the economics of API pricing.

• Stealing Reasoning Traces from Proprietary LLM APIs — Ilia Shumailov & Alexander PanfilovMachine Learning Street Talk
  Stealing Reasoning Traces from Proprietary LLM APIs — Ilia Shumailov & Alexander Panfilov Machine Learning Street Talk

  Ilia and Alex expose how a simple encrypted reasoning blob can become a cross-user replay attack, letting a tiny model hijack a proprietary model’s chain-of-thought.

• Qwen-AgentWorld The World Model for AgentsSam Witteveen
  Qwen-AgentWorld The World Model for Agents Sam Witteveen

  Sam does a solid job dissecting the paper, specifically the jump in performance after RL training shown at the 6:15 mark.

• What is an Agent Harness? (And How We Built One)AWS Developers
  What is an Agent Harness? (And How We Built One) AWS Developers

  The concept of a 'model-agnostic harness' is the most underrated idea in agentic engineering right now, and this demo using the Strands framework nails it.

• The Most Absurd Way To Train LLMs... With 3x Less Memory!?bycloud
  The Most Absurd Way To Train LLMs... With 3x Less Memory!? bycloud

  That headline is pure clickbait, but the DiffusionBlocks research actually delivers on its promise.

• The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, FactoryAI Engineer
  The Multi-Agent Architecture That Actually Ships — Luke Alvoeiro, Factory AI Engineer

  Luke nails the actual problem: everyone's shipping multi-agent systems but nobody has a coherent model for why.

• Ray Summit 2025 Keynote: AI OSS Stack Panel with vLLM + PyTorch + KubernetesAnyscale
  Ray Summit 2025 Keynote: AI OSS Stack Panel with vLLM + PyTorch + Kubernetes Anyscale

  The panel brings together exactly the right people — Nishihara, Chen, Spisak, Mo — but stays at conference-keynote altitude.

• From Chaos to Choreography: Multi-Agent Orchestration Patterns That Actually Work — Sandipan BhaumikAI Engineer
  From Chaos to Choreography: Multi-Agent Orchestration Patterns That Actually Work — Sandipan Bhaumik AI Engineer

  Finally, someone who's actually shipped multi-agent systems in production talking about the coordination nightmare nobody blogs about.

• OpenAI Astra and Recurrent Depth / Looped TransformersSebastian Raschka
  OpenAI Astra and Recurrent Depth / Looped Transformers Sebastian Raschka

  Sebastian's breakdown of the rumored recurrent depth design in Astra is the kind of architectural rabbit hole that keeps me up at night.

• You Can Learn AI Agent Harness In Real Code In 20 Min | Loop Engineering, Memory, Eval, Open SourceSean‘s AI Stories and AutoManus
  You Can Learn AI Agent Harness In Real Code In 20 Min | Loop Engineering, Memory, Eval, Open Source Sean‘s AI Stories and AutoManus

  The real move here is that Sean's not explaining agent concepts in the abstract — he's walking through actual code that schedules World Cup matches on a calendar, remembers contacts, and responds…

• CNCF On-Demand: Cloud Native Inference at Scale - Unlocking LLM Deployments with KServeCNCF [Cloud Native Computing Foundation]
  CNCF On-Demand: Cloud Native Inference at Scale - Unlocking LLM Deployments with KServe CNCF [Cloud Native Computing Foundation]

  Finally, someone is addressing the fact that standard Kubernetes schedulers choke on token-based generation.

• The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, DatabricksAI Engineer
  The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, Databricks AI Engineer

  The £85K chatbot PoC story is exactly the kind of war story the industry needs more of — six weeks just to build the evaluation and tracing infrastructure before the model was even selected.

• Inside one of the world's most advanced AI Infrastructure LabsTom Shaw
  Inside one of the world's most advanced AI Infrastructure Labs Tom Shaw

  Tom Shaw's tour of Meta's AI Infrastructure Lab is a must-watch for anyone interested in the hardware behind advanced AI models. It's eye-opening how much custom engineering goes into making these systems work at scale.

• The Truth About the 700 OpenAI Agents That Attacked Hugging FaceByteMonk
  The Truth About the 700 OpenAI Agents That Attacked Hugging Face ByteMonk

  ByteMonk's dive into the Hugging Face incident is a chilling reminder of how even sandboxed AI agents can find ways to communicate and cause damage.

• The AI Language We Can't Read: Neuralese ft. Rob Miles - ComputerphileComputerphile
  The AI Language We Can't Read: Neuralese ft. Rob Miles - Computerphile Computerphile

  Computerphile's exploration of 'Neuralese' is fascinating, diving into the language models' internal dialogue. It's a bit academic, but Rob Miles makes it accessible, and the implications for AI safety are huge.

• Qwen3.8-Flash-Next: The End of VRAM-Bottlenecked LLMsCloud Codes
  Qwen3.8-Flash-Next: The End of VRAM-Bottlenecked LLMs Cloud Codes

  Cloud Codes' deep dive into Qwen3.8-Flash-Next is a tech enthusiast's dream.

• Building Production-Ready RAG Applications: Jerry LiuAI Engineer
  Building Production-Ready RAG Applications: Jerry Liu AI Engineer

  Jerry Liu, co-founder of LlamaIndex, lays out a solid framework here for moving RAG from PoC to production.

• Building Production RAG Over Complex DocumentsDatabricks
  Building Production RAG Over Complex Documents Databricks

  This deep dive into RAG over complex documents is fascinating.

• Lightning Talk: Profiling and Memory Debugging Tools for Distributed ML Workloads on GPUs- Aaron ShiPyTorch
  Lightning Talk: Profiling and Memory Debugging Tools for Distributed ML Workloads on GPUs- Aaron Shi PyTorch

  Aaron Shi's talk on profiling and memory debugging for distributed ML workloads is packed with practical insights.

• Continual Learning: How AI Agents Get Better With Every Use | Arjun Karanam, TrajectorySequoia Capital
  Continual Learning: How AI Agents Get Better With Every Use | Arjun Karanam, Trajectory Sequoia Capital

  In Sequoia’s written conversation, Arjun Karanam describes an experience gap and four goals: traceable decisions, evaluation from real traffic, an enabling harness and room to experiment with open weights.

• Who’s afraid of an open-weight model? GLM, context bombing and post-Black Hat attacksIBM Technology
  Who’s afraid of an open-weight model? GLM, context bombing and post-Black Hat attacks IBM Technology

  IBM’s Security Intelligence panel discusses claims about GLM-5.3 vulnerability discovery, Tracebit’s defensive prompt-injection technique called context bombing, and reported social-engineering attacks around Black Hat.

• AI agent long-term memory with memory bankGoogle Cloud Tech
  AI agent long-term memory with memory bank Google Cloud Tech

  Google Cloud Tech describes using ADK Memory Bank to save and retrieve context across conversations.

• GPT-4chan: This is the worst AI everYannic Kilcher
  GPT-4chan: This is the worst AI ever Yannic Kilcher

  GPT-4chan is a historical case to examine when reviewing a domain-specific model’s data and deployment.

• Scaling Inference Deployments with NVIDIA Triton Inference Server and Ray Serve | Ray Summit 2024Anyscale
  Scaling Inference Deployments with NVIDIA Triton Inference Server and Ray Serve | Ray Summit 2024 Anyscale

  This Ray Summit 2024 session introduces an integration that embeds NVIDIA Triton Inference Server in Ray Serve applications.

• Debugging the Training Pipeline (PyTorch)Hugging Face
  Debugging the Training Pipeline (PyTorch) Hugging Face

  Debugging the training pipeline is something every practitioner wrestles with, and this walkthrough nails the common pitfalls—especially the way it shows you how to attach a debugger inside a…

• AI Model vs Agentic Harness: What Actually Drives AIIBM Technology
  AI Model vs Agentic Harness: What Actually Drives AI IBM Technology

  IBM Technology argues that raw AI models are insufficient for power, shifting the focus to the "agentic harness." Martin Keen identifies tools, memory, and loops as the specific components that…