Videos
Find videos about LLM models, hardware, runtimes, benchmarks and practical AI workflows.
Evaluates the AMD Ryzen AI Halo as a rival to NVIDIA's DGX Spark, questioning if the $4,000 developer machine makes sense.
Compares Ornith 1.0 9B and Qwen 3.5 9B on 8GB VRAM using Performance, Memory, Agency, OpenAI Human Eval, Expense Tracker, and Memory Match Game tests.
Locally installed test of Agents-A1, an agentic model designed to scale heterogeneous agentic abilities across multiple domains.
Claude Fable 5 optimized llama.cpp for Qwen3.6 35B-A3B on an RTX 3060, achieving a 64.5% prefill speedup via four patches, with results verified through self-benchmarking.
Installs and tests Headroom with Ollama to demonstrate compression of data read by an AI agent.
Compares QWEN3.6-35B-A3B inference speed with and without Multi-Token Prediction on an Apple M5 Max using Pi Coding Agent and LM Studio.
Demonstrates two practical tips and tricks to reduce costs when using the Anthropic Fable 5 model.
Covers updates on LongCat 2.0, Claude Fable 5, Sonnet 5, Agents A1, Gemini Omni Flash, Brain2Qwerty, MusViT, LiveEdit, VidiHand, OmniContact, PhysiFormer, Aspire, Comfy MCP, RDM, MrFlow, SimFoundry, C
Installs and tests google/tabfm-1.0.0-pytorch, a zero-shot tabular foundation model from Google Research.
NVIDIA Research's SpatialClaw is a training-free spatial reasoning agent that writes and revises Python code to understand 3D space.
LongCat 2.0 is compared against Claude, Kimi, GLM and DeepSeek to determine which performs best.
Tests Leanstral 1.5, an open-source code agent model designed for Lean 4.
Fable 5 is tested via Arduino render farm concepts, live GPU image rendering, cluster redundancy, city time travel simulations, computer repair, Ultracode games, and rage cheating games.
Installs and tests Needle running on Cactus, achieving 6000 tokens per second prefill and 1200 tokens per second decode speed.
Dan Shapiro maps AI coding tools like Claude Code and Codex to five self-driving car levels, identifying Level 3 as optimal for reliability while explaining why Level 5 Dark Factories aren't the prima
Covers redeployed Fable 5, BuseyBench, GPT-5.6 Sol preview, Claude Sonnet 5, Gemini Omni Flash, NotebookLM Video Overviews, Claude Science Workbench, Cursor for iOS, X MCP Server, Gemini Meet Notes, G
Compares Ornith 1.0 35B and Qwen3.6-35B-A3B-GGUF on a 16GB VRAM system using Driving Game, Fake Desktop, and Agent Maze coding tests.
A Supermicro server with dual Intel Xeon CPUs benchmarks local LLM inference using 4x Nvidia Tesla V100, 4x Nvidia Tesla P100, and 4x AMD Radeon Instinct MI25 GPUs.
Measures accepted-length speedup of DeepSeek's DFlash drafter on Gemma 12B using speculative decoding on a single GPU via DeepSpec resources.
Compares local GLM 5.2 on M3 Ultra 512GB via Inferencer App against cloud Claude Sonnet 5 and Opus 4.8 using CometAPI in Chat, Terminal, and AI Agent OpenCode modes for speed and capability.
Explores speculative decoding techniques used by Deepseek's dSpark system to enhance inference performance through confidence heads and hardware-aware algorithms.
Tests Laguna XS 2.1, a 33B model with 3B activated parameters per token designed for agentic coding and long-horizon tasks.
Explains FP16 to INT4 quantization via floating-point arithmetic, symmetric/asymmetric methods, block quantization, PTQ/QAT, and measures quality using perplexity and KL divergence.
Nemotron TwoTower employs a frozen reader and diffusion generator in parallel, achieving 2.4x speedup while retaining 98.7% of the original model's quality.
Compares Gemma 4B and Gemma 12B outputs on an e-commerce app, connects Claude Desktop via MCP server, and uses DeepEval, Ragas, and Hugging Face Evaluate with LLM-as-a-judge for assessment.
Explains Google's Open Knowledge Format (OKF), which formalizes Andrej Karpathy's LLM wiki pattern into plain markdown for agents to query folders without plugins, RAG pipelines, or vector databases.
Locally installs and tests Qwopus-3.6-35B-A3B-Coder, a thinking-off, token-efficient coding agent model.
Tests Ornith-1.0-35B performance in OpenClaw, Hermes, and Pi agent harnesses versus Qwen3.6 27B dense using its mixture-of-experts architecture.
Compares Ornith 35B and Qwen 3.6 35B-A3B using llama.cpp, llama-swap, and OpenCode on dual RTX 3090s while building a street-racing car OS, race-control interface, and live race simulator UI.
Tests Z.ai's GLM-5.2 via hosted web app, API, agent harness, and self-hosted setups using webpage building, Cursor integration, Remotion animations, and Inference.net traffic mirroring.
Compares Ornith 1 35B MoE against Qwen 3.6 35B A3B using Performance, Memory, Agency, OpenAI Human Eval, Sand Physics and Dungeon Crawler tests on a 16GB VRAM system.
Compares closed-source Sonnet 5 against open-model Ornith 35B to evaluate whether a local model can outperform its proprietary counterpart.
Tests Claude Sonnet 5 via browser OS, C++ skate game, subway scene, FPS, skydiving simulation, frontend design, city timeline, multimodal coding, low-poly F1, drum kit simulation, and terrain improvem
Tests Claude Sonnet 5 via real-world coding, SVG generation, agentic, and reasoning benchmarks against Opus 4.8 and GLM 5.2, covering CursorBench, token efficiency, pricing, and World of AI Benchmark
Demonstrates running local AI agents with Archestra and Ollama, watching tool calls live, and blocking misbehaving MCP servers in real time using open-source, self-hosted tools.
Demonstrates using natural language prompts in Reflect.run's recorder to generate tests via an AI Agent that auto-fills login credentials and form fields while reasoning through the UI.
Reviews and tests LongCat-2.0, a large-scale MoE language model with 1.6 trillion total parameters.
Locally installs and tests Ornith-1.0, a self-improving family of open-source models for agentic coding.
Demonstrates removing restrictions from Krea 2 Turbo via SGLang Quick Rebalancer, enabling reference images for consistency and targeted edits on the local AI image generation model.
Compares GLM-5.2 and Claude Opus 4.8 via skydiving simulation, Windows XP-style AI app, 3D model generation, and time travel scene tests evaluating reasoning, coding, creativity, and consistency.
Tests Qwen-AgentWorld-35B-A3B against Qwen3.6-35B-A3B using Performance, Memory, Agency, OpenAI Human Eval, Sand Physics, and Dungeon Crawler benchmarks on a system with 16GB VRAM and 32GB DDR4 RAM.
Demonstrates setting up DeepSeek's DSpark drafter on Qwen3-4B locally to reproduce accepted-length speedup using dspark_qwen3_4b_block7 on a single GPU.
Covers GPT-5.6 benchmarks, Minecraft demos, U.S. AI restrictions, Fable 5 embargo, Zhipu AI cybersecurity model, Anthropic enterprise growth, and Grok 4.5 testing rumors.
Tests Ornith-1.0-9B versus Gemma-4-12B in local AI coding tasks to evaluate whether Ornith's benchmark scores align with real-world performance.
Demonstrates wiring AnythingLLM, Pi, and n8n to a local llama.cpp engine via llama-swap for offline chat, document-based RAG, coding assistance, and automation.
Locally installed and tested Qwable-5-27B-Coder, a Qwen3.6-27B base lightly post-trained on 10 traces total.
Tests Ornith 1.0 local coding models via browser OS, subway scene/FPS generation, frontend design, C++, multimodal code, and agentic coding scenarios.
Installs and tests Krea 2, an image generation model, using ComfyUI.
Covers updates on Seedance 2.5, Happyhorse 1.1, GPT 5.6, Seed 2.1, Krea 2, Jalapeno chips, IBM sub nano chip, DanceOPD, Un-0, HIW 500, Lift4D, PerceptionDLM, Aleph ultrasound, Autodata and Sakana fugu
Runs MiniMax M3 locally using Inferencer App on Mac Studio M3 Ultra 512GB to evaluate performance against claims of higher scores than GLM, Kimi and Opus.