Videos
Find videos about LLM models, hardware, runtimes, benchmarks and practical AI workflows.
Benchmarks Moonshot AI's Kimi K3.1 Pro against Claude Opus 5.5, GPT-6 Astra, and Gemini 4 Argon using WoAI Bench for coding, 3D generation, and agentic tasks.
Evaluates Kimi K3.1 Pro via browser OS tests, C++ skate games, robot arms, Blender/Godot wrestling, CAD engines, Subway FPS, image-to-game, pool party, and Steve the PC Repair Man challenges.
Tests Kimi K3.1 Pro through pizza and donor kebab, floor is lava, KarmaBowl hidden bug, financial modeling, image analysis, multilingual tasks, benchmarks, and video understanding evaluations.
Weekly AI news overview covering major industry developments, security alerts, and product releases across leading artificial intelligence platforms and tools.
Demonstrates using Port to manage an AI Software Factory by assigning ownership, grading codebases from Bronze to Gold, and automatically fixing broken CI checks via approved issues.
Demonstrates setting up OpenClaw on a Contabo VPS via 1-click deployment, integrating an AI model provider, Telegram, and Skills, while discussing local hosting challenges.
Compares Qwen 3.8 27B at Q3 and Strata Coder using Performance, Seat Planner, Wave Simulation, Prism Relay, Diorama, and Blender Windmill tests on a 16GB GPU with 32GB RAM.
Tests Ling-3.0-flash-Sante, a healthcare-specialized model built on Ling-3.0-flash, through multiple evaluations and benchmarks.
Reviews DeepSeek-V4.1-Flash-1017 and DeepInfer through math, Unreal games, music, Blender VR, cancer identification, deep research, ideation, specs, benchmarks and Higgsfield integration.
Tests six open-source coding harnesses using four flash models on long-horizon coding, agent orchestration efficiency, and cybersecurity via prompt injection demonstrations.
Tests Tencent’s Hy-Memory-30B-A3B running locally with vLLM and the Hy Memory SDK to demonstrate its capability as a writer for AI agent memory.
Compares qwen3.8:27B on MacBook Pro M5 Max versus Claude Opus across four coding and data tasks, measuring time, cost, tokens, and output quality using Ollama and Claude Code.
Compares MiniMax M3.1 Flash and GPT-6.1 Sol generating a playable Call of Duty-style 3D FPS game using Three.js from scratch with identical prompts.
Evaluates MiniCPM5 1B fine-tuned via mlx-tune using SFT and DPO on a MacBook Pro M5 Max, achieving 84.3% brand compliance on 108 Bitext test messages versus 0% baseline.
Tests Vast-10M using Floor is Lava Game, tattoo generation, large Australian law corpora, and multilingual dating app name tasks.
Explains cognitive drain from managing multiple AI coding agent sessions via constant context switching, advocating for increased harness autonomy rather than additional concurrent sessions.
Installs and tests Flint, Recursive’s mixture-of-experts language model for coding agents, evaluating local inference speed, quality versus speed, benchmarks, and the rec inference engine.
Head-to-head comparison of Qwen 3.8 variants Swift and Austrian Qwen evaluates whether Swift’s reduced thinking data improves speed while maintaining intelligence through challenges and code review.
Demonstrates setting up AnythingLLM on Windows with Ollama to query PDFs locally, fixing the Agent mode web search issue by switching to Chat mode, and testing speed with Qwen 3.8 27B on an RTX 5090.
Evaluates Claude Opus, DeepSeek and local Qwen directing Higgsfield Katana via Higgsfield MCP to remake a viral edit from one selfie into AI video edits.
Evaluates DavidAU Qwen3.8 27B Turbo Brilliance 50X-Reasoning Instruct NEO MAX MTP via Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot tests.
Tests NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 using Portuguese image translation, circuit diagram reasoning, HTML generation, and benchmarks.
Locally installed Ornith-1.5-122B in GGUF format from AtomicChat is tested using six specific prompts covering bugs, recipes, animations, diagrams, scenarios, and languages.
Demonstrates letting an agent propose changes to its own harness using LangChain's Managed Deep Agents and LangSmith Engine, requiring human approval for learned improvements.
Evaluates Nvidia's Nemotron Bolt, tuned for 24GB cards, testing image understanding, everyday tasks, coding difficulty levels, and comparing against Qwen Flash.
Tests Strata 0.2 with Qwen3.8-Flash-Next via Apple Metal, covering setup, model loading, memory needs, and prompt processing plus generation speed.
Evaluates the 6B parameter FLUX 3 Image Klein model using version two of an image model testing suite to assess its strengths and weaknesses.
Hands-on tutorial managing LG, Samsung, or Android TVs using Hermes Agent and Ollama locally via IP address and a GitHub prompt.
LoRA fine-tunes Gemma 3 270M via Unsloth Studio on MacBook MLX GPU, routing support tickets into six queues with accuracy rising from 25.3% to 99.4%.
Tests Mercury 3 using tasks involving LINEAGE registry bugs, marble/cup reasoning, quote attribution, brand names in 80+ languages, and dragon tattoo HTML generation.
Review and tutorial covering MOSS Image usage in ComfyUI, comparison with Krea 2, layout control, LoRA addition, speed optimization, editing, and training.
Tests pplx-decider-v1.2, the successor to pplx-decider-v1.1-27b, built on the Qwen3.8-27B backbone.
Tests ThinkingCap and Swift 1.5 Qwen 3.8 27B fine-tunes in GGUF Q4_K_M on an RTX PRO 6000 using llama.cpp across nine benchmarks covering speed, tool calling, and visual tasks.
Tests looping Qwen3-4B, Qwen3.6-35B-A3B and NVIDIA's Dory on RTX 3060 using MMLU-Pro, CommonsenseQA and SWE-bench Verified benchmarks via llama.cpp.
Qwen 3.8 27B completed Stripe export tasks error-free in PilotDeck but failed numerical accuracy checks in another agent configuration.
Tests Underdog Saluki 27B by ConwayResearch across Performance, Memory, Reasoning, HumanEval Remix, Kanban, Sand Physics, Dungeon Crawler, Blender and Godot tasks.
Tests running DeepSeek V4.1-1017 on Mac Studio M3 Ultra using Inferencer and MLXe with Engram SSD offloading, covering benchmarks, vision inference, coding challenges, and game simulations.
Compares DeepSeek V4.1 Flash and Claude Opus 5.5 solving an auto-updating links task in Texodus, noting costs of 6 cents versus $5.45 and observing DeepSeek’s search behavior during execution.
Glimpse is tested using an epic split train video, a webcam clip, and five seatbelt images to evaluate its ability to question initial impressions and return to visual evidence before answering.
DeepSeek V4.1 Flash is tested on browser workflows, Corvette website generation, C++ games, Peri Robot arm control, Blender, Godot, FPS generation, pool party simulation, and cost breakdown.
Terminal Bench Mini evaluates DeepSeek V4.1 Flash 1017 on Strix/Gorgon Halo via DwarfStar/AI Toolbox Cockpit, showing 44% fewer tokens and 42% less agent time across eight matched tasks.
Benchmarked via WoAI Bench, evaluating DeepSeek V4.1 Flash 1017’s coding, 3D generation, agentic performance, and speed against GPT-6.1 Sol, GPT-6 Luna, Claude Opus 5.5, Gemini 4 Argon, and Kimi K3.
Routes 826 consumer complaints to 9 financial products using Qwen 3.5 0.8B fine-tuned via LoRA and PyTorch MPS on MacBook, improving accuracy from 43.6% to 73.1%.
Tests DeepSeek-V4.1-Flash-1017 on coastal erosion bug fixing, Pop It game building, multilingual condolence generation, scientific math reasoning, taxi image reasoning, and benchmarks.
Tests DeepSeek-V4.1-Flash-1017 against Qwen3.8-Flash-Next on a robotic arm 3D model and compares its World War I voxel art output with the original V4.1 Flash.
Locally installs and tests VoiceStudio for open-source voice cloning, voice design, video dubbing, dictation, and transcription using GitHub repository debpalash/VoiceStudio.
Performance test of Muse-Spark-Open-816B-A42B-MLX-Q4i using Inferencer v2.4.6 on Mac Studio M3 Ultra 512 against DeepSeek claims.
LoRA fine-tunes IBM Granite 4.0 1B on MacBook Pro using PyTorch MPS, improving invoice extraction accuracy from 55% to 77%.
Tests NVIDIA Nemotron 3.5 Bolt on coding and agentic work, comparing performance against Qwen3.8-27B, Qwen3.8-Flash-Next, and Strata Coder.
Locally installed NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is tested on agentic tasks including bug fixing, vision self-fix, mind games, headline translation, and circuit reasoning.