I got a 1.9B model to watch a restaurant camera and tell staff which tables need cleaning 🍽️ It runs on my laptop. No cloud, and it generates zero tokens. Built over a weekend with @liquidai d1-2B

DeepSeek V4.1 Flash 1017 on one Mac Studio M5 Ultra 512 GB, official lossless weights: 65.7 tok/s output with a 128k-token prompt, mlx-serve 26.11.2 defaults. Median of 5 runs, range 62.2–70.0. Other engines and the method below.

I wanted to test Fable 5.2's quality with one-shot prompts without human feedback. Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours! Play it here:

Another open prompt for DeepSeek v4.1 + dsh running locally on 4xDGX Sparks: A reference image + "recreate its architecture, composition, colors, and lighting in 3D, then animate a smooth, cinematic drone shot. Feel free to use anything you feel is better suited for this task. You have Blender or Unreal engine at your disposal, you can even use pur three.js code with html if it works well for you." Went with Blender + Python. Photo → procedural Blender scene → 12-second drone shot. Every roof, lantern, tree and ripple is generated in code. No textures, no imported models. Some stats: Session wall clock: 2 h 52 m Model generation time: 1 h 47 m Output tokens generated: 389k Average output speed: 65 tok/s Prompt tokens processed: 58.8M total Cache hit rate: 97.8% Context: ~478k Great job overall! No?

I already see clear signs that who is refusing AI assistance for programming it is starting to lag behind in a significant way, where this means: they are living in a bubble where losing work will happen faster compared to the other folks.

Bizarro Dimension: †@%$=777x Where the robots built a church for the last meat proxy. AI Tools: PhospheneAI (MiniMax H3 + YuE2, all rendered locally on my Mac), ChatGPT Images If you enjoy my work, a retweet, like, or share helps keep me motivated on this journey. Here’s what I learned from this project 🧵

Big news for virtualization: macOS 27.3 lets Core AI in Mac VMs use the host’s GPU and Neural Engine. In some tests, the new paravirtualized paths can run Core AI tasks up to 7× faster¹ than the guest CPU, achieving 50 to nearly 100% of the host’s native speed. The new Core AI accelerator currently requires private APIs and a restricted entitlement. Here's how to set it up: • Install macOS 27.3 on the host and guest. • Build a VM host app with com\.apple.security.virtualization and com\.apple.private.virtualization entitlements. You’ll need to relax AMFI for the second entitlement. • On VZVirtualMachineConfiguration, add the accelerators using _setAcceleratorDevices with: • _VZMacCoreAIDeviceConfiguration (GPU, adds PVAIDevice) • _VZMacNeuralEngineDeviceConfiguration (existing ANE accelerator, adds Neural Engine for Core AI) On the guest, enable the feature flag: sudo mkdir -p /Library/Preferences/FeatureFlags/Domain && \ sudo defaults write /Library/Preferences/FeatureFlags/Domain/ParavirtualizedCoreAI enable_pv_delegate -dict Enabled -bool true and restart the guest. You can now call Core AI on the guest. If set up properly, ComputeUnitKind.availableKinds should list gpu and neuralEngine. —— There's a second, currently disabled host-side feature flag (ParavirtualizedCoreAIHost/enable_save_restore). When enabled, the Core AI accelerator advertises support for saving and restoring VM state. —— 1. Tested on a Mac with an M4 Max chip using public models (YOLO, CLIP, Depth Anything) and matmul/convolution benchmarks. Crossing the VM boundary adds ~1 ms per inference call. Qwen3-0.6B in Apple’s llm-benchmark ran at about 10% of host speed. Core AI’s paravirtualized path repeatedly sets up host access to large shared KV-cache buffers for each generated token.