built eight inception scenes with a 12b model on a laptop. four worked first time, three didn't. the top failed because it rotated a symmetric cone on its axis, which you cannot see. a top reads as spinning through precession and a mark on the side. youtube.com/shorts/4GPDK59dK4o

had a 12b model on my laptop build six places from inception, one html file each. the street that folds over, the cafe coming apart, the corridor on its side, the van, the snow fortress, the top. cut them together with narration. youtube.com/shorts/qhNpb...

attached a video clip to a local vision model today. the lab samples eight frames in the browser and hands them over as images, so the file never leaves the mac. 5b model, 3.7gb on disk, and it tells you what the clip is. youtu.be/jcIJ-ZwfyNw

trained a 9.4m model to do school maths with no llm at inference, then had a verifier pick one answer out of 96 samples. sampling more made it worse: 8.5% at 192, 8.3% at 288. the extras were plausible wrong answers it had to choose between. weights: hf.co/codelion/sprog-9m

wired two macs together over thunderbolt to see if local inference would go faster. it doesn't. what you get is room for a bigger model. any layer that spans the two machines pays the link, and that's far slower than memory bandwidth.

been benchmarking local models on a mac and tok/s turned out to be the wrong thing to watch. what you actually feel is the wait before the first token. it grows with prompt length and more ram doesn't help.

the 2x-not-10x coding thing rings true. the model writes stuff fast, but i still have to figure out if it did the right thing. half the time it already had a working fix in its first couple tries, i just couldn't tell which without running them. that checking is where the time goes now.

most brands can't say whether chatgpt or perplexity ever name them when someone asks for the best tool in their category. quick check: run ~10 buyer questions through the models, see who gets recommended. free scan that does it for you: merageo.com

resized a photo with the network turned off. 26 small browser tools that run in your tab and never upload your files. no AI, no server round-trip. merasaas.com

meraGEO checks if AI assistants actually cite your brand. give it a domain, it queries ChatGPT, Claude and Gemini and shows the % that name you. ran Hugging Face: 43/100, "low visibility", Claude 60% vs Gemini 30%. free scan. merageo.com/scan/huggingface.co

meraGPT: an always-on hosted box running Nous Research's open Hermes agent. memory lives on the machine as plain-text files. you bring your own OpenAI/Anthropic key so it stays on the box, no token markup. $9.99/mo, no GPU so it's for agent + text work. meragpt.com

Live: an AI dreaming a racing game, with two AI comedians narrating it around the clock. All generated on the fly. Say something in the chat and the hosts react to you. www.youtube.com/@thecrowdeda...

Star Viper - a Star Fox-style space shooter, and the whole game was coded by a local AI running on a Mac. No cloud, no API. Play it free: optiq.itch.io/star-viper

a world model for supertuxkart that runs in your browser at 10fps. no game engine. what got it fast: 48x24x8 latents not pixels, sampler distilled 8 steps to 2, and one onnx graph doing prefill + both denoise steps + kv cache. huggingface.co/spaces/codelion/neural-drive