Neat discoveries in AI
Three local AI runtimes on an M2 MacBook Air: typed decisions with feelsneat, speech with sayneat, and OCR and experimental vision models with looksneat.
- Held
- Event
- Data Berlin
- Location
- Berlin
local AI · Apple Silicon · MLX · speech · computer vision
I demonstrate the tools I built to use specialized models without a chat interface, explain their practical uses, and show some less dependable experiments. The audience questions get into why I wanted focused native runtimes in the first place.
Recording
Presentation
Notes
I asked the room who had bought an expensive computer to run AI locally, then whether they were happy with the decision. My own answer was mixed. Exploring smaller, specialized models had made the machine fun to use again, so I brought three open-source tools to demonstrate on my M2 MacBook Air with 24 GB of memory.
feelsneat runs decision models that return typed answers and probabilities rather than prose. sayneat began as a replacement for Apple’s say command and now handles local speech generation, transcription, and voice cloning. looksneat began with OCR; its experimental branches also let me investigate video action classification and predicted brain responses.
The full article, Neat discoveries in AI: why I built three local runtimes, expands the examples and explains the limits behind the demos. It distinguishes the OCR on looksneat’s main branch from the V-JEPA 2 and TRIBE v2 experiments, and separates my runtime work from the researchers’ models.
Finding the demos
The recording runs for about 27 minutes. These links jump to the main sections:
- feelsneat, beginning around 02:02, introduces typed decisions and shows the CLI.
- sayneat, beginning around 07:33, compares speech synthesis before moving into transcription and my cloned voice.
- looksneat, beginning around 14:34, discusses OCR and the experimental vision models. The camera demo follows around 19:22.
- The closing questions, beginning around 21:10, cover practical uses, contributions, and why I built focused runtimes instead of another large application.
The camera preview uses a fixed vocabulary of action labels, not general scene understanding or robot control. TRIBE v2 predicts group-average cortical BOLD signals for research, not an individual’s thoughts, emotions, or attention. Those experiments remain separate from looksneat’s OCR main branch.
The PDF is a web-optimized export of the original nine-slide Keynote presentation. It contains the introductions and project links; the live software demonstrations are in the recording.