I'm Matt Wood, and this is For Your Information. A live list of riffs and links for you and your agent, drawn from what I'm reading, noticing, questioning, concluding, and revising.
Thinking Fast and Slow in AI: the Role of MetacognitionMini-AGI with continual learning represents the kind of system-level adaptation the paper argues requires metacognition to bridge narrow AI toward more general intelligence
Real-Time Video Generation on TrainiumBoth address making computationally intensive AI workloads feasible on constrained or specific hardware targets through optimization, suggesting a broader trend of democratizing high-performance AI inference
Dust: Pretraining Transformers Without BackpropagationZeroth-order/backprop-free training dramatically reduces memory footprint (no gradient storage), directly enabling continual learning on constrained hardware like 8GB VRAM
Context Language ModelsContext Language Models' editable context and multi-agent support directly enables continual learning scenarios like Mini-AGI, where context must persist and evolve across learning episodes without full retraining
openTPU: Open-Source AI AcceleratoropenTPU running Qwen3/LLaMA on Kintex-7 FPGA cards complements the Mini-AGI goal of running capable AI on constrained hardware (8GB VRAM) — both democratize access to LLM inference on commodity/accessible hardware
Kev: Tiny Decision Models on QwenBoth pursue the same core goal of enabling capable AI model training and inference on consumer/local hardware with tight VRAM constraints, validating the viability of small local model approaches
Related
The Ultimate Guide to Multi-Harness RLMini-AGI's continual learning approach on constrained hardware shares the multi-harness RL concern of efficiently training agents across varied tasks and environments without catastrophic forgetting
TerrainSR: Terrain Super-Resolution ModelMini-AGI targets continual learning on constrained hardware (8GB VRAM); TerrainSR's ONNX support similarly targets efficient deployment, both addressing the challenge of running specialized deep learning models on accessible hardware
mattwood.fyi