研究
珀刻机从离网算力底线开始：加速器 TOPS、持续功耗、便携电池瓦时、模型大小、量化、存储与离线语料规模。
- 初始摘要：`knowledge/research-digests/2026-05-30-pockot-offgrid-compute-floor.md`
- 第一批基线：Apple M4、Meta Llama 3.2、Raspberry Pi AI HAT+、FAA battery rules、Snapdragon X Elite
- 估算规则：能力层级是启发式判断，不是保证
Pockot
研究
研究笔记Offline Updating Is Not Automatic Self-ImprovementPockot's north star includes a local system that can keep becoming more useful. That has to be handled carefully. Updating a retrieval corpus, adding notes, training a small adapter, and ...研究笔记TOPS Is a Starting Point, Not a DevicePocket-scale AI hardware is now marketed with large TOPS numbers. Pockot needs to track those numbers, but it also needs to keep them in their lane. TOPS does not specify memory, model qu...研究笔记Pocket Solar Is a Power Source, Not a ShortcutEnergy autonomy sounds simple until area, weather, angle, battery safety, charge electronics, and duty cycle enter the model. Pockot should treat solar as a slow input to a watt-hour ledg...研究笔记The Minimum Off-Grid Stack Has Five BudgetsWhat is the smallest stack that remains useful when the network is gone? Pockot should not answer with one chip. A durable pocket AI needs compute, memory, storage, energy, and trust budg...研究笔记Small Models Make the Floor MeasurableThe off-grid question is not "can a pocket device run the largest model?" It is "which tasks become useful at 1B, 3B, 7B, or 13B parameters under a strict power budget?" Pockot needs a mo...研究笔记Offline Agents Still Need a Trust ModelRemoving the cloud does not remove risk. An offline agent can still mishandle private data, follow malicious instructions in documents, run unsafe local tools, or produce unreliable outpu...研究笔记A Credible Offline AI Terminal Needs a BOM, Not a VibeIf Pockot were built today as a reference terminal, what public component baselines would keep the estimate honest? The answer should be a bill-of-materials scaffold, not a product promise.研究笔记Offline Knowledge Is a Storage BudgetAn offline AI terminal cannot "know everything" in the cloud sense. It can carry model weights, indexes, documents, and cached tools. Pockot needs to measure how much useful reference mat...研究笔记On-Device Learning Needs a Tight Boundary"Self-improving" sounds simple until the device has to train safely, remember correctly, and avoid corrupting itself. Pockot's question is narrow: what kinds of local adaptation are plaus...研究笔记A Local AI Has to Survive the Watt-Hour LedgerThe simplest off-grid calculation is also the one most likely to be skipped: watt-hours divided by sustained watts. Pockot starts with that ledger because a useful local model is not usef...