Open weights do not mean data control: ultra-large models are too costly to host locally
EmetInteractive · reddit · 2026-08-26
A Reddit user argues that open weights do not equate to data control. Citing Kimi's 2.8T parameter model, which recommends 64+ accelerators, the author notes that only hyperscalers or well-funded labs can run it locally. Most users still send prompts to third-party (often foreign) endpoints, which is no different from using a closed API. For enterprise applications, the critical factor is whether data leaves the local environment. The author asks how others handle regulated or private data: do you self-host or rely on smaller, manageable models?
More from AGI Musings
- Pedro Domingos: Python is becoming the new assembly code — pmddomingos · 2026-08-26
- Paper: Characterizing Agentic Flooding of Government Services — RexDouglass · 2026-08-26
- AI supply chain paradox: Context providers give it away for free, yet vendors charge them — Ashwinizer · 2026-08-26
- GPT-Astra Assists Chip Design, Significantly Lowering Engineering Barriers — SumitGup · 2026-08-26
- Ox Alpha signal: AI economic rents commoditizing faster than expected — FinanceYF5 · 2026-08-26
- China cracks down on AI companion bots over emotional dependency fears — nordicinst · 2026-08-26