Here are 5 good reasons to run your AI locally

The points of the article on confidentiality and the reduction of the carbon footprint are correct. Consequently, On the other hand. Consequently, to be complete:
• Developer access: There is no major difference in terms of possibilities for a developer, because the cloud models (GPT-4, GPT-5, etc.) are also accessible via APIs. In addition, The barrier is therefore not technical but linked to the internet connection and the cost of use.
• Power significantly lower locally:
1- Model size: GPT-4/5 Cloud runs on approximately 200 to 1000 billion “neurons” (parameters) while local model on RTX 4080 is generally limited to 7 to 13 billion. For example, or 33 billion in highly compressed version.
2- Contextual memory: GPT-4/5 Cloud can treat the equivalent of 300 to 800 pages of text in a single conversation (128k to 1m tokens). against 10 to 20 pages for a local model (4k here 5 good reasons run to 8k tokens).
3- Material resources: GPT-4/5 Cloud uses clusters with 640 GB of ultra-fast true. hundreds of specialized processors connected by NVLink/NVSWITCH, as well as ~ 1.9 TB of System memory and several NVME local storage teraoctes per node. An RTX 4080 has 16 GB of VRAM. around 1/40ᵉ of the raw power of a single “node” GPT-4, with typically 32 to 64 GB of system memory and 1 to 2 TB of NVME storage on a high-end PC.

In summary. if the local execution of an AI offers autonomy and confidentiality, it involves working with much smaller and less efficient models than those available via the cloud. Everyone has their own choice.