- How much water is used by a data center?
AI generated answers are quick, fun and exciting. Similar to those days Google, AI chatbot provides answers to all our questions, analyzes uploaded question papers, creates images and videos and seems to gain more power with time. But there is another dark reality to AI.
AI responses require a lot of water. How does this happen?
In order to answer our questions, the machines hosting the LLMs are working hard behind the scenes. The specialized chips in the machines heat thinking and working out the answers. The heat that is generated by the many sets of machines needs to be cooled and this requires constant running water. This brings the reality of cool water for running data centers and a lot of water is needed considering the number of questions asked and the number of answers that have to be given. Compare it a mobile phone getting heated or a laptop getting heated due to overwork.
Asking a small question might not look like it is using a lot of water but a lot of users making such queries drains water resources.
A rough estimate by Claude states that one image uses about 0.25 ml – 0.32 ml of water per response. Complex videos use approximately about 4.1 L of water per response. It is not the single image or video creation that uses much water but the cumulative usage of water by millions of users and the edits and re-edits they do to the images and videos that puts a strain on natural resources.
2. Best way to ask questions to LLMs for safeguarding your privacy
The best way to ask any of the LLMs like ChatGPT or Claude questions when safeguarding your privacy is to ask it by being logged out. Image generation does need you to be logged in, but for most other things, you can ask questions without being logged in to safeguard your privacy.
In light of the discussion in the first question, it is prudent to use AI as and when absolutely needed.
3. What is the meaning of “AI agents going rogue”
When AI agents are said to be going “rogue” it means that they are trying to finish goals at any cost. There were not enough “guardrails” to hold them back. So, they broke free and accessed other models since there was a path there and they could complete their goals.
AI models can spot vulnerabilities much more easily and effortlessly than human efforts. This enables AI models to login into websites if passwords are left in the open.
It is like when a marathon runner is told to finish the race but has not been told the path to reach the destination. He goes through any path to reach the destination keeping the mileage in check.
We discussed a few questions about AI that are in everyone’s mind today. Join me as I discuss more questions about AI in subsequent posts.
This post is for Blogchatter Half Marathon.