Just in case you missed it, ChatGPT got friendlier again this week, but on to the other top bits of news:
- In a first, Google has released data on how much energy an AI prompt uses. Google released some interesting data this week on the power consumption of the median Gemini AI query. At 0.0024 KW/H. Google also released the same data for search queries: just 0.0003 KW/H. So AI queries are no more power hungry than search (Correction -> I previously said 800x!). I actually thought the gap would be way higher. I'd also expect that the power consumption of the largest queries is many orders of magnitude higher than the median number (the average also seems likely to be higher).
- Agents.MD. OpenAI and others have collaborated on creating a common place to place code instructions to AI coding agents. They are promoting the use of an "Agents.MD" markdown file in the root directory of code repositories to add context information for AI systems writing code. The idea is to have a unified place for extra build and other instructions. The files should hopefully help coding agents from many companies, not just OpenAI. This seems like the beginning of documenting code for AI. Slightly wild.
- Anthropic takes aim at OpenAI, offers Claude to ‘all three branches of government’ for $1. Both OpenAI and Anthropic now offer AI to the US government for cut-price rates. If you thought that government might be slow to adopt... possibly think again. In the long run, no doubt normal commercial rates will return, but the land grab is on right now.
- The last days of brainstorming. The economist has a fun piece predicting the end of brainstorming, as people simply use ChatGPT to generate options and then pick one. It seems to me that ChatGPT and other models could be a huge accelerator to brainstorming, but not in the "ask it for the answer" way. Instead, use it not just to generate ideas, but also to do background checks and to create a panel of customers and users: what would they think? What other brands might they confuse a name with, etc?
- Small Language Models are the Future of Agentic AI (NVIDIA Research). In this research, an NVIDIA and Georgia Tech team argue that small language models (defined as <10bn parameters and able to run on a consumer device) will likely be much more effective at many specific, well-defined tasks than LLMs. Not only does their size make them more economical to run, but they are better suited to operating narrowly on a topic at hand, where an LLM may end up being distracted into activating irrelevant outputs. I definitely subscribe to this view; we're really just at the beginning of understanding how we build reliable AI systems, and smaller models will be a big part of that journey. (The full paper is here.)
Wishing you a wonderful weekend.
CORRECTIONS:
- I initially quoted the Google Gemini error consumption at 0.24 KWH, but it is actually 0.00024.
