How I keep up with AI progress
15 points by atharva
15 points by atharva
Another strategy is to just wait. The really useful ideas will bubble to the surface in time, and you will have saved the mental bandwidth of keeping up with every little twist and turn along the way.
I think this is a good strategy. By its nature there isn’t the risk of being “left behind” here due to lack of skill. You can get to the forefront of current LLM stuff (agents and/or local LLMs with formal grammar token restriction) in just a few hours. Spend the time reading Knuth books instead, in his spirit of wanting to get to the bottom of things instead of staying on top of things.
You can get to the forefront of current LLM stuff (agents and/or local LLMs with formal grammar token restriction) in just a few hours.
I disagree with this. It takes a serious time investment to be at the “forefront” in terms of skill using these tools. But if you just mean “get up to speed enough to be decently productive” that might be true.
Spend the time reading Knuth books instead
In my article, I state that the work of keeping up is about 15-20 minutes per day. This is not enough time to do any justice to Knuth’s books ;)
There is value in staying on the cutting edge for those building with AI. Given that the capabilities are evolving rapidly, it helps to know and anticipate where the technology is at in order to make an informed choice to build a good product (this is true of any technology, not just AI!). I’m also against neomania, and a good builder must get an intuition for what is worth paying attention to.
I address the point about mental bandwidth at the end. A big motivation for writing this is to carve out a path for those who want to actually understand the fundamentals of the technology they are working with without needing to be terminally online in unskillful ways.
Given that the capabilities are evolving rapidly,
Could you elaborate?
Could you elaborate?
Not your parent, but from a similar comment I made elsewhere:
For example, you could point to any of these things in the last ~8 months as being significant changes:
All or any of these could have made for a big or small impact. For example, I’m big on agentic tools, skeptical of MCPs, and don’t think we yet understand subagents. That’s different from those who, for example, think MCPs are the future.
I find https://old.reddit.com/r/LocalLLaMA/ serves as a pretty good aggregator for new model releases and the like.