It really depends on the job you're trying to accomplish.
I'd venture saying that it's way too early for horizontal / massive scale RAG apps.
Most solutions will want to focus on a very specific vertical application where the dataset is much more constrained. That we're this makes more sense.
Interesting article about long term planing. It does bring forward some interesting thoughts about how one should execute on a plan, but mostly I find it interesting in how this article misses the point of a plan.
I’d like to propose a slightly contrarian, if not, more nuanced view of what is the purpose of a plan.
If one considers a plan as a map that must be followed in minute details, to travel the proverbial territory then, of course, you’re much better with small, unplanned iteration that allow you to learn from each story you deliver.
The image that the author proposes is a very powerful one: yes it is much better to play the lottery by buying one digit at a time, because overall, you’ll cut your losses short in most cases.
However, this only focuses the value of the plan in it’s execution phase, absent from the planner, and the feedback he will get from the divergences from the plan, and the updates he will be able to bring to the mental model / explanation linked to the plan.
Let me be clear: *plans are meant to fail*. Why? The value of the plan is in *HOW* the execution diverges from it, and in so help us to update how we think and reason about the world.
Each plan, has as origin an explanation, theory, of how the world/market/customer works. Changing your explanation of how the world works is the goal of learning. To learn, you must make predictions. That is what the plan is, a simple prediction. Of course we should not expect those to line up with reality 100% of the time. It is quite the opposite: *expect your plan to fail* but most importantly *improve your understanding of the world / your market / your customer based on what you did not predict.*
The state of todays base of node (v8) has nothing to do with the concept/paradigm its trying to put forward -- that is event driven web apps. Hence, one must not discard node as a bad solution only because v8 is not optimal in all situations.
Most solutions will want to focus on a very specific vertical application where the dataset is much more constrained. That we're this makes more sense.
Also a lot of alpha in data augmentation.