Thank you for such a great article. The level of detail is impressive—I even felt it was close to a survey paper. I liked it so much that I shared it with my colleagues.
Seems like the other path forward is split your context into parallel queries and recombine answers. Considering that (prompt-context-query) sandwich repeats prompt and query whole, but you can split many kinds of context which have no inherent serial dependencies, you might convert a 1MT problem into 8 150kT problems that can be merged with orchestration. Agents orchestrate extensively already, perhaps an economical path forward.
Thank you for such a great article. The level of detail is impressive—I even felt it was close to a survey paper. I liked it so much that I shared it with my colleagues.
Thank you for sharing.
Seems like the other path forward is split your context into parallel queries and recombine answers. Considering that (prompt-context-query) sandwich repeats prompt and query whole, but you can split many kinds of context which have no inherent serial dependencies, you might convert a 1MT problem into 8 150kT problems that can be merged with orchestration. Agents orchestrate extensively already, perhaps an economical path forward.