ecal
@0xeca1
llms, neovim, nixos
rag feels antiquated, there should be systems that dynamically swap context and continue generation according to what’s relevant to the next (n) token(s) would be slower ofc but much smarter
I really like Chollet’s take on pattern recognition vs intelligence (as presented on mlst podcast). sure scaling alone has made great advancements so far, and i bet there’s a lot more to come, but even the most complex outputs are still undeniably derived from interpolation…
There'll be 1-2 years where we'll be like "Aidan was right, AGI is here, but… it's meh?" as models burn through the hardest benchmarks yet this is still mainly memoization. IMO shattered, but no Fields in sight. AlphaFold 5: cancer survivability stagnant. And Chollet triumphant.