I turned your average chat bot experience into The Sims 3D experience - your interaction with the LLM expands into spatial domain to map the inputs and outputs as stats showing the self evaluating parameters of the LLM - the basic setup is developed manually then the complex logic is further advanced through manual looping - theres an exportable textual report document referencing important parts of the architecture and the code. Fully vibe coded w @antigravity soon available on github and via vercel. To support: CA 0xc692d990BDd836e87759A9c67476e5f1a656BB07 $LLMSIM via @clanker_world on @base Cc @jessepollak @brian_armstrong @farcaster_xyz @baseposting @0xDeployer

CURIE introduced custom evals like LLMSim and LMScore to grade nuanced outputs (like equations, summaries, YAML, code). Even the best models (Claude 3, Gemini, GPT-4) scored just ~32%. Proteins? Total fail. LLMs can read papers — solving them is another matter. #LLMbenchmarks #ArtificialInteligence #Google