InBetter MLbyJaideep Ray·Oct 3Should models output logits or probabilities ?A classification model does not always need to return probabilities. Use logits for loss calculation, score changes, and some ranking…
InBetter MLbyJaideep Ray·Jul 25What do Self-Improving Agents learn from their mistakes ?Darwin Gödel Machines
InBetter MLbyJaideep Ray·Jul 17Agent Compute: Think Longer, Branch Wider, or Verify Harder ?where should an agent spend its next unit of compute ?
InBetter MLbyJaideep Ray·Jul 8Language as the Gradient: From text generation to autonomous executionBuilding General-Purpose Problem SolversA response icon1A response icon1
InBetter MLbyJaideep Ray·Jul 6The Retrieval Divergence Problem: Why Rank Fusion Matters More in the LLM Era ?1. The Widening Gap in retrievers
InBetter MLbyJaideep Ray·May 27AlphaEvolve: a reference architecture for turning LLMs into algorithm-discovery enginesAlphaEvolve is best understood as an agent loop.A response icon1A response icon1
InBetter MLbyJaideep Ray·May 3Differentiable Top-k RoutingMost large-scale ML systems make a hard top-k decision: scores → pick k winners → discard the rest. Hard top-k breaks gradient flow at…
InBetter MLbyJaideep Ray·May 3How Big Is an LLM? Count the Facts It RemembersFrontier labs rarely disclose parameter counts. As a result, practitioners rely on indirect proxies such as API latency, pricing, and…
InBetter MLbyJaideep Ray·Apr 12Over optimization & Reward hacking in AutoresearchAutoresearch failures
