TickerTain — Make Analysis Fun Again

▶ 30:15 · AI Infrastructure · KV cache inefficiency demands new training infrastructure for personalized models

now playing · AI Infrastructure

Memory & Storagetailwindscore 8/10dan biderman

Continual learning via weight updates beats RAG for enterprise AI adoption

Externalized memory (RAG/context engineering) hits scaling limits with token costs and retrieval failures, while training adapters on private data enables 100x inference reduction and impli…

Enterprise AI Adoptiontailwindscore 7/10jessy lin

Frontier labs' AGI focus leaves enterprise continual learning underserved

Frontier labs prioritize generic AGI capabilities (coding/math) with clean supervision, while enterprise needs involve ambiguous, private, conflicting preferences that require integrated re…

AI Infrastructuretailwindscore 7/10dan biderman

KV cache inefficiency demands new training infrastructure for personalized models

KV caches for long contexts consume massive HBM (80GB for a single Wikipedia article vs 100GB for entire Llama-70B weights), proving current inference architecture is bit-inefficient. Offli…

←

▶ 30:15 · AI Infrastructure · KV cache inefficiency demands new training infrastructure for personalized models

now playing · AI Infrastructure

Memory & Storagetailwindscore 8/10dan biderman

Continual learning via weight updates beats RAG for enterprise AI adoption

Externalized memory (RAG/context engineering) hits scaling limits with token costs and retrieval failures, while training adapters on private data enables 100x inference reduction and impli…

Enterprise AI Adoptiontailwindscore 7/10jessy lin

Frontier labs' AGI focus leaves enterprise continual learning underserved

Frontier labs prioritize generic AGI capabilities (coding/math) with clean supervision, while enterprise needs involve ambiguous, private, conflicting preferences that require integrated re…

AI Infrastructuretailwindscore 7/10dan biderman

KV cache inefficiency demands new training infrastructure for personalized models

KV caches for long contexts consume massive HBM (80GB for a single Wikipedia article vs 100GB for entire Llama-70B weights), proving current inference architecture is bit-inefficient. Offli…