russlan

5 ms·

1. 1.▲
   by russlan 1mo ago

   The useful contribution is not zero token cost; it is removing generative rewriting from memory. Preserving original traces avoids a subtle auditability failure: once an LLM compresses an interaction, retrieval is grounded in the summary&#x
2. 2.▲
   What 4,523 AI investment forecasts revealed about model disagreement(ipulseai.substack.com)

   3 points by russlan 2mo ago|0 comments
3. 3.▲
   by russlan 2mo ago

   imo model switching only works if you normalize outputs and keep model specific weirdness behind well designed adapters
4. 4.▲
   by russlan 2mo ago

   generic model benchmarks are less useful than workflow benchmarks. Nowadays it's all about workflows , think prompt injections, tool permissions is a big one, logs and lately what anthropic claims is the model distillation risks