Can we change an AI model’s architecture without retraining it from scratch? - Amazon Science

kept by eddie oct 13

Priming lets researchers reuse pretrained Transformer weights to build efficient hybrid state‑space models at <1% training cost, achieving up to 2× throughput with comparable accuracy.

How LLM Inference Works
LLM inference works by tokenizing input, computing embeddings through transformer layers, then generating tokens autoregressively with KV caching and quantization optimizations.
also on LLaMA, #machine-learning, #developer-tools
Mistral 3 family of models released
Mistral released three small dense models and a large MoE model under Apache 2.0, claiming best performance-to-cost in their size categories.
also on Mistral, #machine-learning
Just say no to OpenAI and Anthropic
The author urges rejecting OpenAI and Anthropic due to privacy risks and oligarchic control, recommending open-weight LLMs served via private enclave providers like Tinfoil or Maple.
also on Mistral, #ai-coding, #developer-tools
Corporate America Is Getting Hooked on Open-Source A.I.
AT&T and other corporations are shifting significant portions of their AI workloads to free, open‑source models to cut costs and reduce vendor dependence.
also on LLaMA, #developer-tools
the broken economics of databases - by almog gavra
Database vendors charge high prices but earn slim profits because they must recover large fixed costs, constantly innovate to avoid commoditization, and compete against cloud providers who own the underlying infrastructure.
also on Amazon Web Services, #developer-tools