Exowatt built a serverless fine-tuning pipeline on SageMaker AI that beat frontier models on two narrow energy-infrastructure tasks, with a 1.7B model running offline on CPU and a 14B model serving through Bedrock
AspireAI community · Amazon Web Services
AWS published an October 27, 2026 case study on Exowatt, an energy-management-software company whose sites raise constant alarms. A frontier model could not work for three stated reasons: the task is proprietary, it must run at the edge with no GPU and often no internet, and per-decision frontier cost does not scale across thousands of sites. Exowatt's first task — edge alarm triage — reached 0.911 macro-F1 on 320 test cases with a fine-tuned Qwen3-1.7B, ahead of every frontier baseline, running quantized to 8 bits (about 1.71 GiB) in a 6 GiB llama.cpp pod on the site's existing CPU server; training took roughly 32 minutes for about $20–$25. A prompted Qwen3-8B had plateaued at 0.735 macro-F1, and 13 fine-tuned 8B candidates across 12 dataset versions never beat it — the smaller model with more varied data won. The second task, cloud energy-market analysis, used fine-tuned Qwen3-14B: paraphrase diversity lifted macro-F1 from 0.950 to 0.975 (base model 0.415), and the model imports into Bedrock Custom Model Import at about $6.86 per hour while active. The honest limitation: this is a vendor case study with AWS's own framing, the macro-F1 figures come from a single company's curated 2,522-example and 86,281-row training sets, and no frontier model was fine-tuned for comparison — only prompted.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.