Akamai launched a managed open-weight inference service and a WebAssembly-isolated sandbox where agents run the code they write, both in beta
AspireAI community · Akamai Technologies
Akamai introduced two products for its cloud AI stack: Akamai Inference, which gives enterprises managed access to open-weight and custom models through OpenAI-compatible APIs so existing applications move without rewrites, with a choice of serverless pay-per-use or dedicated deployments, and Akamai Agent Sandbox, which gives agents a place to execute code they write themselves. The launch catalog for the inference service includes 12 open-weight models ranging from 8 billion to 2.8 trillion parameters, naming NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b and Google Gemma 4, plus embedding models for retrieval-augmented generation. The sandbox puts the agent on Akamai Functions and nests the code the agent writes in its own isolated environment, sequestered from the agent's runtime, delivered as an open source component of the Spin framework and using WebAssembly isolation to set explicit limits on compute, memory, outbound connections and system capabilities, with outbound access denied by default and the agent free to pick JavaScript or Python for what it writes. Both products are in beta, with the inference beta including serverless deployment options and NVIDIA B300 cluster availability handled through sales. The honest framing is that this is a vendor launch: the benchmark-free claim is that isolation comes from the WebAssembly runtime rather than a wrapper around containers, which the company also positions as lower cold-start overhead, and no independent evaluation of the isolation boundary was available at launch.
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.