World Wire

Retrieval quality in an AI search product is bounded by two things: how good the embedding model is, and how cheaply you can run it across an index. This week, Perplexity Engineering team published Fast Embeddings on GPUs , an under-the-hood account of the second — the serving infrastructure behind pplx-embed and the ranking models used across Perplexity Search, Computer and the API Platform. Perplexity team states that embedding inference on the GPU side has largely converged across engines on mature Hopper and Blackwell hardware. The wins sit in the runtime and harness around the model: CUDA graph management, an async result-tracking abstraction, and a Rust request path. Two traffic patterns, one engine Perplexity frames embedding serving as two workloads. Batch embedding happens when building or re-indexing the vector database, where throughput minimizes cost. Online embedding happens at query time, where a short query must be embedded fast. Scoring sits in between: ...

The hard part of running an open-weights model locally was never the model. It was everything before it, reading VRAM specs, guessing which quantization fits, setting context length and GPU layer counts, then finding out at load time that the file is three gigabytes too large. Nous Research has collapsed that sequence into a single click inside Hermes Desktop . The new easy setup flow reads your hardware, selects a model that fits it, downloads the weights, and configures the inference runtime for you. Is it deployable? Yes. Hermes Desktop is the MIT-licensed, free build of the open-source Hermes Agent, runs on macOS 12+, Windows 10/11, and any Linux distribution, and needs no account at all for local models. What actually shipped The announcement is narrow and concrete: Hermes Desktop now sets up local models in one click, reading your hardware, choosing a model, downloading it, and configuring the runtime. The flow appears automatically on first launch and is reachabl...

This week, Adaption Labs released Invent a Dataset . The feature generates a structured, training-ready dataset from a description of the behavior you want a model to learn. You do not arrive with a seed corpus, a predefined schema, or a labeling guide. Is it deployable? Yes, with one caveat. Invent a Dataset is live now in the Adaption app and through the Python SDK and REST API . Generated rows download as JSONL, JSON, CSV, or Parquet, so the artifact is a portable file you own and can train on anywhere. Generation itself runs on Adaption’s hosted platform and consumes credits. No self-hosted generation path is documented. The problem being targeted Most dataset workflows begin with data that already exists. Teams then spend weeks labeling, filtering, and reshaping it to approximate the target task. Adaption’s argument is that this caps model quality at how closely the available data matches the intended behavior. For proprietary and specialized tasks, the r...

Multi-agent workflows have changed the shape of local inference. A lead agent decomposes a task and spawns subagents. What looked like one user request becomes dozens of independent model calls. Pointed at a single local engine, those calls compete for the same execution slots. The queue grows while a workstation, laptop, or DGX Spark on the same network sits idle. NVIDIA Personal AI Router (PAIR) targets exactly that bottleneck. Announced this week, PAIR is a virtual inference router. It discovers compatible machines on a home network and schedules independent inference requests across them. It is not a new inference engine. Ollama or LM Studio still executes the model on whichever node PAIR selects. Is it deployable? Yes. PAIR ships today as a public beta (v0.1.1) with signed installers for Windows, macOS, and Linux, and the full source is on GitHub under Apache 2.0. It runs entirely on the local network, with internet needed only to download models. No new API ...