Google's EmbeddingGemma 2 drew attention on Hacker News, where Simon Willison argued open-weight embedding models under Apache 2.0 are the right choice for applications storing millions of vectors. Proprietary hosted embeddings create lock-in: when vendors deprecate a model, users must pay to re-embed their stored corpora.
A demo photo gallery that lets users search images by text description, with all embedding and retrieval running in the browser on the GPU. It uses Google's newly released EmbeddingGemma 2 (text, image, audio, and video in one vector space) ported to ruNNtime, a TypeScript WebGPU inference library. It solves on-device, private semantic photo search without a server, for developers and privacy-conscious users.
A demo photo gallery that lets users search images by text description, with all embedding and retrieval running in the browser on the GPU. It uses Google's newly released EmbeddingGemma 2 (text, image, audio, and video in one vector space) ported to ruNNtime, a TypeScript WebGPU inference library. It solves on-device, private semantic photo search without a server, for developers and privacy-conscious users.
ZeroEntropy's API will shut down on September 4 following its acquisition by Notion, forcing production users of its embedding and reranking models to migrate. The models are open-sourced under Apache 2.0, but no major inference providers currently host them, leaving users to consider self-hosting or switching models.
ZeroEntropy's API will shut down on September 4 following its acquisition by Notion, forcing production users of its embedding and reranking models to migrate. The models are open-sourced under Apache 2.0, but no major inference providers currently host them, leaving users to consider self-hosting or switching models.
A Reddit user compared per-token cloud API pricing for embedding models in a RAG indexing project and found OpenAI's text-embedding-3-small ($0.02/1M tokens) cheaper than open-model APIs like Snowflake Arctic Embed L v2 ($0.038) and Qwen3 Embedding 4B ($0.13). The post argues open models only save money when self-hosted on a GPU at very high utilization, prompting discussion on cost-versus-quality tradeoffs.
A Reddit user compared per-token cloud API pricing for embedding models in a RAG indexing project and found OpenAI's text-embedding-3-small ($0.02/1M tokens) cheaper than open-model APIs like Snowflake Arctic Embed L v2 ($0.038) and Qwen3 Embedding 4B ($0.13). The post argues open models only save money when self-hosted on a GPU at very high utilization, prompting discussion on cost-versus-quality tradeoffs.
A Reddit user criticizes AI companies for publicly detailing performance improvements (e.g., Logan Kilpatrick on Gemini 3.6 Flash) while rarely releasing comparable notes for safety filter changes, guardrail updates, or refusal behavior shifts. The post highlights a perceived asymmetry in transparency between capability and safety updates.
A Reddit user criticizes AI companies for publicly detailing performance improvements (e.g., Logan Kilpatrick on Gemini 3.6 Flash) while rarely releasing comparable notes for safety filter changes, guardrail updates, or refusal behavior shifts. The post highlights a perceived asymmetry in transparency between capability and safety updates.
Simon Willison published a hands-on teardown of OpenAI's ChatGPT Work, finding it is actually two products: Work Cloud on chatgpt.com and Work Local, a separate desktop app. Both are gated to $20/month-and-up subscribers. Willison warns the product combines private data, tool access, and untrusted web content — the 'lethal trifecta' for prompt-injection risk.
Simon Willison published a hands-on teardown of OpenAI's ChatGPT Work, finding it is actually two products: Work Cloud on chatgpt.com and Work Local, a separate desktop app. Both are gated to $20/month-and-up subscribers. Willison warns the product combines private data, tool access, and untrusted web content — the 'lethal trifecta' for prompt-injection risk.