Frontiers Capital

📢 Announcing a new generation of natively multilingual Voyage rerankers: rerank-2 and rerank-2-lite! Adding rerank-2 and rerank-2-lite on top of @OpenAI’s latest embedding model (v3 large) improves accuracy by 13.89% and 11.86%, 2.3x and 1.7x the improvement attained by the latest @cohere reranker (English v3), respectively. Voyage rerankers work exceptionally well with Voyage embedding models - in the same benchmark, rerank-2 is the only reranker to improve atop voyage-multilingual-2. You can get a full list of the embedding models we provide at https://lnkd.in/gb6a3i2i For more details, check out our latest blog post: https://lnkd.in/g78DP787 Start building with @VoyageAI today - the first 200M tokens are on us! We’d also love to support academic retrieval research and benchmarking. Please email us at contact@voyageai.com for more free tokens.

CSM will be in Japan to meet some of our partners in October. Please DM or message me at tk@csm.ai if you want to meet. Please share this with any of your relevant colleagues or friends.

Follow our portfolio company Sentient Foundation

Proud of our founder Raquel Urtasun !

#OpenSource has been the driving force of progress in Gen AI lately. Quite deservingly, AI at Meta's #Llama3.1 has been getting great attention from the community. While people analyze and digest the detailed 92 page report, there is a new release for the community that also needs some community love: Mint-1T is a new multi-modal dataset - largest dataset ever released. 10X more data than any previous open datasets. Multi-modal is the direction. Unix vs Linux history unfolding - albeit in AI world! Source: arxiv.org/pdf/2406.11271 Authors: Anas Awadalla Le(Tycho) Xue Oscar Lo Manli Shu Hannah Lee Etash Guha Matt Jordan Sheng S. Mohamed Awadalla Silvio Savarese Caiming Xiong, Ran X., Yejin Choi, Ludwig Schmidt

Our portfolio company Sentient featured on Bankless

In just few days after release, Llama-3 models have been downloaded over 1.2 million times. And there are over 600 derivate models of it on Hugging Face now! And for people who don’t get the value of #opensource and how AI at Meta gains from releasing models openly: The open source community has extended Llama 3 8B's context from 8k to nearly 100k tokens by combining PoSE, continued pre-training, and RoPE scaling methods!

How wonderful is open source? Well, somebody from somewhere in the world picks up something, massively improves it and releases it for somebody else to pick it up. Magic happens with the beautiful force of compounding! Cool new example: A developer (named Justine) picks up Llamafile project and optimizes with new matmul kernels and makes Llamafile run ridiculously faster on lower end compute (literally, toaster level!) How good it gets: 30%-500% faster inference at FP16!! Why is it important: in the world where AI (and its narrative) is controlled by big companies with massive data centers or closed APIs, it’s refreshing to see a path towards Decentralized AI. https://lnkd.in/gKjS9d4t