💡 What are the nerds up to?
➜ Amazon developed Diffuse to Choose, an improved "Virtual Try-All" to overlay any product you want onto any image you want (see how that new bed would look in your room, or how a sweater would look on your back) – something that could, over time, become standard customer experience in retail. - GitHub
➜ Brush up on the many different ways to attack LLMs, to see how bad actors might want to exploit your LLM-based solutions. - PortSwigger
➜ “What used to be an effective protection can now be solved in a few hundred lines of Python. In what other domains is this also the case?” – asked cybersec researcher Clint Gibler regarding this article about using multi-modal LLMs to easily break CAPTCHAs. - LinkedIn
➜ Ollama helps you get up and running with LLMs locally – they’ve now added Python and JavaScript libraries. - GitHub, Ollama
➜ “Clay is like ChatGPT for Earth - a platform and community with a generative AI model at its core.” Clay provides a foundational model of Earth, which uses a Vision Transformer architecture adapted to understand geospatial and temporal relations on Earth Observation data, and you can experiment with it to:
–”Generate semantic embeddings for any location and time.
–Fine-tune the model for downstream tasks such as classification, regression, and generative tasks.
–Use the model as a backbone for other models.”
- Clay
➜ Open-source fans just got a nice treat – FireLlaVa is the first LLaVa (Large Language and Vision Assistant) model with a commercially permissive OSS license, which hackers can use to put together a model that recognizes photos and images just like GPT-4V. - Fireworks
➜ The race to improve the Transformer architecture is going strong – Eagle 7B is a small model based on the RWKV-v5 architecture (up to100x lower inference cost than conventional Transformers with the Attention mechanism), ranking as the world’s greenest 7B model, it’s close to Llama 2 and Mistral in English evaluations. - RWKV
