Published:
I gave Gemini 4 Argon the same web form three times: through Antigravity’s browser agent, through Playwright MCP, and through macOS Apple Events. All three got it right, but they went about it very differently.
Published:
Extract every layer’s activations from Gemma 3 with JAX and Keras on a free TPU, then watch it answer in English before Arabic, and steer it into Arabic with one vector.
Published:
The chip built from scratch for AI, and why it might be your next move.
Published:
Applying the method from “Refusal in Language Models Is Mediated by a Single Direction” to Google’s Gemma 3 family, with experiments accelerated on TPUs.
Published:
How the Model Context Protocol, Google’s Agent Development Kit and the Agent2Agent protocol fit together, with a currency agent built step by step.
Published:
From an AI Product Management Course
A journey of understanding, evaluating, and gaining deeper awareness of what truly makes a difference in the world of products.
Published:
Reflections from an AI Product Management Course
Published:
Published:
Image source [3]
Published:
Continuing our educational series focusing on Arabic language handling with large language models, in this part, we will explore how to fine-tune the Gemma2-9b model on an Arabic dataset using the Keras library, Keras_nlp, and LoRA technique. We will cover how to set up the environment, load the model, make necessary modifications, and train the model using model parallelism to distribute model parameters across multiple accelerators.
Published:
As part of our tutorial series focusing on handling Arabic with large language models, we will continue in the second part by exploring specialized methods for fine-tuning the Gemma-2b model on an Arabic dataset to enhance its performance. We will use the Transformers library and qLoRA (Quantized Low-Rank Adaptation) technique to reduce memory usage.
Published:
When working on a project that requires handling the Arabic language, you might wonder whether to use Retrieval-Augmented Generation (RAG) or to fine-tune a model with a new Arabic dataset. In this series of tutorials with two parts, we will explore both options: using RAG and fine-tuning a model with Arabic data, specifically Wikipedia. Throughout the project, we will focus on open-source models, utilizing Gemma 2 Instruct and an open-source embeddings model. We will also leverage the LangChain framework to streamline the process of building the RAG and fine-tuning the model. Let’s dive into the practical implementation.
Published:
Filling out forms can be boring and time-consuming. This often leads to user frustration and incomplete submissions. However, conversational AI, like the Gemini 1.5 Pro language model, is changing how we interact with forms.
Published:
The world of data science can be daunting for newcomers, filled with complex terminology and intricate concepts. But what if you had an AI assistant by your side, ready to explain these concepts in simple terms and guide you through the learning process? This is where the power of Retrieval Augmented Generation (RAG) comes into play. In this blog post, we’ll embark on a hands-on journey, building an AI-powered explainer for data science concepts using the RAG model and the Gemma language model.