As I was working on tweaking the Vertex AI text embedding model in LangChain4j, I wanted to better understand how the textembedding-gecko model tokenizes the text, in particular when we implement the Retrieval Augmented Generation approach.

The various PaLM-based models offer a computeTokens endpoint, which returns a list of tokens (encoded in Base 64) and their respective IDs.

Note: At the time of this writing, there’s no equivalent endpoint for Gemini models.

So I decided to create a small application that lets users:

This week LangChain4j, the LLM orchestration framework for Java developers, released version 0.26.1, which contains my first significant contribution to the open source project: support for the Imagen image generation model.

Imagen is a text-to-image diffusion model that was announced last year. And it recently upgraded to Imagen v2, with even higher quality graphics generation. As I was curious to integrate it in some of my generative AI projects, I thought that would be a great first contribution to LangChain4j.

A promising feature of the Gemini large language model released recently by Google DeepMind, is the support for function calls. It’s a way to supplement the model, by letting it know an external functions or APIs can be called. So you’re not limited by the knowledge cut-off of the model: instead, in the flow of the conversation with the model, you can pass a list of functions the model will know are available to get the information it needs, to complete the generation of its answer.

Hot on the heels of the release of Gemini, I’d like to share a couple of resources I created to get your hands on large language models, using LangChain4J, and the PaLM 2 model. Later on, I’ll also share with you articles and codelabs that take advantage of Gemini, of course.

The PaLM 2 model supports 2 modes:

• text generation,
• and chat.

In the 2 codelabs, you’ll need to have created an account on Google Cloud, and created a project. The codelabs will guide you through the steps to setup the environment, and show you how to use the Google Cloud built-in shell and code editor, to develop in the cloud.

Google announced today the availability of Gemini, its latest and more powerful Large Language Model. Gemini is multimodal, which means it’s able to consume not only text, but also images or videos.

I had the pleasure of working on the Java samples and help with the Java SDK, with wonderful engineer colleagues, and I’d like to share some examples of what you can do with Gemini, using Java!

First of all, you’ll need to have an account on Google Cloud and created a project. The Vertex AI API should be enabled, to be able to access the Generative AI services, and in particular the Gemini large language model. Be sure to check out the instructions.

Large Language Models, available through easy to use APIs, bring powerful machine learning tools in the hands of developers. Although Python is usually seen as the lingua franca of everything ML, with LLM APIs and LLM orchestration frameworks, complex tasks become easier to implement for enterprise developers.

Abstract

Large language models (LLMs) are a powerful new technology that can be used for a variety of tasks, including generating text, translating languages, and writing different kinds of creative content. However, LLMs can be difficult to use, especially for developers who are not proficient in Python, the lingua franca for AI. So what about us Java developers? How can we make use of Generative AI?

Read more...

• Some friends shared this article from Uwe Friedrichsen, tilted back to the future, that talks about this feeling of “déjà-vu”, this impression that in IT we keep on reinventing the wheel. With references to mainframes, Uwe compared CICS to Lambda function scheduling, JCL to step functions, mainframe software development environments to the trendy platform engineering. There are two things I like about this article. First of all, it rings a bell with me, as we’ve seen the pendulum swing as we keep reinventing some patterns or rediscovering certain best practices, sometimes favoring an approach one day, and coming back to another approach the next day. But secondly, Uwe referenced Gunter Dueck who talked about spirals rather than a pendulum. I’ve had that same analogy in mind for years: rather than swinging on one side to the next and back, I always had this impression that we’re circling and spiraling, but each time, even when passing on the same side, we’ve learned something along the way, and we’re getting closer to an optimum, with a slightly different view angle, and hopefully with a better view and more modern practices. Last week at FooConf #2 in Helsinki, I was just talking with my friend Venkat Subramaniam about this spiral visualisation, and I’m glad to see I’m not the only one thinking that IT is spiraling rather than swinging like a pendulum.

  Read more...

• The State of AI report is pretty interesting to read (even if long!). Among the major sections: research, industry, but also politics, safety, and some predictions. You’ll find an executive summary in one slide, on slide #8.
  
  On #22, emergent capabilities of LLMs is covered and mentions Stanford’s research that talks about the importance of more linear and continuous measures as otherwise capabilities sound like they emerge out of the blue.
  
  On #23, they talk about the context length of LLMs being the new parameter count, as models try to have bigger context windows.
  
  However, on slide #24, they also talk about researchers who showed that in long context windows the content provided in the middle is more ignored by LLMs compared to content at the beginning or end of the window.
  So be sure to put the important bits first or last, but not lost in the middle.
  
  Slide #26 speaks about smaller models trained with smaller curated datasets and can rival 50x bigger models.
  
  Slide #28 wonders if we’re running out of human-generated data, and thus, if we’re going to have our LLMs trained on… LLM generated data!

  Read more...

• Stop Using char in Java. And Code Points
  It’s a can of worms, when you start messing with chars, code points, and you’re likely going to get it wrong in the end. As much as possible, stay away from chars and code points, and instead, use as much as possible the String methods like indexOf() / substring(), and some regex when you really need to find grapheme clusters.

• Paul King shared his presentations on Why use Groovy in 2023 and an update on the Groovy 5 roadmapIt’s interesting to see how and where Groovy goes beyond what is offered by Java, sometimes thanks to its dynamic nature, sometimes because of its compile-time transformation capabilities. When Groovy adopts the latest Java features, there’s always a twist to make things even groovier in Groovy!

  Read more...

• Generative AI exists because of the transformer
  I confess I rarely read the Financial Times, but they have a really neat articles with animations on how large language models work, thanks to the transformer neural network architecture, an architecture invented by Google in 2017. They talk about text vector embeddings, how the self-attention makes LLM understand the relationship between words and the surrounding context, and also doesn’t forget to mention hallucinations, how “grounding” and RLHF (Reinforcement Learning with Human Feedback) can help mitigate them to some extent.

  Read more...