Hello, I’m Andrew Briz! I lead an engineering team in the POLITICO Newsroom that develops tools, prototypes, and infrastructure for news creation and distribution. This quarter, I joined Axel Springer’s inaugural Generative AI fellowship cohort. I've paused my day job to research, document, and develop AI use-cases and partnerships in journalism. I’ll occasionally blog about my findings, news, and developments. If you’re interested, follow me on all the socials.
One of the significant AI stories this week was the launch of Google’s open-source LLM Gemma. As an open-source model, it can be customized, fine-tuned, and deployed on infrastructure that meets your needs. But what makes this release interesting is its impact on the ongoing debate between open-source and closed-source. Until now, the major players have been divided, with OpenAI, Anthropic, and (previously) Google offering closed models with limited API access, while Facebook, Stability, and others have taken the opposite approach. However, with the introduction of Gemma, Google is taking a hybrid model. This is not the first time they have adopted this approach - their open-source Chromium browser serves as the foundation for Microsoft Edge and Brave (which happens to be my preferred browser). Clearly, this dual approach is effective for them - they sell access to casual users while still benefiting from collaboration with the broader LLM community for improvements, which can then be integrated into their closed models.
Ultimately, this is good for journalism. Reporters are often protective about their notes and sensitive materials. As a developer creating tools for them, using third-party services means I can’t guarantee the safety of that content. Open-source models allow us to leverage these models in environments we have complete control over — even air-gapped ones.
The unanswered question is, “Is Gemma any good?” Unfortunately, I haven’t had enough time to answer that question, but you can! If you’re interested in trying Gemma (or any other open model), I was shown a cool tool called Ollama this week. It’s easy to download on any OS, and you can use it to run LLMs locally with a single command! If the command line scares you, this is a simple first tool to start your journey.
MindsDB is an open source technology and tech company that enables data teams to interact with AI using an SQL-like syntax. I seldom write SQL, but I find its conceptual premise fascinating. Using an abstraction layer to lower the barriers of leveraging Generative AI and other models for a specific community is a key strategy for adoption, and MindsDB does that for data crunching.
The code ends up looking something like
SELECT question, answer
FROM openai_model
WHERE question = 'Where is Stockholm located?';
Which looks like a more complicated way to ask ChatGPT a question, but imagine you have a full table of thousands of cities. You could easily JOIN that table and have an LLM answer the question for each one. If you’re familiar with SQL, you can read more about this example in their docs.