Gus Martins

From Experimentation to Enterprise Deployment: Operationalizing Generative AI I recently had a fascinating conversation with Shane McAllister and Stanimira Vlaeva, developer advocates at #MongoDB, about the hurdles of taking #generative AI from ideas and hackathons to real-world enterprise use. A key takeaway? We need to move from a model-centric mindset to a platform-centric approach. This strategic shift streamlines the complexities of generative AI development and deployment (think MLOps, model evaluation, prompt management) and ensures crucial elements like data strategy and responsible AI are baked in. Huge thanks to #StackOverflow team for hosting this insightful discussion on such a relevant topic! #generativeAI #MLOps #responsibleAI #datascience #enterpriseAI #cloudAI #machinelearning #deeplearning ##GoogleVertexAI  #Google #gcp ##GoogleCloud #lifeatgoogle #googledevelopers

12

Yesterday we had a fun community meetup about vision tasks in Argilla. We shared a load of references, so I thought I would be cool to drop it all here: In the meeting, David Berenstein and I shared some notebooks that demonstrate what's possible with vision models and human feedback. And in the discussion we explored what's possible with Argilla and the vision features we're releasing and building. 📰 Posts: This is a straightforward guide on how vision LLMs work with some snippets to get started: https://lnkd.in/ec5HX-zS. Overview of inference with CLIP: https://lnkd.in/ewA-ryQz 👩‍🔬 Papers: OCR 2.0 is useful review of the current state of OCR and a new more efficient approach: https://lnkd.in/ewH7jSPr The DocVQA dataset contains documents and questions about the document: https://lnkd.in/eVhTPHtv ColPali is an efficient model for visual document retrieval based on Pali Gemma using a Colbert approach: https://lnkd.in/e9F_QD3F 👷 Notebooks: Image Search Finetune Demo - With Argilla ImageField: https://lnkd.in/ez_Qdjjh Document RAG with ColPali and Qwen 2 VL: https://lnkd.in/edtTFSUj Image Preferences datasets in Argilla David Berenstein :https://lnkd.in/eDUdx_jc

32

Excited to be at FACCT 2024 - right next door to the gorgeous beach in Rio! If you are, also, here - let's grab a coffee. Many interesting conversations and talks about Fairness and Transparency in AI - Will be curating and sharing some of my favorite papers later, but a couple of thematic thoughts based on what I've seen so far: ⚖  Fairness: Unsurprisingly, metrics driven fairness evaluation is very limited and flawed- there is a large push for participatory processes (e.g. involving people who will be affected by algorithms in development and following up with them). Not something we are seeing in proposed legal frameworks; remains to be seen how this will or could be implemented at scale. 👩‍⚖️ EU AI Act: Many active concerns and discussions about how audits/evaluations will look like in the context of the Act. No perfect answers yet, but, I am excited to see where this research and conversation leads. Interesting additional flag is about how EFFECTIVE Human Oversight required by Article 14 is, also, not a well-defined concept. 📚 Dataset Transparency: I'm seeing less of a push for Open Training Data, than I expected. One rightful concern is around privacy, but a subtle additional point that came up in discussion is around how its important to know how to test the data not just have a blind dump of it. More insights to come - let me know what aspects you are interested in.

16