Filipe Gracio, PhD

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

Looking forward to being a part of our upcoming Data Management for Precision Medicine event! See you all soon! Join us at: https://lnkd.in/ehXVibJr Read more on my perspectives on Data and Precision Medicine in my latest article: https://lnkd.in/eaA9THkf #DataManagement #HealthcareData #LargeDataModels #Omics #PopulationHealth #PopulationOmics #PrecisionMedicine

6

Interesting paper about the limitations of the generative AI. Actually Yann LeCun talks about this limitation as inherited problem in the existing models for long time. Knowing the limitation helps to align expectations. https://lnkd.in/dCid9B_J #generativeai #llm #artificailintelligence #ai

7

3 Comments

There is so much research going into bringing the power of LLMs into the Enterprise context. The main techniques are k-shot learning or prompt engineering, RAG (Retrieval Augmented Generation) & Model Fine tuning (not to mention many variations of these e.g. Agentic RAG, GraphRAG etc.) However, I came across an interesting take on this from IBM and they are calling it "InstructLab" - where the claim is that they are able to teach the LLM "new skills" grounded in an Enterprise context. On the surface it looks like a k-shot learning + fine turning on steroids. Approach seems to be to utilize a small "skills training dataset" in order to generate a large synthetic dataset (perhaps utilizing a specialized LLM) that is then incorporated into the LLM weights natively. This could be a very interesting concept as long as the cost of adding new skills is not very high (when compared to fine tuning). Exciting times! #genAI #enterpriseAI

40

2 Comments

Good article on practical tips to improve the #RAG pipeline - across indexing, retrieval and generation. • Indexing - level up with simple, semantic or #LLM based chunking • Retrieval - implement hybrid search, combining vector and keyword searches • Query Rewriting - optimize queries for better retrieval • Metadata - boost retrieval accuracy with rich contextual data • Autocut - filter out irrelevant information from retrieved content • Reranking - prioritize the most relevant content, enhancing retrieval precision https://lnkd.in/eJMvu3XJ

7

Really interesting read on the evolution of Data Warehouses, Data Lakes and Lake Houses and how Databricks and Snowflake are battling for the top spot. https://lnkd.in/egg4P5VG This is one of the more complete analysis I've seen on the evolution of the analytics space and it's big gorilla's. In the end, it gets into the AI discussion of big LLM's vs. small models and how SLM's will have a place in the analytics world, which I completely agree with.

9

1 Comment

While we've posted in the past on PolyBlocks performance, we now report its coverage on all the 210 PyTorch models recently used in the evaluation of the PyTorch-2 paper from ASPLOS 2024 (https://lnkd.in/gj6cZCYn). These benchmarks are publicly available and drawn from HuggingFace, TIMM, and TorchBench suites. PolyBlocks can successfully compile and execute 185 of these 210 models (push button, no source changes)! These models span various AI domains with some containing nearly 25,000 tensor operators. 90% of the models that run with PyTorch's standard runtime or compiler work with PolyBlocks! Benchmarks courtesy: https://lnkd.in/gtcTNQzj

58

This is a great first step in understanding how scaling dynamics extend beyond language to other types of sequential data. Time is a universal axis, understanding how models learn optimally over data collected over time is a crucial step in the broader AI research agenda.

26

Latest updates of dbt, one of the best transformation tools

5

Looking forward to great discussion on this critical topic! #GenerativeAI #largelanguagemodels #artificialintelligence #airevolution

19

1 Comment

Exciting news for customers: Fine-tuning for Anthropic’s Claude 3 Haiku model is now available in Amazon Bedrock—currently the only managed service to offer this capability. This means our customers can now customize Claude 3 Haiku with their own data to build GenAI applications tailored to their specific needs, while keeping their data secure. Since we launched the preview a few months ago, we've already seen some great use cases. One of our customers, Robin AI, a legal AI company, is using this feature to optimize legal workflows and help lawyers speed up contract processes. With a fine-tuned Claude 3 Haiku in Bedrock, they are helping lawyers finish contract review processes 90% faster. Will be great to see how other customers use this new capability. https://lnkd.in/g2jU8Kfv

324

5 Comments

Simple messaging, Andrew Ng knows how to simplify complicated topics, here he speaks of AI Agents ... not new but a nice breakdown of agent reasoning design patterns. Consider Reflection, Tool Use, Planning and multi agent collab can help aid solns to be more robust. Some excellent examples that are presented simply. I had the rare pleasure to work with Andrew and many of his graduate students in 2014-2015 around much smaller DNNs for ASR and IRC where we thought a billion params was a lot... how far we have come. https://lnkd.in/gK8DAQMv

25

2 Comments

Good summary

3

1 Comment

Noteworthy #opensource project of the week: #optillm is an OpenAI API compatible optimizing inference proxy which implements several state-of-the-art techniques that can improve the accuracy and performance of #LLMs. The current focus of this project is on implementing techniques that emulate "reasoning" (I really hate this term with reference to language models, but let's use it to make a larger audience clear about what optillm does) when addressing coding, logical and mathematical queries. It includes out-of-the-box implementations of algorithms from 11 popular papers, including R* and CoT with Reflection. This is basically similar to what the latest model from OpenAI does at inference time, but here you can apply different techniques also to older OpenAI models or local OS models and stay more in control. Written in #Python, optillm is easy to integrate with other tools and frameworks. Please let me know if you have a chance to use it and your feedback (and give it a star if you should find it useful). Disclaimer: I am one of the contributors to this project. Link to the official optillm's GitHub repo in the comments. #AI #GenerativeAI

14

4 Comments

https://lnkd.in/dBkJGXwN A really detailed paper comparing latest results in linear mat-mul free progress. Worth reading and if you are a Ph.D. Researcher something to consider as a great investment for your carreer

6