Jagane Sundar

Technologist, AI and LLM Developer

[Jagane Sundar @jagane]

Jev is a useful new LLM use case, it brings efficiencies to inference https://t.co/KyTgpnp97h

[Jagane Sundar @jagane]

The Data agent OpenAI is launching Thursday inside ChatGPT Work, OpenAI's competitor to Anthropic's Cowork works by combining all the various contexts in their state, as opposed to generating a net new context layer itself. I believe this approach is superior to Databricks Adaptive Instructed Retriever (improved RAG?). My own first hand experience with my agentic Gmail chat and search indicates this as well. #databricks #agentic #bi

[Jagane Sundar @jagane]

Your point about spec driven development and being able to express intent are well taken, @GamielGran . LLMs are weak at making good architectural choices - especially ones where there are subtle second order and third order effects. It is possible that advancements in Reasoning ability and oodles more tokens will improve this..

[Jagane Sundar @jagane]

@GamielGran Thanks, Gamiel. At the moment, I am treating this simply as an experiment to get accurate information on models, application designs, agentic clients etc. I consider this as a proxy for general Business Intelligence applications.

[Jagane Sundar @jagane]

https://t.co/Rfa1XVt2PZ

[Jagane Sundar @jagane]

@Mascobot @a16z Nice, @mascobot - at OXMIQ, we did a similar project. We took it one step further - we overcame the 2 PCIe GPU limitation enforced by NVIDIA NCCL so that you can use all 384 GB to load a single model - https://t.co/4Nye1KnoOG

[Jagane Sundar @jagane]

Four 96GB RTX6000 Blackwell GPUs in a workstation. 384GB of memory, now this is serious AI horsepower! https://t.co/4Nye1KnoOG #blackwell #rtx6000 #deepseek32 #opencode #aider #oxmiq #nvidia

[Jagane Sundar @jagane]

Google Vertex AI vs Google Gemini API

Building a RAG Chat application using Google’s Gemini model. Two ways - Generative AI on Google Vertex AI (https://t.co/ytnPnMjin1...) and Gemini API (https://t.co/5rcDhHc5E5...)
Vertex AI API better for production - https://t.co/nADuI06v3C

[Jagane Sundar @jagane]

OpenAI O1 Model Token Count Explodes when Reasoning
https://t.co/P99mTC4m2g has a free course on Reasoning with o1 presented by Colin Jarvis that is worthwhile. User tokens=15, Reasoning tokens=1152, generated tokens=121!

[Jagane Sundar @jagane]

New Medium Article: Hadoop/HDFS Events to SQS in real time

Hello all - I just finished writing a new article on medium. As the title indicates, I show how to send a stream of HDFS events (create, close, append, metadata, etc. to AWS SQS - https://t.co/Ar4kX7e0Of

[Matei Zaharia @matei_zaharia]

ChatGPT is super cool, but it can still get facts very wrong, like this example we used in 2021 (https://t.co/R7XPUpinjh). It even generates a second wrong fact (not named after governor). LLMs probably need to use tools like retrieval to avoid being BS-as-a-service.

[InfinStor @InfinStor]

Announcing MLflow Parallels - open source project for running AI compute in Kubernetes

Develop and run complex AI workloads in Kubernetes and log in MLflow. Bridge MLflow and K8s!

No arcane k8s or docker knowlege required

https://t.co/eSEmocJBqA

#mlflow #kubernetes #ai #ml

[Jagane Sundar @jagane]

HIPAA for AI Data with InfinStor MLflow - interested? Sign up for my Wednesday Jul 27 @ 11 Pacific webinar:
https://t.co/8dMrOx4rFK
#mlflow #ai #hipaa #databricks #sagemaker #azureml #MLOps

[Jagane Sundar @jagane]

Here's a really good book on MLflow - Practical Deep Learning at Scale with MLflow https://t.co/Y9tfPEQmod
#mlflow #databricks #infinstor #azureml #sagemaker #mlops

[Jagane Sundar @jagane]

MLflow is a terrific ML experiment tracking system, but it does not manage S3 credentials for artifact storage, i.e. you need to distribute S3 access keys separately. Our product InfinStor MLflow solves that problem #mlflow #infinstor #azureml #databricks

[Jagane Sundar @jagane]

Regulatory compliance, e.g. HIPAA, while using data for AI requires sophisticated tools - InfinStor MLflow. If this is interesting to you, attend my Jul 27 webinar 'HIPAA Compliance for AI Data Using InfinStor MLflow'
https://t.co/8dMrOx4rFK
#mlflow #HIPAA #TensorFlow #PyTorch

[Jagane Sundar @jagane]

I'm presenting a webinar titled 'HIPAA Compliance for AI Data Using InfinStor MLflow' on Jul 27 @ 11 AM Pacific. If this interests you, please signup at: https://t.co/TSDFa8j8HT

[Jagane Sundar @jagane]

Automatically record jupyterlab runs to MLflow - cool open source project by @jnathp #mlflow #Jupyter #sagemaker #infinstor #databricks #azureml #snowflake

[Jagane Sundar @jagane]

Why choose MLflow over other proprietary ML Experiment tracking systems? Because - a new version of pytorch or tensorflow or any other ML library will not break compatibility with your ML tracking system! #mlflow #infinstor #databricks #sagemaker #azureml

[Jagane Sundar @jagane]

'Data is the new oil' is old news. 'AI models are the new oil'. Try MLflow - it is a terrific open source tool for managing models. #mlflow #mlops