The Evolving Role of Data Scientists in the Era of Large Language Models
Plus, the Latest Headlines in AI and How to Overcome the Hurdles of Scaling Generative AI
What’s up, Community!
Shout out to Skalskip from RoboFlow for last Friday’s Hacky Hour session! Skalskip came by to talk about AutoDistill.
What is AutoDistill? It’s a library that allows for creating computer vision models without manual data labelling.
It uses the concept of "distillation" to transfer knowledge from large foundation models to smaller, more efficient models. The process involves automatic image labelling, training a new model on the labelled data, and testing the new model.
AutoDistill has integration with YOLO-NAS, and you can check it out here.
The rest of summer will be packed with awesome virtual events. Stay up to date by joining the Deep Learning Daily Discord!
🧐 What’s in this edition?
🗞️ The Sherry Code (News headlines)
📰 The Deci Digest (Research and repositories)
🤨 The Evolving Role of Data Scientists in the Era of Large Language Models (Opinion)
A poll - Let me know how I did this week
🗞️ The Sherry Code: Your Weekly AI Bulletin
Shout out to Sherry for sharing her top picks for the AI news headlines you need to know about!
Sherry is an active member of the community. She’s at all the events, shares resources on Discord, and is an awesome human being.
Show some support and follow her on Instagram, Twitter, and Threads.
• OpenAI has introduced the GPT-4 API, automatically upgrading GPT-3 models starting January 4, 2024. Developers should switch to gpt-3.5-turbo-instruct for a seamless transition. GPT-3.5 Turbo and GPT-4 fine-tuning are available with priority access for users with older models.
• AWS is introducing the AWS Generative AI Innovation Center, investing $100 million to support small businesses in enhancing operations through affordable generative AI services. The center offers free workshops, training, and tools like Amazon CodeWhisperer and Amazon Bedrock. Already collaborating with Highspot, Lonely Planet, Ryanair, and Twilio on generative AI solutions.
• Elon Musk launches AI firm xAI as he looks to take on OpenAI: Musk explained his plan for building a safer AI. Rather than explicitly programming morality into its AI, xAI will seek to create a "maximally curious" AI."If it tried to understand the true nature of the universe, that's the best thing I can come up with from an AI safety standpoint," Musk said.
• Applied AI is being transformed by the emergence of AI Engineers, who leverage accessible Foundation Models and open-source APIs. Their crucial role includes efficiently applying AI technologies, model evaluation, and tool utilization. The demand for AI Engineers is growing, and success in this field relies on a strong grasp of engineering principles and practical experience without needing a Ph.D.
Overcoming the Hurdles of Scaling Generative AI: Cost-Effective and Eco-Friendly Solutions
In the rapidly evolving world of AI, scaling generative models presents unique challenges.
From high operational costs to environmental concerns, these hurdles can seem daunting. However, this blog post from Deci offers insightful solutions to make scaling these models cost-effective and eco-friendly.
Key Points:
💰 Generative AI models have high operational costs due to their computational demands.
🌍 These models contribute to global greenhouse gas emissions due to their substantial energy consumption.
🎯 Building specialized models can make them more computationally efficient.
⚡ Inference acceleration through quantization and compilation can reduce costs and improve efficiency.
🛠️ Deci's platform offers tools for reducing inference costs and ensuring fast and efficient deployment.
📚 Want to delve deeper into these insights? Read the full blog post here to understand scaling generative AI models comprehensively.
📰 The Deci Digest
• Introducing Keras Core, a preview release of Keras 3.0. Keras Core is a multi-backend version of Keras, supporting TensorFlow, JAX, and PyTorch. It offers a new stateless API for layers, models, metrics, and optimizers and is fully compatible with native workflows in JAX, PyTorch, and TensorFlow. This blog also discusses various aspects of Keras Core, including its compatibility with different data pipelines, its support for pretrained models, and its plans.
• McKinsey's research estimates that generative AI could add $2.6 trillion to $4.4 trillion annually to the global economy across 63 use cases. Along with productivity gains, it is also poised to impact the workforce.
• @alibaba_cloud releases its latest AI image generation model, Tongyi Wanxiang (‘Wanxiang’ means ‘tens of thousands of images’). The Generative AI model can produce rich images in various styles in response to Chinese and English text prompts.
• @AnthropicAI introduces Claude 2. Like its predecessor, Claude 2 can search, summarize, write, code, and answer questions on specific subjects. However, Claude 2 outperforms in various aspects.
• Large language models (LLMs) are one of the key technologies pushing AI forward. This survey report by the @mlopscommunity, explores LLMs’ diverse applications, the challenges in deploying them, and the solutions being implemented.
• For enterprises, the rise of generative AI will change how we interact with all the software and applications and, eventually, how they interact with customers. In this @HarvardBiz, read how enterprises can adapt.
• Prompt engineering is all about experimenting. With gpt-prompt-engineer, you only have to input a description of your task and some test cases, and the system will generate, test, and rank prompts to find the most effective and relevant ones.
🖼️What's in an image beyond pixels and colours?
What secrets lie within the rich data tapestry woven from our sophisticated AI models?
Cracking the code of image data is no easy feat.
The tools for Exploratory Data Analysis (EDA) tailored for image data are few and far between. Most can handle the organized structure of tabular data but stumble when faced with the diverse universe of image data.
Image data presents a panorama of challenges: the difference in colour schemes, the variety of shapes, the complexity of labels and annotations, and the variations in masks and bounding box formats.
How can we use traditional EDA tools to decode such complex variables?
Enter DataGradients 🚀. A new open-source solution that brings the right tools to the right data.
With DataGradients, navigate datasets for image segmentation and object detection tasks without getting lost in the data wilderness.
Ready for a deep dive? Explore DataGradients through this tutorial.
🤨 The Evolving Role of Data Scientists in the Era of Large Language Models
As the landscape of Large Language Models (LLMs) continues to evolve, the roles of product teams, engineers, and machine learning engineers are well-defined, and their value is clear.
But what about data scientists?
Do they primarily shine in evaluations, or are there other areas where their skills make a significant impact?
I posed this question on social media, and it elicited a range of insightful responses that shed light on the evolving role of data scientists in the age of LLMs.
Eugene Yan highlighted the multifaceted role of data scientists, from evaluations to data collection, cleaning, bias and toxicity detection, and augmenting with other ML techniques.
Jayeeta Putatunda emphasized the critical role of data scientists in deciding data requirements, tuning LLMs, detecting drift, and implementing evaluations. She also noted the increasing overlap between data science and machine learning engineering roles, with data scientists now involved in deploying end-to-end LLM pipelines.
Matthew Blasa pointed out the importance of data scientists in evaluating ML and LLM processes, ensuring that the building processes and data products are well-considered.
Greg Coquillo added that data scientists bring deep statistical knowledge, making them instrumental in setting re-training criteria and improving algorithms behind LLMs.
Ashish Patel underscored the impact of data scientists in identifying patterns, predicting trends, and making data-driven decisions, thereby influencing product strategy, user experience, and process optimization.
Dean Pleban mentioned a conversation with Hamel Husain on The MLOps Podcast, which revealed that in the world of LLMs, data development and understanding become more critical and model development less so.
This insight aligns with the argument in a recent paper titled "What Should Data Science Education Do with Large Language Models?"
The paper suggests that LLMs are transforming the responsibilities of data scientists, shifting their focus from hands-on coding and conducting standard analyses to assessing and managing analyses performed by these automated AIs.
However, not everyone agrees with this optimistic view of the future of data science.
Bakulesh Rane suggested that the rise of LLMs could render data science and analytics obsolete. While this perspective may seem extreme, it underscores the transformative potential of LLMs and the need for data scientists to adapt and evolve.
The role of data scientists in the era of LLMs is not limited to evaluations. Their skills and expertise are vital in various areas, from data collection and cleaning to bias detection, algorithm improvement, and making data-driven decisions.
As LLMs continue to evolve, so too, must the role of data scientists and the education that prepares them for these roles.
The future of data science lies at the intersection of AI and human intelligence, with each playing a complementary role in enhancing the overall capabilities and potential of data-driven decision-making.
The insights from industry professionals and the academic paper highlight the transformative potential of LLMs and the need for data scientists to adapt and evolve in this new era of data-driven technologies.
That’s it for this week!
Let me know how I’m doing.
Cheers,