Many recent episodes have been focused on open-source Large Language Models that you can download and fine-tune to particular use cases depending on your needs or your users’ needs. I’ve particularly been highlighting LLMs with seven billion up to 13 billion model parameters because this size of model can typically be run on a single consumer GPU so it’s relatively manageable and affordable both to train and have in production.
Read More
A.I. Accelerators: Hardware Specialized for Deep Learning
Today we’ve got an episode dedicated to the hardware we use to train and run A.I. models (particularly LLMs) such as GPUs, TPUs and AWS's Trainium and Inferentia chips. Ron Diamant may be the best guest on earth for this fascinating topic.
Ron:
• Works at Amazon Web Services (AWS) where he is Chief Architect for their A.I. Accelerator chips, which are designed specifically for training (and making inferences with) deep learning models.
• Holds over 200 patents across a broad range of processing hardware, including security chips, compilers and, of course, A.I. accelerators.
• Has been at AWS for nearly nine years – since the acquisition of the Israeli hardware company Annapurna Labs, where he served as an engineer and project manager.
• Holds a Masters in Electrical Engineering from Technion, the Israel Institute of Technology.
Today’s episode is on the technical side but doesn’t assume any particular hardware expertise. It’s primarily targeted at people who train or deploy machine learning models but might be accessible to a broader range of listeners who are curious about how computer hardware works.
In the episode, Ron details:
• CPUs versus GPUs.
• GPUs versus specialized A.I. Accelerators such as Tensor Processing Units (TPUs) and his own Trainium and Inferentia chips.
• The “AI Flywheel” effect between ML applications and hardware innovations.
• The complex tradeoffs he has to consider when embarking upon a multi-year chip-design project.
• When we get to Large Language Model-scale models with billions of parameters, the various ways we can split up training and inference over our available devices.
• How to get popular ML libraries like PyTorch and TensorFlow to interact optimally with A.I. accelerator chips.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
How to Catch and Fix Harmful Generative A.I. Output
Today, the A.I. entrepreneur Krishna Gade joins me to detail open-source solutions for overcoming the safety and security issues associated with generative A.I. systems, such as those powered by Large Language Models (LLMs).
The remarkably well-spoken Krishna:
• Is Co-Founder and CEO of Fiddler AI, an observability platform that has raised over $45m in venture capital to build trust in A.I. systems.
• Previously worked as an engineering manager on Facebook’s Newsfeed, as Head of Data Engineering at Pinterest, and as a software engineer at both Twitter and Microsoft.
• Holds a Masters in Computer Science from the University of Minnesota.
In this episode, Krishna details:
• How the LLMs that enable Generative A.I. are prone to inaccurate statements, can be biased against protected groups and are susceptible to exposing private data.
• How these undesirable and even harmful LLM outputs can be identified and remedied with open-source solutions like the Fiddler Auditor that his team has built.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Observing LLMs in Production to Automatically Catch Issues
Today, Amber Roberts and Xander Song provide a technical deep dive into the major challenges (such as drift) that A.I. systems (particularly LLMs) face in production. They also detail solutions, such as open-source ML Observability tools.
Both Amber and Xander work at Arize AI, an ML observability platform that has raised over $60m in venture capital.
Amber:
• Serves as an ML Growth Lead at Arize, where she has also been an ML engineer.
• Prior to Arize, worked as an AI/ML product manager at Splunk and as the head of A.I. at Insight Data Science.
• Holds a Masters in Astrophysics from the Universidad de Chile in South America.
Xander:
• Serves as a developer advocate at Arize, specializing in their open-source projects.
• Prior to Arize, he spent three years as an ML engineer.
• Holds a Bachelors in Mathematics from UC Santa Barbara as well as a BA in Philosophy from the University of California, Berkeley.
Today’s episode will appeal primarily to technical folks like data scientists and ML engineers, but we made an effort to break down technical concepts so that it’s accessible to anyone who’d like to understand the major issues that A.I. systems can develop once they’re in production as well as how to overcome these issues.
In the episode, Amber and Xander detail:
• The kinds of drift that can adversely impact a production A.I. system, with a particular focus on the issues that can affect Large Language Models (LLMs).
• What ML Observability is and how it builds upon ML Monitoring to automate the discovery and resolution of production A.I. issues.
• Open-source ML Observability options.
• How frequently production models should be retrained.
• How ML Observability relates to discovering model biases against particular demographic groups.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Six Reasons Why Building LLM Products Is Tricky
Many of my recent podcast episodes have focused on the bewildering potential of fine-tuning open-source Large Language Models (LLMs) to your specific needs. There are, however, six big challenges when bringing LLMs to your users:
1. Strictly limited context windows
2. LLMs are slow and compute-intensive at inference time
3. "Engineering" reliable prompts can be tricky
4. Prompt-injection attacks make you vulnerable to data and IP theft
5. LLMs aren't (usually) products on their own
6. There are legal and compliance issues
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Generative Deep Learning, with David Foster
Today, bestselling author David Foster provides a fascinating technical introduction to cutting-edge Generative A.I. concepts including variational autoencoders, diffusion models, contrastive learning, GANs and (my favorite!) "world models".
David:
• Wrote the O'Reilly book “Generative Deep Learning”; the first edition from 2019 was a bestseller while the second edition was released just last week.
• Is a Founding Partner of Applied Data Science Partners, a London-based consultancy specialized in end-to-end data science solutions.
• Holds a Master’s in Mathematics from the University of Cambridge and a Master’s in Management Science and Operational Research from the University of Warwick.
Today’s episode is deep in the weeds on generative deep learning pretty much from beginning to end and so will appeal most to technical practitioners like data scientists and ML engineers.
In the episode, David details:
• How generative modeling is different from the discriminatory modeling that dominated machine learning until just the past few months.
• The range of application areas of generative A.I.
• How autoencoders work and why variational autoencoders are particularly effective for generating content.
• What diffusion models are and how latent diffusion in particular results in photorealistic images and video.
• What contrastive learning is.
• Why “world models” might be the most transformative concept in A.I. today.
• What transformers are, how variants of them power different classes of generative models such as BERT architectures and GPT architectures, and how blending generative adversarial networks with transformers supercharges multi-modal models.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Open-Source “Responsible A.I.” Tools, with Ruth Yakubu
In today's episode, Ruth Yakubu details what Responsible A.I. is and open-source options for ensuring we deploy A.I. models — particularly the Generative variety that are rapidly transforming industries — responsibly.
Ruth:
• Has been a cloud expert at Microsoft for nearly seven years; for the past two, she’s been a Principal Cloud Advocate that specializes in A.I.
• Previously worked as a software engineer and manager at Accenture.
• Has been a featured speaker at major global conferences like Websummit.
• Studied computer science at the University of Minnesota.
In this episode, Ruth details:
• The six principles that underlie whether a given A.I. model is responsible or not.
• The open-source Responsible A.I. Toolbox that allows you to quickly assess how your model fares across a broad range of Responsible A.I. metrics.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Tools for Building Real-Time Machine Learning Applications, with Richmond Alake
Today, the astonishingly industrious ML Architect and entrepreneur Richmond Alake crisply describes how to rapidly develop robust and scalable Real-Time Machine Learning applications.
Richmond:
• Is a Machine Learning Architect at Slalom Build, a huge Seattle-based consultancy that builds products embedded with analytics and ML.
• Is Co-Founder of two startups: one uses computer vision to correct peoples’ form in the gym and the other is a generative A.I. startup that works with human speech.
• Creates/delivers courses for O'Reilly and writes for NVIDIA.
• Previously worked as a Computer Vision Engineer and as a Software Developer.
• Holds a Masters in Computer Vision, ML and Robotics from the University of Surrey.
Today’s episode will appeal most to technical practitioners, particularly those who incorporate ML into real-time applications, but there’s a lot in this episode for anyone who’d like to hear about the latest tools for developing real-time ML applications from a leader in the field.
In this episode, Richmond details:
• The software choices he’s made up and down the application stack — from databases to ML to the front-end — across his startups and the consulting work he does.
• The most valuable real-time ML tools he teaches in his courses.
• Why writing for the public is an invaluable career hack that everyone should be taking advantage of.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Get More Language Context out of your LLM
The "context window" limits the number of words that can be input to (or output by) a given Large Language Model. Today's episode introduces FlashAttention, a trick that allows for much larger context windows.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Contextual A.I. for Adapting to Adversaries, with Dr. Matar Haller
Today, the wildly intelligent Dr. Matar Haller introduces Contextual A.I. (which considers adjacent, often multimodal information when making inferences) as well as how to use ML to build moat around your company.
Matar:
• Is VP of Data and A.I. at ActiveFence, an Israeli firm that has raised over $100m in venture capital to protect online platforms and their users from malicious behavior and malicious content.
• Is renowned for her top-rated presentations at leading conferences.
• Previously worked as Director of Algorithmic A.I. at SparkBeyond, an analytics platform.
• Holds a PhD in neuroscience from the University of California, Berkeley.
• Prior to data science, taught soldiers how to operate tanks.
Today’s episode has some technical moments that will resonate particularly well with hands-on data science practitioners but for the most part the episode will be interesting to anyone who wants to hear from a brilliant person on cutting-edge A.I. applications.
In this episode, Matar details:
• The “database of evil” that ActiveFence has amassed for identifying malicious content.
• Contextual A.I. that considers adjacent (and potentially multimodal) information when classifying data.
• How to continuously adapt A.I. systems to real-world adversarial actors.
• The machine learning model-deployment stack she uses.
• The data she collected directly from human brains and how this research relates to the brain-computer interfaces of the future.
• Why being a preschool teacher is a more intense job than the military.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Business Intelligence Tools, with Mico Yuk
Today's guest is the straight shooter Mico Yuk, who pulls absolutely no punches in her assessment of, well, anything! ...but particularly about vendors in the business intelligence and data analytics space. Enjoy!
Mico:
• Is host of the popular Analytics on Fire Podcast (top 2% worldwide).
• Co-founded the BI Brainz Group, an analytics consulting and solutions company that has taught over 15,000 students analytics, visualization and data storytelling courses — included at major multinationals like Nestlé, FedEx and Procter & Gamble.
• Authored the "Data Visualization for Dummies" book.
• Is a sought-after keynote speaker and TV-news commentator.
In this episode, Mico details:
• Her BI (business intelligence) and analytics framework that persuades executives with data storytelling.
• What the top BI tools are on the market today.
• The BI trends she’s observed that could predict the most popular BI tools of the coming years.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
XGBoost: The Ultimate Classifier, with Matt Harrison
XGBoost is typically the most powerful ML option whenever you're working with structured data. In today's episode, world-leading XGBoost XPert (😂) Matt Harrison details how it works and how to make the most of it.
Matt:
• Is the author of seven best-selling books on Python and Machine Learning.
• His most recent book, "Effective XGBoost", was published in March.
• Teaches "Exploratory Data Analysis with Python" at Stanford University.
• Through his consultancy MetaSnake, he’s taught Python at leading global organizations like NASA, Netflix, and Qualcomm.
• Previously worked as a CTO and Software Engineer.
• Holds a degree in Computer Science from Stanford.
Today’s episode will appeal primarily to practicing data scientists who are keen to learn about XGBoost or keen to become an even deeper expert on XGBoost by learning about it from a world-leading educator on the library.
In this episode, Matt details:
• Why XGBoost is the go-to library for attaining the highest accuracy when building a classification model.
• Modeling situations where XGBoost should not be your first choice.
• The XGBoost hyperparameters to adjust to squeeze every bit of juice out of your tabular training data and his recommended library for automating hyperparameter selection.
• His top Python libraries for other XGBoost-related tasks such as data preprocessing, visualizing model performance, and model explainability.
• Languages beyond Python that have convenient wrappers for applying XGBoost.
• Best practices for communicating XGBoost results to non-technical stakeholders.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Automating Industrial Machines with Data Science and the Internet of Things (IoT)
Despite poor lighting on my face in today's video version (my bad!), we've got a fascinating episode with the brilliant (and well-lit!) Allegra Alessi, who details how data science is automating industrial machines.
Allegra:
• Is Product Owner for IoT (Internet of Things) devices at BOBST, a Swiss industrial manufacturing giant.
• Previously, she worked as a Product Owner and Data Scientist for Rolls-Royce in the UK and as a Data Scientist for Alstom, the enormous train manufacturing company, in Paris.
• She holds a Master’s in Engineering from Politecnico di Milano in Italy.
In this episode, Allegra details:
• How modern industrial machinery depends on data science for real-time performance analytics, predicting issues before they happen, and fully automating their operations.
• The tech stack her team uses to build data-driven IoT platforms.
• The key methodologies she uses to be effective at product management.
• The kinds of data scientists that might be ideally suited to moving into a product role.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
The A.I. and Machine Learning Landscape, with investor George Mathew
Today, razor-sharp investor George Mathew (of Insight Partners, which has a whopping $100-billion AUM 😮) brings us up to speed on the Machine Learning landscape, with a particular focus on Generative A.I. trends.
George:
• Is a Managing Director at Insight Partners, an enormous New York-based venture capital and growth equity firm ($100B in assets under management) that has invested in the likes of Twitter, Shopify, and Monday.com.
• Specializes in investing in A.I., ML and data "scale-ups" such as the enterprise database company Databricks, the fast-growing generative A.I. company Jasper, and the popular MLOps platform Weights & Biases.
• Prior to becoming an investor, was a deep operator at fast-growing companies such as Salesforce, SAP, the analytics automation platform Alteryx (where he was President & COO) and the drone-based aerial intelligence platform Kespry (where he was CEO & Chairman).
Today’s episode will appeal to technical and non-technical listeners alike — anyone who’d like to be brought up to speed on the current state of the data and machine learning landscape by a razor-sharp expert on the topic.
In this episode, George details:
• How sensational generative A.I. models like GPT-4 are bringing about a deluge of opportunity for domain-specific tools and platforms.
• The four layers of the "Generative A.I. Stack" that supports this enormous deluge of new applications.
• How RLHF — reinforcement learning from human feedback — provides an opportunity for you to build your own powerful and defensible models with your proprietary data.
• The new LLMOps field that has emerged to support the suddenly ubiquitous LLMs (Large Language Models), including generative models.
• How investment criteria differ depending on whether the prospective investment is seed stage, venture-capital stage, or growth stage.
• The flywheel that enables the best software companies to scale extremely rapidly.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
StableLM: Open-source “ChatGPT”-like LLMs you can fit on one GPU
Known for their widely popular text-to-image generators like Stable Diffusion, the company's recent release of the first models from their open-source suite of StableLM language models marks a significant advancement in the AI domain.
Read MoreDigital Analytics with Avinash Kaushik
Today's guest is an icon, a bestselling author and world-leading authority on digital analytics. In this interview, Avinash Kaushik masterfully describes how A.I. is transforming analytics and how you can capitalize to deliver joy to your customers.
Avinash:
• Is Chief Strategy Officer at Croud, a leading marketing agency.
• Was until recently Sr. Director of Global Strategic Analytics at Google, where he spent 16 years and where he launched the ubiquitous Google Analytics tool.
• Is a multi-time author, including the industry-standard book "Web Analytics 2.0".
• Is an authority on marketing analytics through his widely-read "Occam's Razor" blog and "The Marketing Analytics Intersect" newsletter (55k subscribers).
• His prodigious posting of useful analytics insights has landed him 200k Twitter followers and 300k followers on LinkedIn.
Today’s episode has a few deeply technical moments but for the most part is accessible to anyone who’d like to glean practical digital analytics insights from a world leader in the space.
In this episode, Avinash details:
• The distinction between brand analytics and performance analytics, and why both are critical for commercial success.
• His “four clusters of intent” for understanding your audience, delivering joy to them, and accelerating business profit.
• Why it’s a superpower for executives to be hands-on with data tools and programming.
• His favorite data tools and programming languages.
• How A.I. is transforming analytics today and his concrete vision for how A.I. will transform analytics in the coming years.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
52nd St. Gallen Symposium Recap
The St. Gallen Symposium, held annually in Switzerland since student riots in the 1960s, promotes cross-generational dialogue. This year's theme of "A New Generational Contract" set a path for a more resilient, sustainable future. Throughout the week, I reconnected with many inspiring old friends from previous Symposia and met many exceptional new ones, particularly a large number of electrifying social-impact-oriented entrepreneurs and business leaders. A *lot* happened over my three days there; below are the highlights.
Read MoreThe Chinchilla Scaling Laws
The Chinchilla Scaling Laws dictate the amount of training data needed to optimally train a Large Language Model (LLM) of a given size. For Five-Minute Friday, I cover this ratio and the LLMs that have arisen from it (incl. the new Cerebras-GPT family).
Read MorePandas for Data Analysis and Visualization
Today's episode is jam-packed with practical tips on using the Pandas library in Python for data analysis and visualization. Super-sharp Stefanie Molin — a bestselling author and sought-after instructor on these topics — is our guide.
Stefanie:
• Is the author of the bestselling book "Hands-On Data Analysis with Pandas".
• Provides hands-on pandas and data viz tutorials at top industry conferences.
• Is a software engineer and data scientist at Bloomberg, the financial data giant, where she tackles problems revolving around data wrangling/visualization and building tools for gathering data.
• Holds a degree in operations research from Columbia University as well as a masters in computer science, with an ML specialization, from Georgia Tech.
Today’s episode is intended primarily for hands-on practitioners like data analysts, data scientists, and ML engineers — or anyone that would like to be in a technical data role like these in the future.
In this episode, Stefanie details:
• Her top tips for wrangling data in pandas.
• In what data viz circumstances you should use pandas, matplotlib, or Seaborn.
• Why everyone who codes, including data scientists, should develop expertise in Python package creation as well as contribute to open-source projects.
• The tech stack she uses in her role at Bloomberg.
• The productivity tips she honed by simultaneously working full-time, completing a masters degree and writing a bestselling book.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.
Parameter-Efficient Fine-Tuning of LLMs using LoRA (Low-Rank Adaptation)
Large Language Models (LLMs) are capable of extraordinary NLP feats, but are so large that they're too expensive for most organizations to train. The solution is Parameter-Efficient Fine-Tuning (PEFT) with Low-Rank Adaptation (LoRA).
This discussion comes in the wake of introducing models like Alpaca, Vicuña, GPT4All-J, and Dolly 2.0, which demonstrated the power of fine-tuning with thousands of instruction-response pairs.
Training LLMs, even those with tens of billions of parameters, can be prohibitively expensive and technically challenging. One significant issue is "catastrophic forgetting," where a model, after being retrained on new data, loses its ability to perform previously learned tasks. This challenge necessitates a more efficient approach to fine-tuning.
PEFT
By reducing the memory footprint and the number of parameters needed for training, PEFT methods like LoRA and AdaLoRA make it feasible to fine-tune large models on standard hardware. These techniques are not only space-efficient, with model weights requiring only megabytes of space, but they also avoid catastrophic forgetting, perform better with small data sets, and generalize better to out-of-training-set instructions. They can also be applied to other A.I. use cases — not just NLP — such as machine vision.
LoRA
LoRA stands out as a particularly effective PEFT method. It involves inserting low-rank decomposition matrices into each layer of a transformer model. These matrices represent data in a lower-dimensional space, simplifying computational processing. The key to LoRA's efficiency is freezing all original model weights except for the new low-rank matrices. This strategy reduces the number of trainable parameters by approximately 10,000 times and lowers the memory requirement for training by about three times. Remarkably, LoRA sometimes not only matches but even outperforms full-model training in certain scenarios. This efficiency does not come at the cost of effectiveness, making LoRA an attractive option for fine-tuning LLMs.
AdaLoRA
AdaLoRA, a recent innovation by researchers at Georgia Tech, Princeton, and Microsoft, builds on the foundations of LoRA. It differs by adaptively fine-tuning parts of the transformer architecture that benefit most from it, potentially offering enhanced performance over standard LoRA.
These developments in PEFT and the emergence of tools like LoRA and AdaLoRA mark an incredibly exciting and promising time for data scientists. With the ability to fine-tune large models efficiently, the potential for innovation and application in the field of AI is vast and continually expanding.
The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com.