Gradient Dissent: Conversations on AI

About This Episode

Shreya Shankar is a computer scientist, PhD student in databases at UC Berkeley, and co-author of "Operationalizing Machine Learning: An Interview Study", an ethnographic interview study with 18 machine learning engineers across a variety of industries on their experience deploying and maintaining ML pipelines in production.

Shreya explains the high-level findings of "Operationalizing Machine Learning"; variables that indicate a successful deployment (velocity, validation, and versioning), common pain points, and a grouping of the MLOps tool stack into four layers. Shreya and Lukas also discuss examples of data challenges in production, Jupyter Notebooks, and reproducibility.

Show notes (transcript and links): http://wandb.me/gd-shreya

---

💬 *Host:* Lukas Biewald

---

*Subscribe and listen to Gradient Dissent today!*

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Sarah Catanzaro is a General Partner at Amplify Partners, and one of the leading investors in AI and ML. Her investments include RunwayML, OctoML, and Gantry.

Sarah and Lukas discuss lessons learned from the "AI renaissance" of the mid 2010s and compare the general perception of ML back then to now. Sarah also provides insights from her perspective as an investor, from selling into tech-forward companies vs. traditional enterprises, to the current state of MLOps/developer tools, to large language models and hype bubbles.

Show notes (transcript and links): http://wandb.me/gd-sarah-catanzaro

---

⏳ Timestamps:

0:00 Intro

1:10 Lessons learned from previous AI hype cycles

11:46 Maintaining technical knowledge as an investor

19:05 Selling into tech-forward companies vs. traditional enterprises

25:09 Building point solutions vs. end-to-end platforms

36:27 LLMS, new tooling, and commoditization

44:39 Failing fast and how startups can compete with large cloud vendors

52:31 The gap between research and industry, and vice versa

1:00:01 Advice for ML practitioners during hype bubbles

1:03:17 Sarah's thoughts on Rust and bottlenecks in deployment

1:11:23 The importance of aligning technology with people

1:15:58 Outro

---

📝 Links

📍 "Operationalizing Machine Learning: An Interview Study" (Shankar et al., 2022), an interview study on deploying and maintaining ML production pipelines: https://arxiv.org/abs/2209.09125

---

Connect with Sarah:

📍 Sarah on Twitter: https://twitter.com/sarahcat21

📍 Sarah's Amplify Partners profile: https://www.amplifypartners.com/investment-team/sarah-catanzaro

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan

---

Subscribe and listen to Gradient Dissent today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Cristóbal Valenzuela is co-founder and CEO of Runway ML, a startup that's building the future of AI-powered content creation tools. Runway's research areas include diffusion systems for image generation.

Cris gives a demo of Runway's video editing platform. Then, he shares how his interest in combining technology with creativity led to Runway, and where he thinks the world of computation and content might be headed to next. Cris and Lukas also discuss Runway's tech stack and research.

Show notes (transcript and links): http://wandb.me/gd-cristobal-valenzuela

---

⏳ Timestamps:

0:00 Intro

1:06 How Runway uses ML to improve video editing

6:04 A demo of Runway’s video editing capabilities

13:36 How Cris entered the machine learning space

18:55 Cris’ thoughts on the future of ML for creative use cases

28:46 Runway’s tech stack

32:38 Creativity, and keeping humans in the loop

36:15 The potential of audio generation and new mental models

40:01 Outro

---

🎥 Runway's AI Film Festival is accepting submissions through January 23! 🎥

They are looking for art and artists that are at the forefront of AI filmmaking. Submissions should be between 1-10 minutes long, and a core component of the film should include generative content

📍 https://aiff.runwayml.com/

--

📝 Links

📍 "High-Resolution Image Synthesis with Latent Diffusion Models" (Rombach et al., 2022)", the research paper behind Stable Diffusion: https://research.runwayml.com/publications/high-resolution-image-synthesis-with-latent-diffusion-models

📍 Lexman Artificial, a 100% AI-generated podcast: https://twitter.com/lexman_ai

---

Connect with Cris and Runway:

📍 Cris on Twitter: https://twitter.com/c_valenzuelab

📍 Runway on Twitter: https://twitter.com/runwayml

📍 Careers at Runway: https://runwayml.com/careers/

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan

---

Subscribe and listen to Gradient Dissent today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Jeremy Howard is a co-founder of fast.ai, the non-profit research group behind the popular massive open online course "Practical Deep Learning for Coders", and the open source deep learning library "fastai".

Jeremy is also a co-founder of #Masks4All, a global volunteer organization founded in March 2020 that advocated for the public adoption of homemade face masks in order to help slow the spread of COVID-19. His Washington Post article "Simple DIY masks could help flatten the curve." went viral in late March/early April 2020, and is associated with the U.S CDC's change in guidance a few days later to recommend wearing masks in public.

In this episode, Jeremy explains how diffusion works and how individuals with limited compute budgets can engage meaningfully with large, state-of-the-art models. Then, as our first-ever repeat guest on Gradient Dissent, Jeremy revisits a previous conversation with Lukas on Python vs. Julia for machine learning.

Finally, Jeremy shares his perspective on the early days of COVID-19, and what his experience as one of the earliest and most high-profile advocates for widespread mask-wearing was like.

Show notes (transcript and links): http://wandb.me/gd-jeremy-howard-2

---

⏳ Timestamps:

0:00 Intro

1:06 Diffusion and generative models

14:40 Engaging with large models meaningfully

20:30 Jeremy's thoughts on Stable Diffusion and OpenAI

26:38 Prompt engineering and large language models

32:00 Revisiting Julia vs. Python

40:22 Jeremy's science advocacy during early COVID days

1:01:03 Researching how to improve children's education

1:07:43 The importance of executive buy-in

1:11:34 Outro

1:12:02 Bonus: Weights & Biases

---

📝 Links

📍 Jeremy's previous Gradient Dissent episode (8/25/2022): http://wandb.me/gd-jeremy-howard

📍 "Simple DIY masks could help flatten the curve. We should all wear them in public.", Jeremy's viral Washington Post article: https://www.washingtonpost.com/outlook/2020/03/28/masks-all-coronavirus/

📍 "An evidence review of face masks against COVID-19" (Howard et al., 2021), one of the first peer-reviewed papers on the effectiveness of wearing masks: https://www.pnas.org/doi/10.1073/pnas.2014564118

📍 Jeremy's Twitter thread summary of "An evidence review of face masks against COVID-19": https://twitter.com/jeremyphoward/status/1348771993949151232

📍 Read more about Jeremy's mask-wearing advocacy: https://www.smh.com.au/world/north-america/australian-expat-s-push-for-universal-mask-wearing-catches-fire-in-the-us-20200401-p54fu2.html

---

Connect with Jeremy and fast.ai:

📍 Jeremy on Twitter: https://twitter.com/jeremyphoward

📍 fast.ai on Twitter: https://twitter.com/FastDotAI

📍 Jeremy on LinkedIn: https://www.linkedin.com/in/howardjeremy/

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan

Jerome Pesenti is the former VP of AI at Meta, a tech conglomerate that includes Facebook, WhatsApp, and Instagram, and one of the most exciting places where AI research is happening today.

Jerome shares his thoughts on Transformers-based large language models, and why he's excited by the progress but skeptical of the term "AGI". Then, he discusses some of the practical applications of ML at Meta (recommender systems and moderation!) and dives into the story behind Meta's development of PyTorch. Jerome and Lukas also chat about Jerome's time at IBM Watson and in drug discovery.

Show notes (transcript and links): http://wandb.me/gd-jerome-pesenti

---

⏳ Timestamps:

0:00 Intro

0:28 Jerome's thought on large language models

12:53 AI applications and challenges at Meta

18:41 The story behind developing PyTorch

26:40 Jerome's experience at IBM Watson

28:53 Drug discovery, AI, and changing the game

36:10 The potential of education and AI

40:10 Meta and AR/VR interfaces

43:43 Why NVIDIA is such a powerhouse

47:08 Jerome's advice to people starting their careers

48:50 Going back to coding, the challenges of scaling

52:11 Outro

---

Connect with Jerome:

📍 Jerome on Twitter: https://twitter.com/an_open_mind

📍 Jerome on LinkedIn: https://www.linkedin.com/in/jpesenti/

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan, Lavanya Shukla

---

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

D. Sculley is CEO of Kaggle, the beloved and well-known data science and machine learning community.

D. discusses his influential 2015 paper "Machine Learning: The High Interest Credit Card of Technical Debt" and what the current challenges of deploying models in the real world are now, in 2022. Then, D. and Lukas chat about why Kaggle is like a rain forest, and about Kaggle's historic, current, and potential future roles in the broader machine learning community.

Show notes (transcript and links): http://wandb.me/gd-d-sculley

---

⏳ Timestamps:

0:00 Intro

1:02 Machine learning and technical debt

11:18 MLOps, increased stakes, and realistic expectations

19:12 Evaluating models methodically

25:32 Kaggle's role in the ML world

33:34 Kaggle competitions, datasets, and notebooks

38:49 Why Kaggle is like a rain forest

44:25 Possible future directions for Kaggle

46:50 Healthy competitions and self-growth

48:44 Kaggle's relevance in a compute-heavy future

53:49 AutoML vs. human judgment

56:06 After a model goes into production

1:00:00 Outro

---

Connect with D. and Kaggle:

📍 D. on LinkedIn: https://www.linkedin.com/in/d-sculley-90467310/

📍 Kaggle on Twitter: https://twitter.com/kaggle

---

Links:

📍 "Machine Learning: The High Interest Credit Card of Technical Debt" (Sculley et al. 2014): https://research.google/pubs/pub43146/

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan, Anish Shah, Lavanya Shukla

---

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Emad Mostaque is CEO and co-founder of Stability AI, a startup and network of decentralized developer communities building open AI tools. Stability AI is the company behind Stable Diffusion, the well-known, open source, text-to-image generation model.

Emad shares the story and mission behind Stability AI (unlocking humanity's potential with open AI technology), and explains how Stability's role as a community catalyst and compute provider might evolve as the company grows. Then, Emad and Lukas discuss what the future might hold in store: big models vs "optimal" models, better datasets, and more decentralization.

-

🎶 Special note: This week’s theme music was composed by Weights & Biases’ own Justin Tenuto with help from Harmonai’s Dance Diffusion.

-

Show notes (transcript and links): http://wandb.me/gd-emad-mostaque

-

⏳ Timestamps:

00:00 Intro

00:42 How AI fits into the safety/security industry

09:33 Event matching and object detection

14:47 Running models on the right hardware

17:46 Scaling model evaluation

23:58 Monitoring and evaluation challenges

26:30 Identifying and sorting issues

30:27 Bridging vision and language domains

39:25 Challenges and promises of natural language technology

41:35 Production environment

43:15 Using synthetic data

49:59 Working with startups

53:55 Multi-task learning, meta-learning, and user experience

56:44 Optimization and testing across multiple platforms

59:36 Outro

-

Connect with Jehan and Motorola Solutions:

📍 Jehan on LinkedIn: https://www.linkedin.com/in/jehanw/

📍 Jehan on Twitter: https://twitter.com/jehan/

📍 Motorola Solutions on Twitter: https://twitter.com/MotoSolutions/

📍 Careers at Motorola Solutions: https://www.motorolasolutions.com/en_us/about/careers.html

-

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Angelica Pan, Lavanya Shukla, Anish Shah

-

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Jehan Wickramasuriya is the Vice President of AI, Platform & Data Services at Motorola Solutions, a global leader in public safety and enterprise security.

In this episode, Jehan discusses how Motorola Solutions uses AI to simplify data streams to help maximize human potential in high-stress situations. He also shares his thoughts on augmenting synthetic data with real data and the challenges posed in partnering with startups.

Show notes (transcript and links): http://wandb.me/gd-jehan-wickramasuriya

-

⏳ Timestamps:

00:00 Intro

00:42 How AI fits into the safety/security industry

09:33 Event matching and object detection

14:47 Running models on the right hardware

17:46 Scaling model evaluation

23:58 Monitoring and evaluation challenges

26:30 Identifying and sorting issues

30:27 Bridging vision and language domains

39:25 Challenges and promises of natural language technology

41:35 Production environment

43:15 Using synthetic data

49:59 Working with startups

53:55 Multi-task learning, meta-learning, and user experience

56:44 Optimization and testing across multiple platforms

59:36 Outro

-

Connect with Jehan and Motorola Solutions:

📍 Jehan on LinkedIn: https://www.linkedin.com/in/jehanw/

📍 Jehan on Twitter: https://twitter.com/jehan/

📍 Motorola Solutions on Twitter: https://twitter.com/MotoSolutions/

📍 Careers at Motorola Solutions: https://www.motorolasolutions.com/en_us/about/careers.html

-

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Cayla Sharp, Angelica Pan, Lavanya Shukla

-

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Will Falcon is the CEO and co-founder of Lightning AI, a platform that enables users to quickly build and publish ML models.

In this episode, Will explains how Lightning addresses the challenges of a fragmented AI ecosystem and reveals which framework PyTorch Lightning was originally built upon (hint: not PyTorch!) He also shares lessons he took from his experience serving in the military and offers a recommendation to veterans who want to work in tech.

Show notes (transcript and links): http://wandb.me/gd-will-falcon

---

⏳ Timestamps:

00:00 Intro

01:00 From SEAL training to FAIR

04:17 Stress-testing Lightning

07:55 Choosing PyTorch over TensorFlow and other frameworks

13:16 Components of the Lightning platform

17:01 Launching Lightning from Facebook

19:09 Similarities between leadership and research

22:08 Lessons from the military

26:56 Scaling PyTorch Lightning to Lightning AI

33:21 Hiring the right people

35:21 The future of Lightning

39:53 Reducing algorithm complexity in self-supervised learning

42:19 A fragmented ML landscape

44:35 Outro

---

Connect with Lightning

📍 Website: https://lightning.ai

📍 Twitter: https://twitter.com/LightningAI

📍 LinkedIn: https://www.linkedin.com/company/pytorch-lightning/

📍 Careers: https://boards.greenhouse.io/lightningai

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Anish Shah, Cayla Sharp, Angelica Pan, Lavanya Shukla

---

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​

Aaron Colak is the Leader of Core Machine Learning at Qualtrics, an experiment management company that takes large language models and applies them to real-world, B2B use cases.

In this episode, Aaron describes mixing classical linguistic analysis with deep learning models and how Qualtrics organized their machine learning organizations and model to leverage the best of these techniques. He also explains how advances in NLP have invited new opportunities in low-resource languages.

Show notes (transcript and links): http://wandb.me/gd-aaron-colak

---

⏳ Timestamps:

00:00 Intro

00:57 Evolving from surveys to experience management

04:56 Detecting sentiment with ML

10:57 Working with large language models and rule-based systems

14:50 Zero-shot learning, NLP, and low-resource languages

20:11 Letting customers control data

25:13 Deep learning and tabular data

28:40 Hyperscalers and performance monitoring

34:54 Combining deep learning with linguistics

40:03 A sense of accomplishment

42:52 Causality and observational data in healthcare

45:09 Challenges of interdisciplinary collaboration

49:27 Outro

---

Connect with Aaron and Qualtrics

📍 Aaron on LinkedIn: https://www.linkedin.com/in/aaron-r-colak-3522308/

📍 Qualtrics on Twitter: https://twitter.com/qualtrics/

📍 Careers at Qualtrics: https://www.qualtrics.com/careers/

---

💬 Host: Lukas Biewald

📹 Producers: Riley Fields, Cayla Sharp, Angelica Pan, Lavanya Shukla

---

Subscribe and listen to our podcast today!

👉 Apple Podcasts: http://wandb.me/apple-podcasts​​

👉 Google Podcasts: http://wandb.me/google-podcasts​

👉 Spotify: http://wandb.me/spotify​