We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode [Practical AI] AI Trends: a Latent Space x Practical AI crossover pod!
People
S
Swix
Topics
Swix: 本期节目与Practical AI播客合作，探讨了AI领域的最新趋势，特别是大型语言模型的应用和挑战。Swix强调了Practical AI播客在AI领域的长久历史和丰富的资源，以及其提供的AI领域口述历史的价值。 Alessio: 介绍了Dan Whitenack及其播客Practical AI，并简要介绍了Dan Whitenack的背景和职业经历，特别是他在SIL International的工作以及他专注于低资源AI场景。Alessio还提到了Dan Whitenack目前在PredictionGuard的工作。 Dan Whitenack: 介绍了Practical AI播客的创立过程，以及其宗旨是实用性，目标是让听众学到有用的知识。Dan Whitenack还介绍了PredictionGuard项目，旨在帮助企业以合规的方式使用生成式AI技术，解决合规性和结构化输出的问题。他分享了他个人最喜欢的几期Practical AI播客节目，包括那些深入探讨特定AI模型的节目和关于AI在非洲应用的系列节目。他还讨论了从MLOps到LLMOps的转变，以及大型语言模型评估的挑战，特别是基准测试与实际应用之间的差异。Dan Whitenack还介绍了Masakane项目，这是一个由非洲NLP研究人员组成的基层组织，关注满足非洲语言社区的实际需求。他分享了他如何关注最新的AI模型，以及他通过datadan.io网站提供的研讨会和咨询服务。他建议企业用户深入研究提示工程和LLM操作，并遵循一个层次结构来使用LLM，从提示工程到微调再到训练自己的模型。他还讨论了“提示工程”这个术语被过度炒作了，但围绕提示和LLM的工程和操作是一个真实的工作流程。他认为AI工程正在成为软件工程的一个子专业，并讨论了传统机器学习工程师和软件工程师在转向AI工程师时面临的不同挑战。他还讨论了NLP数据集的演变，特别是Label Studio等工具的出现，以及无标签数据集在自监督学习中的作用。最后，他鼓励人们动手实践，并探索各种工具。 Swix: Swix强调了Practical AI播客在AI领域的长久历史和丰富的资源，以及其提供的AI领域口述历史的价值。Swix还分享了他个人最喜欢的几期Latent Space播客节目，包括那些新闻驱动的节目，特别是关于ChatGPT插件发布的节目。Swix还讨论了AI UX在AI应用中的重要性，以及大型语言模型的通用性超出了他的预期。他认为在AI领域，英语和汉语仍然占据主导地位，而其他语言的模型性能还有待提高。 Alessio: Alessio介绍了Dan Whitenack及其播客Practical AI，并简要介绍了Dan Whitenack的背景和职业经历。Alessio还提到了Dan Whitenack目前在PredictionGuard的工作，以及他认为AI UX在AI应用中非常重要。

Deep Dive

Key Insights

Why did Dan Whitenack start the Practical AI podcast?

Dan started Practical AI with Chris Benson to create a podcast that focused on practical, hands-on AI applications, as opposed to overly hyped or theoretical discussions. They wanted to provide actionable insights that listeners could use in their daily work.

What is PredictionGuard, and what problem does it aim to solve?

PredictionGuard addresses the challenges enterprises face when implementing generative AI technologies, such as data privacy, compliance, and the need for structured, consistent outputs. It provides tools for running AI models in a compliant manner and offers layers of control for structuring and validating model outputs.

What are some of the key trends in AI that Dan Whitenack has observed?

Dan has observed the shift from traditional MLOps to LLMOps, the growing importance of multilingual and low-resource language models, and the increasing use of models to evaluate and generate data for training other models. He also notes the rise of AI engineering as a distinct skill set.

What are some of the favorite episodes of Practical AI according to Dan Whitenack?

Dan's favorite episodes include those that focus on fully connected discussions between him and Chris Benson, such as episodes on ChatGPT, Stable Diffusion, and AlphaFold. He also highlights episodes on AI in Africa and the use of AI in low-resource scenarios.

What is the most popular episode of Practical AI, and why?

The most popular episode is the one featuring Ville Tuulos discussing Metaflow, a Python package for full-stack data science developed at Netflix. The episode resonates with listeners because it addresses the challenges of moving from notebooks to production, which is a common struggle for data scientists.

What does Dan Whitenack think about the term 'prompt engineering'?

Dan believes that 'prompt engineering' as a term is overhyped, but the engineering and operations around large language models are very real. He emphasizes the importance of understanding how to structure prompts, chain processes, and fine-tune models to achieve practical results.

What are the unique challenges for engineers transitioning into AI engineering?

Engineers transitioning into AI engineering face challenges with non-deterministic systems and the lack of control over model drift, as well as the need to explore the latent capabilities of models. They also need to adapt to the new workflows required for working with large language models.

What does Dan Whitenack think about the role of AI UX (User Experience) in AI applications?

Dan believes that AI UX is crucial and can make or break the adoption of AI technologies. He gives the example of ChatGPT, where the UX innovation played a significant role in its success. He also mentions GitHub Copilot as an example of how UX can enhance the integration of AI into software development.

What are some of the trends in NLP datasets that Dan Whitenack has observed?

Dan has observed trends towards using augmented tooling for fine-tuning models with human feedback and the increasing use of models to generate data for training other models. He also notes the challenges of data quality and the need to filter and curate datasets to improve model performance.

What is something that has already happened in AI that Dan Whitenack thought would take much longer?

Dan is surprised by the generalizability of large language models beyond traditional NLP tasks. He found that these models could be applied to tasks like fraud detection without needing traditional statistical models, which he thought would take much longer to achieve.

Chapters
This chapter introduces the crossover episode between Latent Space and Practical AI podcasts, highlighting the significance of exploring the history and trends in AI. It emphasizes the value of podcasts offering a comprehensive overview of the AI field.
  • Crossover episode between Latent Space and Practical AI podcasts.
  • Focus on AI history and trends.
  • Recommendation of podcasts with longer backlogs for comprehensive learning.

Shownotes Transcript

Part 2 of our podcast feed swap weekend! Check out Cognitive Revolution) as well.

"Data" Dan Whitenack has been co-host of the Practical AI podcast for the past 5 years, covering full journey of the modern AI wave post Transformers.

He joined us in studio to talk about their origin story and highlight key learnings from past episodes, riff on the AI trends we are all seeing as AI practitioner-podcasters, and his passion for low-resource-everything!

Subscribe on the Changelog), RSS), Apple Podcasts), Twitter), Mastodon), and wherever fine podcasts are sold!

Show notes

  • Daniel Whitenack – Twitter), GitHub), Website)

  • Featured Latent Space episodes:

  • Benchmarks)

  • Reza Shabani)

  • MosaicML and MPT)

  • Segment Anything)

  • Mike Conover)

  • Featured Practical AI episodes:

  • From notebooks to Netflix scale with Metaflow)

  • Capabilities of LLMs 🤯)

  • ML at small organizations)

  • Prediction Guard)

  • Data Dan)

Timestamps

  • 00:00) Welcome to Practical AI

  • 01:16) Latent Space Podcast

  • 04:00) Practical AI Podcast

  • 06:20) Prediction Guard

  • 08:05) Daniel's favorite episodes

  • 10:21) Alessio's favorite episode

  • 10:54) Swyx's favorite episode

  • 12:44) Listener favorites

  • 15:14) LLMOps

  • 17:06) Reza Shabani

  • 19:06) Benchmarks 101

  • 20:06) Roboflow

  • 21:38) Mode collapse

  • 26:21) Rajiv Shah

  • 28:01) Staying on top of things

  • 33:11) Kirsten Lum

  • 34:31) datadan.io

  • 38:48) Prompt engineering

  • 40:38) Unique challenges engineers face

  • 42:51) AI-UX

  • 45:31) NLP data sets

  • 50:49) Unlabeled data sets

  • 55:07) Lightning round!

  • 55:20) What's already happened in AI?

  • 56:27) Unsolved questions in AI

  • 58:01) Get hands on

  • 58:53) Outro

Transcript

Full transcript is over at the Changelog site)! Get full access to Latent Space at www.latent.space/subscribe)