[Practical AI] AI Trends: a Latent Space x Practical AI crossover pod!

Episode Summary

Executive Summary: This crossover episode between Latent Space and Practical AI explores the evolution of AI podcasts, the shift from hype to practical usage, and how enterprises are adopting LLMs. Dan Whitenack discusses Practical AI’s history, favorite episodes, evaluation challenges, multilingual/low-resource AI, model/data curation, and why AI engineering now centers on workflow, UX, and guardrails rather than just prompt writing.

Main Topics: Podcast crossover and AI oral history: The episode frames both shows as part of a longer AI media history, emphasizing the value of podcast backlogs as an oral record of how AI thinking evolved over time. Practical AI’s origin and mission: Dan explains how Practical AI emerged from conversations with the Changelog network and co-host Chris Benson, with a goal of making AI coverage useful, grounded, and non-hype-driven. From MLOps to LLMOps / AI engineering: The discussion centers on how AI work is shifting from traditional model training and deployment toward prompt engineering, chaining, guardrails, evaluation, and operational workflows around foundation models. Evaluation, benchmarks, and model-generated data: The hosts discuss the limits of traditional benchmarks, the rise of LLM-based evaluation, and the growing use of model-generated or AI-assisted data in training and benchmarking. Low-resource languages and global AI communities: Dan highlights his work on low-resource languages and grassroots efforts like Masakane, arguing that AI systems must reflect linguistic and cultural diversity rather than only English-centric assumptions. Enterprise adoption and practical tooling: The conversation covers how enterprises are still early in LLM adoption, why they need better tooling and UX, and how products like Prediction Guard and Label Studio help bridge the gap from demos to production. Model selection, open source, and ecosystem navigation: The speakers discuss how to choose among rapidly proliferating models on Hugging Face, the importance of openness and licensing, and how model behavior depends heavily on training data and use case.

Key Arguments: Practical AI was created to be useful and grounded, in contrast to earlier AI content that was more hype-driven. The most valuable podcast episodes are often the ones that capture a moment in AI history, because they preserve how people reacted before outcomes were known. Enterprise AI adoption is still shallow; many organizations have not yet integrated LLMs deeply into workflows. Prompt engineering alone is overhyped; the real discipline is broader AI engineering, including chaining, guardrails, data augmentation, and output validation. Traditional benchmarks are insufficient for production LLMs because many real-world tasks are open-ended rather than multiple-choice. Model-based evaluation is becoming more important, especially for LLMs and machine translation, where metrics like COMET outperform simple static scores. Low-resource languages and non-English contexts remain underrepresented, so foundation models need more linguistic diversity to be broadly useful. UX is a major part of AI product success; ChatGPT and Copilot show that interface design can be as important as model capability. Data scientists and software engineers face different challenges when moving into AI engineering: one group must adapt to non-determinism, the other to workflow and product design. The data mix behind foundation models is still opaque, so practitioners should experiment with multiple models and not rely solely on leaderboard rankings.

Data Points: Practical AI episode count: over 200 episodes - Dan notes the show has covered AI consistently for more than 200 episodes. Practical AI audience favorite episode lead: ~30% more listens than any other episode - The Metaflow episode was described as the clear listener favorite by a wide margin. ChatGPT plugins live event attendance: 4,000 people - Swix says the ChatGPT app store / plugins release episode drew about 4,000 live listeners. Practical AI history: 5 years - Dan says he has been in the hosts’ ears for about five years covering AI before it was cool. Changelog network age: 11 or 12 years - Dan references the Changelog podcast network as having been around for roughly 11–12 years. Model size reference: 175 billion or 1 trillion parameters - Used to illustrate how much of a model’s latent capability may remain unexplored. Common Crawl quality issue: 100% - Dan cites a study where 100% of data labeled as Latin-character Arabic was not actually Arabic. Human-generated data share in a released dataset: less than 1% - The discussion of Meta’s released multimodal dataset notes that less than 1% was human-generated. Hugging Face model scale: 200,000 models - A rough estimate mentioned while discussing how many models are available to navigate on Hugging Face. Model release example: 13 billion parameters - Swix mentions a newly released News Research model fine-tuned on 300,000+ instructions. Instruction tuning dataset size: over 300,000 instructions - Referenced in the discussion of a newly released model on Hugging Face.

Pivotal Quotes: "“Prompt engineering as like a term is probably too hyped. I think engineering and ops around large language models though is a real thing.”" — Dan Whitenack: Dan distinguishes between superficial prompt-selling and the broader operational discipline needed for production AI. "“We need to develop a new workflow around these things.”" — Dan Whitenack: He explains that enterprise teams are still rebuilding intuition after the shift from classic ML to LLM-based systems. "“The last mile of showing something on the screen and making it consumable, easily usable by people is perhaps as valuable as the actual training of the model itself.”" — Swix: This frames the importance of AI UX and product design in making model capabilities useful.

Implications: Listeners should expect AI work to become more workflow-, UX-, and evaluation-heavy, not just model-centric. The future of practical AI will depend on better tooling, multilingual inclusion, and experimentation beyond benchmarks and prompt tricks.