PodcastAI Careers Insights from Ex Meta Staff EngClaimIlya Reznik says fine-tuning can help produce a particular output shape, but prompt engineering should be tried first.1:27
Threads / Open weights against the API / 2025
Open weights against the API in 2025
18 sessions
PodcastAI Careers Insights from Ex Meta Staff Eng
Scaling your data and AI from 0-100 with open sourceClaimPycafe offers a self-hosted Python server that can be installed with pip for organizations with security or privacy requirements.13:19
PodcastI Let An AI Play Pokémon! - Claude plays Pokémon CreatorPushed backFine-tuning is not usually the best first step; prompting should generally be pushed further before fine-tuning is considered.13:31
PodcastWe're All Finetuning IncorrectlyPushed backThe idea that prompting or prompt fine-tuning is inherently undesirable was softened; prompting is useful before a system reaches its ceiling.14:30
Building Out GPU CloudsPushed backDemetrios Brinkmann noted that fine-tuning is often criticized as difficult and potentially harmful when done incorrectly, while Mohan Atreya argued that it can become a zero-code workflow.45:57
Office Hours on DuckDB, AWS Glue, and IcebergPushed backEric said the distinction between Delta Lake and Iceberg was not one he could fully explain from memory, while Hossam framed them as competing open-table-format approaches.15:27
PodcastPackaging MLOps Tech Neatly for Engineers and Non-engineersClaimJukka Remes's team built an MLOps platform from loosely integrated open-source components to support client development, deployment, and operation of models.3:59
PodcastThe Creator of FastAPI's Next ChapterClaimSebastián Ramírez says FastAPI grew from repeated difficulty building APIs for machine-learning applications.8:42
PodcastThe Missing Data Stack for Physical AIClaimRerun's open-source project logs and visualizes multimodal data that changes over time and provides SDKs for Python, Rust, and C++.34:59
PodcastThe Truth About LLM TrainingClaimProsus evaluates models for domain understanding, language capabilities, cost, fine-tuning and distillation potential, and inference performance.1:50
The Hidden Infrastructure Behind Every AI AgentClaimEvery time an AI agent takes an action, queries a model, calls an API, or fetches a tool, something has to handle that traffic.1:21
Reading groupSmall Language Models are the Future of Agentic AIClaimSonam Gupta describes an agent as a system in which a language model reasons and uses tools or APIs to complete autonomous tasks.7:37
PodcastIs Open Source Software Actually Secure?ClaimHudson Buzby says large organizations generally use OpenAI or Anthropic to some extent but are looking toward open-source models because of cost, provider trust, data privacy, and the desire for an open-source environment.12:39
An AI Company By AccidentPushed backThe speakers distinguish between product-market fit and having an actual commercial product, with Russ d'Sa saying LiveKit had the former but initially only had an open-source project.5:38
Fine-Tuning is BrokenPushed backTanmay Chopra disputes the idea that current fine-tuning endpoints are performing true fine-tuning, saying they mainly perform additional pretraining with the same next-token loss and output vocabulary.10:10
Fine-Tuned Models Are Getting Out of HandPushed backDemetrios questioned whether fine-tuning small language models adds unnecessary complexity compared with a RAG-based system.0:57