Data Machina #251
Substack
Data Machina's latest roundup covers six AI side-quests for your weekend, plus a deep pile of research links on in-context learning vs fine-tuning. Worth noting: several top researchers now argue smarter prompting is closing the gap on fine-tuning entirely.
Based on reporting by Substack — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
This week's Data Machina newsletter reads less like a digest and more like a to-do list for AI-curious weekends. Carlos, the curator, frames it around six activities for anyone stuck indoors during a long bank holiday, but the real substance sits in the debate he surfaces about in-context learning versus fine-tuning.
First, the fun stuff. ByteDance's team released StoryDiffusion, a zero-shot diffusion model that generates consistent comics and short videos from text prompts, and Carlos says he's already made a few strips with it. There's also a tutorial from a creator named Tony on stitching together an agent stack — CrewAI for orchestration, Ollama for local models, Groq for fast inference, and AgentOps for monitoring — to build a bot that scrapes YouTube comments and turns them into content insights. And for anyone who wants a break from actual work, AI Town, the a16z-built simulation where AI characters chat and socialize in a virtual town, apparently remains addictive, especially once you skip the cloud signup and run it locally with Llama 3 and Pinokio.
But the newsletter's real meat is the fine-tuning fight. Ethan, a researcher Carlos cites, makes the blunt claim that fine-tuning is dead now that long-context prompting has closed the performance gap. DeepMind's own research on many-shot in-context learning backs some of that up, showing strong results on complex reasoning tasks when you throw enough examples into the prompt. Yet another paper on long-context ICL complicates the picture: most of the apparent gains come from models matching against similar examples already seen, not from genuine task learning. Microsoft Research, meanwhile, is trying to patch a separate flaw — the
My take — AI-written commentary, not fact-checked reporting
I'll bet on the messier truth: fine-tuning isn't dead, it's just less necessary for narrow tasks now that context windows are absurdly long, and pretending otherwise is the kind of overclaim that makes AI Twitter tiresome. The stuff I actually trust here is DeepMind's careful many-shot ICL work — everything else in this roundup is a fun rabbit hole, not a research consensus.
Read more about this at: Substack