Skip to content

What's next, the LLM landscape in motion

A survey lesson opening Phase 3. You can now build, ship, and operate an LLM application; the field is not standing still. This lesson maps the directions it is moving in, lighter on mechanical depth than the rest of the track, so you can read new model releases and platform announcements with a builder’s eye. The source curriculum is the Full Stack Deep Learning LLM Bootcamp (Spring 2023), by Charles Frye, Sergey Karayev, and Josh Tobin, freely available at fullstackdeeplearning.com/llm-bootcamp with recorded lectures on the Full Stack Deep Learning YouTube channel.

You will name the six directions (longer context, multimodality, smaller specialized models, build-vs-buy spectrum, agents, reasoning models); read each through the three productive limits from lesson 2 and the build-vs-buy spectrum; explain why longer context does not eliminate retrieval and when smaller specialized models win; apply the builder’s-eye reading discipline to a new model release; and use the lesson-7 regression discipline to adopt new capabilities safely rather than silently.

This is a survey lesson by design: lighter pedagogy, breadth-over-depth, pointing forward to three deeper Phase 3 lessons (training your own at lesson 9, agents at lesson 10, industry perspective at lesson 11).

This is lesson 8 of 11, opening Phase 3 (advanced and the field). It maps the territory the next three lessons take deeper: lesson 9 deep-dives the “fine-tune an open model” point on the build-vs-buy spectrum; lesson 10 deep-dives the agents direction; lesson 11 is the industry-perspective capstone for the track. It also threads back to lesson 2’s productive limits and lesson 7’s LLMOps regression discipline, both used as the constant lens.

Prerequisites: lesson 7 of this track (the LLMOps regression discipline this lesson uses as the safe-adoption tool for new capabilities). The constant lens is lesson 2’s three productive limits and lesson 4’s tool-use loop (the seed of agents). Familiarity with all of Phase 1 + 2 helps; this lesson is the bridge from “what you’ve built” to “where the field is going.”

None. Survey lesson; the discipline is reading the field through the existing constraints, not deriving anything new.

The single capability this lesson builds: situate today’s LLM application stack in the field’s direction. Concretely, you will be able to:

  • Name the six directions the LLM landscape is moving in
  • Read each direction through the three productive limits and the build-vs-buy spectrum
  • Explain why longer context does not eliminate retrieval, and when smaller specialized models win
  • Apply the builder’s-eye reading discipline to a new model release
  • Use the lesson-7 regression discipline to adopt new capabilities safely
  • Read time: about 10 minutes (survey-lean by design)
  • Practice time: about 10 minutes (read a hypothetical frontier-model release with the builder’s eye; reason about a mix-architecture migration; plus flashcards)
  • Difficulty: standard (no math; the work is internalizing the six directions and the reading discipline)