Exploring the Squishy Side of Language Models

The Daily Interesting — Art & Design Edition

March 21, 2026

The idea

I find it helpful to think about language models as a “squishy”...
What do I mean by “squishy”?
Language models feel like something organic and biological rather than something mechanical. It feels like we grew them, rather than built them.
...So given that these models are untamed, squishy, biological things, we have a core tension we need to navigate when we’re designing and building with them: squish vs. structure
We’re trying to make an unpredictable and opaque system adhere to our rigid expectations for how computers behave. We currently have a mismatch between our old and new mental models for computer systems.[1][2]

— Maggie Appleton (BarnsworthBurning extract)[1]

Who said it (and the context)

Maggie Appleton is a product designer, anthropologist, and self-described "mediocre developer" who creates visual essays on programming, design, and AI. She leads design at Elicit, an AI research lab building tools for researchers using language models, drawing from her background in cultural anthropology to explore human-AI interactions.[2][4] In her essay "Squish Meets Structure," she reframes LLMs as biological entities, urging designers to balance their emergent creativity with reliable constraints.[2]

A small expansion

Appleton's metaphor reveals LLMs' "squishy" unpredictability—prone to surprises, hallucinations, and emergent behaviors—as both a creative boon and a design challenge, requiring "Goldilocks" interfaces that tame chaos without stifling potential.[2]

Another lens on the idea

Andrej Karpathy, former OpenAI and Tesla AI director, similarly describes neural networks as "a very complicated alien artefact," echoing the organic otherness of LLMs and the need to probe their black-box nature through structured experimentation.[2]


Why this matters across art and design

  • Visual identity
    Generate brand explorations with LLMs (e.g., mood boards from prompts), but layer structure like token limits or style guides to ensure consistent, non-hallucinated outputs that evolve organically.[2]

  • Product & UX
    Build "tiny reasoning engines" for specific tasks—chained prompts for search or tone analysis—graduating from squishy ideation to structured delivery, reducing user overload in AI features.[2]

  • Architecture
    Use LLMs for parametric sketches or site analyses, constraining inputs (e.g., material libraries) to balance generative variety with buildable coherence, mimicking biological adaptation.[2]

  • Studio art
    Harness squish for puns, similes, or style shifts in text-to-image tools; add structure via iterative refinement loops to turn unpredictable outputs into intentional compositions.[2]


Two quick drills

  1. Squish-Structure Spectrum (10 minutes)
    Prompt an LLM for 3 creative variants of a design element (high squish: high temperature). Then refine with constraints (low squish). Note the Goldilocks balance—what feels alive yet reliable?[2]

  2. Prompt Chain Audit (15 minutes)
    Break a complex task (e.g., generate UX copy) into 3 sub-steps: observe > reflect > output. Test chain vs. single prompt—which yields more structured, trustworthy results?[2]


Thread to remember

Builds on tool defaults (42), modular components (51), and river persistence (52)—Appleton equips us to design with squishy AI tools, channeling their alien emergence into structured creativity, much like alive-by-default media (11).


Sources

Try this today: Where in your work could a dash of "squish" spark fresh ideas?