A journey through pixels, originality and AI | Eric Hu

Summary

Eric Hu, a designer and typography-focused creative, reflects on how digital images relate to “creation” vs. “discovery.” Using the atomic unit of digital media—the pixel—he explains how images on screens are ultimately finite (for example, a small 5×5 grid yields only a limited number of possible bitmaps), even if the space of possible images can feel infinite at human scale.

He connects this to broader questions in philosophy and computing, comparing the randomness of pixel patterns to ideas like the infinite monkey theorem, and questioning whether mathematical objects/numbers are “made” or “found.”

He then uses noise as a practical bridge: humans can perceive meaning in noisy images through pattern recognition, while machines struggle without training. He describes how AI denoising evolved via supervised learning (e.g., labeling images as “cat” while adding noise), and how more complex generative models can “hallucinate” meaningful outputs when prompted. In this framing, AI image generation operates in latent space, where prompts navigate an enormous probability space to produce outputs.

Finally, Hu argues that art isn’t just the raw output of random data—context and human interpretation matter. He discusses concerns around AI, including plagiarism/consent/data ethics, and worries less about art itself and more about how payers/employers might misuse the tools. He concludes by reframing “creator” identity through an alchemical metaphor: the maker selects from an immense space of possibilities and assigns meaning to what emerges.

---

Artistic techniques / creative processes shown (or discussed)

• Typography & logo-oriented design foundation
• Working with cultural institutions and music-related design
• AI image generation workflow concepts
  • Navigating latent space using prompts
  • Generating small seed/noisy latent images that evolve toward the prompt signal
• AI denoising / image restoration approach (conceptual)
  • Train models on labeled images
  • Add progressively heavier noise and learn to recover the underlying subject
• Intentional “ugly” exploration with generative tools
  • Use generation to produce imperfect or unappealing outputs as a creative strategy
  • Experiment beyond imitation—less “copying,” more “breaking it”

---

Concepts / ideas emphasized

• Pixel-level mechanics of digital imagery

  • Pixels as on/off light states (binary) and RGB color channels
  • The number of possible images is finite given constraints (resolution, color depth)

• Creation vs. discovery

  • Images may feel “infinite” conceptually, but are bounded by parameters
  • Math and numbers are treated as existing in some sense, while labeling/mapping makes them usable

• Human pattern recognition vs. machine perception

  • Humans extract meaning from noise via context
  • Computers require training or explicit guidance

• Context as a core ingredient of art

  • Random pixel arrangements aren’t art on their own; interpretation and circumstance create “meaning”

• Ethics of data and consent

  • Concern about training on freely shared work without proper consent

---

Steps / workflow (as described conceptually)

• Digital image construction framing

  • Choose a pixel grid size (example: 5×5)
  • Set each pixel’s state (off/on)
  • Optionally introduce color via RGB channel values (e.g., 0–255 ranges)

• AI denoising training (as described)

  • Start with labeled images (e.g., “cat”)
  • Add noise to inputs
  • Train a model to reconstruct the original subject despite noise
  • Increase noise difficulty until recognition still succeeds

• Generative AI prompting (as described)

  • Provide a text prompt
  • Sample a latent random representation (described as a long number that becomes a small noisy image)
  • Use the prompt as a nudge to evolve toward a more relevant output

---

Creators / contributors featured

• Eric Hu (speaker)
• Dan Abramoff (mentioned as breaking down the math/naming analogy)
• NVIDIA researchers (credited with doing/advancing denoising and related deep learning work)
• “Mike Jones” (mentioned in an analogy about phone numbers; not otherwise credited)