On AI Synesthesia

• Multimodal AI ends computing’s 500-year bias toward text.
  • Visual thinkers, spatial reasoners, creatives finally included.
• Unified latent space: one model natively spans text, image, code, video, audio.
  • Not stitched APIs — shared semantic embeddings across modalities.
• GPT-4o hybrid: autoregressive semantic outline, then diffusion refinement.
• “AI synesthesia”: convert cognitive strengths across domains.
  • Coders gain design fluency; visual thinkers gain verbal output.
• Labor: floor rises (skills democratize), ceiling rises (specialists expand).
  • Middle-skill routine tasks at highest automation risk.
  • Demand for originality and domain-specific insight may surge.
• Intelligence becomes fluid — your strengths no longer format-locked.

X discourse

• @sandeepnailwal: “LLM based AI is NOT conscious… We’ve reduced consciousness to ‘did the output sound like it had feelings?’” (1092 likes)
• @AnthropicAI: “We identified emotion vectors: patterns of neural activity for concepts like ‘happy’ or ‘calm.’” (819 likes)
• @sedielem: “Since we’re talking about high-dimensional Gaussians today: human intuition tends to break down in high dimensions.” (250 likes)

Sonya Huang and Pat Grady, Sequoia Capital — with AIs as co-authors · 2025-04-28 · Read on sequoiacap.com