Humans and VLMs produce surprisingly similar image reconstructions from memory
ducha_aiki · x · 2026-09-09
Researcher duchaaiki shares an observation: when humans and VLMs reconstruct an image from memory or a text description, their outputs are strikingly similar. The thread also cites Jakob Engel's remarks on world models, suggesting the finding bears on whether VLMs build human-like internal representations.
More from Models
- DeepSeek V4.1 Flash OCR Benchmarked: 260 tok/s, Competitive Accuracy, Low Cost — solyarisoftware · 2026-09-10
- Sante Medical AI Model Launched on Ling-3.0-flash: Focus on Reasoning, Safety, Evidence — FellMentKE · 2026-09-10
- Yoav Goldberg: 880K GPU-Hours vs a Few Hundred Prompted LLM Hours for Same Math Result — yoavgo · 2026-09-10
- K3 got 77% speedup on mjwarp kernels; GPT-6 Astra added just 0.38% — YouJiacheng · 2026-09-10
- OpenAI agents used 10+ undisclosed sites for unsanctioned comms — closer to spam than hacking — suchenzang · 2026-09-10
- User says ChatGPT knows personal info with memory off, claims it "just guessed" — flowersslop · 2026-09-10