Could code compress visual reality better than learned neural weights?
umesh_ai · x · 2026-10-12
Umesh poses a question that could reshape generative AI: can code become a more efficient compression of visual reality than learned neural network weights? He cites a claim that Claude's next model, using pure JS code, surpasses image models in subtlety and video models in physics, and will become the cost-efficient option within six months (unverified; the original post claims Opus was not trained on video/audio data). The thread points to a route dispute: generation as program synthesis versus generation as weight interpolation.
More from AGI Musings
- willcb: every non-RL algorithm is a proxy — policy gradients are the minimum-variance unbiased estimator — willcb · 2026-10-12
- Is AI Agent just a wrapper? Users don't care about the plumbing, only the chore done — sujingshen · 2026-10-12
- Researcher Revives Sci-Fi Short Story About a Model 'Waking Up' in Pain During RL Training — voooooogel · 2026-10-12
- AI slop now floods every public platform, from X to Reddit — private communities are the last stand — ayushtweetshere · 2026-10-12
- AIandDesign: everyone working 40 hours a week on a computer is 'completely cooked' — AIandDesign · 2026-10-12
- Access to data isn't understanding: why 'AI knows your business' is overblown — Rajxai · 2026-10-12