A few loose ideas are worth testing against reality.
Voice, silence, and direct neural input
Typing on a phone keyboard is a poor way to move thoughts out of a head. Voice input is already better in many situations: Stanford data puts speech at about 153 words per minute, compared with roughly 52 for typing. But speaking has obvious limits. You may need quiet, you may not want to disturb people around you, or you may not want to be overheard. Brain-computer interfaces address both the speed problem and the privacy problem by reading neural signals directly. The current numbers are not science fiction—one decoding rate is near 90 words per minute, and a Meta prototype has reported a 3% error rate—but they are still early. If non-invasive BCIs can be made practical for mobile devices, they could also offer accessibility benefits for people with disabilities.
Foldables may find a new reason to exist
AI assistants make more sense when they have room to work. On a desktop, using an assistant is naturally asynchronous: you can hand off a task and keep doing something else. On a phone, switching between apps or using a cramped split screen is clumsy. A foldable can provide something closer to tablet-style split-screen space, which is much better for AI-assisted multitasking. Google's Pixel Fold, for instance, can run AI meeting-note generation and document editing side by side. Newer foldables are being designed around AI interaction, and as costs come down they could move from a premium niche into everyday work and life.
Emotion as hidden pruning
In a value function, emotion is key. Emotion may be an important part of the invisible pruning that happens when we make decisions. This fits Antonio Damasio's somatic marker hypothesis: emotional and bodily signals help filter options before conscious reasoning, speeding up choice. People with prefrontal damage that impairs emotional processing often have trouble deciding. AI work is beginning to incorporate simulated emotion modules, with the aim of improving decision efficiency and making system behavior closer to human thinking.
AGI needs machines that keep learning
The key to AGI is continual learning. Human intelligence is lifelong; most current AI follows a train-deploy pattern, stays static after deployment, requires full retraining, and is prone to forgetting. A successful AGI should be deployable as an agent that keeps learning—not one that is instantly an expert. It should learn every day, or every moment, instead of waiting for another centralized training run. That means incremental learning: absorbing new knowledge after deployment while retaining old skills. If that can be made to work, it could break the limits of narrow AI and move toward general intelligence.