← Accueil actu | Salle du marché
2026-09-10 · Source：Unite AI
A new analysis argues that AI's four hardest problems — data scarcity, weak economics, rogue agents and catastrophic forgetting — are framing errors, and existing data remains largely untapped.
Four obstacles dominate the current discourse around frontier AI development: the sense that the industry is running out of data, doubts over whether AI financially scales, the risk of agents going rogue, and the tendency of models to catastrophically forget whenever developers attempt to modify them. A new analysis argues that none of these walls is as solid as it appears — the problems persist largely because the field is looking at them the wrong way.
The most contested of these is data scarcity. The analysis pushes back directly on the premise: we are not running out of data, because the data that exists has never been fully used. In this view, the perceived shortage reflects incomplete exploitation of available material rather than a genuine exhaustion of the world's information supply.
The belief itself is shaping behavior. According to the report, AI leaders have internalized the idea that data is running out, and their response has been to have models generate new data and then train on it — a synthetic feedback loop adopted as a workaround for a wall that, the argument goes, may not actually be there. The distinction matters: if the constraint is one of utilization rather than supply, then the industry's pivot to self-generated training material addresses a problem that better use of existing data might solve.
The remaining three problems receive the same reframing treatment. The economics question — that AI does not financially scale — the safety question of agents going rogue, and the technical problem of catastrophic forgetting, where changing a model erases what it previously knew, are all presented as difficulties that look insurmountable under the current framing but become tractable once viewed from the right angle.
For an industry that treats these four problems as fixed constraints on its trajectory, the argument's core claim is procedural rather than technical: before accepting data exhaustion, weak economics, rogue agents and forgetting as hard walls, the field should first verify that the walls are real. In this framing, they are not.
Mots-clés: Synthetic Data · Catastrophic Forgetting · AI Safety · Data Scaling · AI Agents
萬安算交所 AIXX · AIXX · Bourse compute vérifiée · Accéder à la salle