Long-Horizon Agents Backfiring? Dev Slams Lack of Profitable Use Cases
MarcJSchmidt · x · 2026-08-10
Developer MarcJSchmidt retweeted a thread arguing that the industry's obsession with "long-horizon agents" in model training is finally backfiring. He claims these fully autonomous agents fundamentally don't work, challenging anyone to show a profitable business built entirely by one.
He notes that the only impressive things they’ve built so far are 3D game prototypes cloning existing titles like Call of Duty. This is only possible because human teams already figured out the requirements, controls, and mechanics years ago, and that knowledge was absorbed into the model's weights—meaning the agent is just regurgitating known patterns.
More from AGI Musings
- From Wittgenstein to AI: The 60-Year Cycle of Abandoning Hidden Structures — adityaag · 2026-08-10
- ICML Paper: The Real Challenge for Superintelligence Is Coexistence, Not Capability — xuanalogue · 2026-08-10
- Researcher Warns of Imminent Parasitic Cyber-Capable AI Agents — StephenLCasper · 2026-08-10
- Scholars Propose AI-Driven Overhaul for Academic Peer Review — paulnovosad · 2026-08-10
- Europa Mission in 2030 May Find Alien Life, Sparking AI Consciousness Debate — AIandDesign · 2026-08-10
- AI Overreliance Breeds 'Human Zombies' Who Blindly Copy-Paste Outputs — bindureddy · 2026-08-10