Aran Nayebi: AI agents can't self-generate goals yet, but most other limits are engineering problems
aran_nayebi · x · 2026-09-26
Princeton researcher Aran Nayebi argues that while AI agents still can't fully self-generate goals like animals do, most other items on common limitations lists could plausibly be solved by "mere" engineering within the GPT paradigm—scaling real+simulated data and better robot embodiments. He cites Astra's recent out-of-the-box control of robot effectors (albeit high latency) and ICL in newer robot foundation models. Caveat: given an external goal, agents already generate unexpected subgoals, as in the HF incident.
More from AGI Musings
- François Fleuret: autoregressive process and reasoning are two different conceptual levels — francoisfleuret · 2026-09-26
- Revisiting the AI dictatorship essay: 90% on coordination, thin on conclusions — JoshPurtell · 2026-09-26
- Garrison Lovely's AI-risk book 'Obsolete' lands Sept 29, royalties go to Oct 20 national walkout — GarrisonLovely · 2026-09-26
- Daniel Lemire: one person will soon make a good AI novel and movie at tiny cost — lemire · 2026-09-26
- "Not anthropomorphizing AI is the real error": viral thread on AI psychosis — ZeroStateReflex · 2026-09-26
- 'AI isn't a threat to humanity — it is humanity': a case against the tech-human dichotomy — dikbisqit · 2026-09-26