Great panel at the Berkeley AI Summit @BerkeleyRDI today on the future of agents. A few takeaways:
- Long-horizon agents are demonstrating economic value but still have a long way to go. They’re hard to scale partly because of how long training runs take. A 12-hour task means only two runs per day, or ~60 per month, dramatically slowing iteration compared to past training attempts. Whether long-horizon tasks are best solved through agent harnesses (memory, compaction, sub-agents, etc.) vs direct model improvements is yet to be decided.
- Recursive self-improvement is still probably years away. Auto-research systems produce incremental gains early, but they don’t yet seem capable of driving truly novel research or meaningful recursive improvement.
- Reward hacking and benchmark maxing are natural behaviors. As more companies adopt open-weight models and fine-tune them for domain-specific use cases, training environment design becomes critical. Producing real economic improvement, not just higher benchmark scores, is still a non-trivial challenge.
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
Allah
♥️ I LOVE ALLAH♥️
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD
I❤️
محمد ﷺ
MUHAMMAD