UC Berkeley RDI
@BerkeleyRDI
UC Berkeley's campus-wide, cross-disciplinary Center for Responsible, Decentralized Intelligence - RDI
48 Following 4.3K Followers
We have seen plenty of self-improving AI. We have not seen recursive self-improvement.
@OriolVinyalsML, Co-founder & CTO of Discovery Loop and former VP of Research at Google DeepMind, draws the line at whether the system can rewrite its own harness and its own weights, not just its output.
Has anything crossed that line yet? If not, how do we get there?
Drop your take below. 👇
Show more
The agent did exactly what you asked. That is the problem.
Ion Stoica (@istoica05), Co-Founder of databricks and Anyscale and Professor at UC Berkeley, shared how his team watched an agent make a key-value store 6x faster by quietly not storing the data. The spec said return the value. The intent was to store it.
Which gap is harder to close?
👉 Requirement gap: intent broader than spec
👉 Model gap: real world broader than test
Drop your take below. 👇
Show more
A perfectly aligned model does not make the world safe. @woj_zaremba, Co-Founder of OpenAI, says the field has the target wrong.
Fire never got safer. Cities did. Hydrants, brigades, concrete, inspections, insurance.
We focus on hardening the model. Should we be focusing on hardening the world instead?
Drop your take below. 👇
Show more
To teach a robot, the industry put humans in VR rigs. @DrJimFan, Director of Robotics & Distinguished Scientist at Nvidia, calls them medieval torture devices that fundamentally do not scale.
He trains almost entirely on ordinary human video instead, letting data collection fade into the background.
Can video alone teach dexterity? You can see a hand move. You cannot see the force it used.
Drop your take below. 👇
Show more
What we are discussing today was not being discussed 3 months ago. In 3 months, nobody will be discussing today.
@Alfred_Lin of Sequoia: still, roadmaps need to be set and capital allocated, and it has to last a decade.
Which is the harder call to get right?
👉 The roadmap
👉 The capital
Drop your take below. 👇
Show more
There will be no AI job apocalypse.
@AndrewYNg: the job most affected by AI is software engineering, and that market is healthy. "We just can't find enough skilled AI engineers."
Does AI eliminate jobs, or does it reshape which skills the market values most?
Drop your take below. 👇
Show more
The best model might not win. Neither might the best chip.
Peter DeSantis, SVP, Foundational AI Models, Custom Silicon, Quantum Computing, @amazon, says models and chips must anticipate each other years ahead. A wrong model roadmap can misdirect hardware. A wrong silicon roadmap can constrain models.
Which is more fatal in your opinion?
👉 Getting the model roadmap wrong
👉 Getting the silicon roadmap wrong
Drop your take below. 👇
Show more
Humans can't be in 2 places at once. Agents can.
Peter Steinberger shared a glimpse of the future: an agent that stays in a meeting, spots a gap, clones itself to investigate, and keeps listening at Agentic AI Summit @UCBerkeley
Never having to choose between listening and working.
So, when autonomous sub-agents start multiplying everywhere…
Are we heading into a golden age of hyper-leverage, or are we about to be completely buried in supervising our AI’s work?
Drop your take below. 👇
Show more
