Introducing a16z Originals: Working Theory.
I sat down with @johnschulman2 and @markchen90 to demystify RSI: AI helping build better AI.
Models can already run experiments. But research taste—the judgment to know which questions are worth asking—remains harder to teach. As John points out, knowing the literature isn’t the same as having been wrong and learned from it.
We also explored how AI could help solve alignment, and why the benefits of RSI don’t depend on fully autonomous research.
I came away optimistic about AI’s potential to accelerate discovery, including solutions to some of the problems it creates, and believe that human judgment remains essential for the foreseeable future.
Huge thanks to Mark and John for joining us.
We cannot defend America’s infrastructure at human speed against attacks moving at machine speed.
In today’s WSJ, I make the case for using AI to protect the systems that keep our water flowing, hospitals operating and lights on.
Here’s the action plan: https://t.co/KsLLstiwFL
- Start with the few hundred operators and shared suppliers whose failure could cascade nationally.
-The White House must set outcome-based goals, like ensuring control systems aren't exposed to the Internet.
- Use AI to find and fix flaws. Contain what cannot be patched. Test whether essential services can keep running under attack.
- CEOs of critical infrastructure firms should be accountable for these outcome-based results.
- The White House, infrastructure operators, AI labs and technology providers each have a role. AI labs, for example, should work with critical infrastructure operators to build digital replicas so high risk patches can be tested offline.
AI can give defenders an advantage. We need to organize to use it.
https://t.co/anGftvo9Dm
The ideas are not new. What wasn’t possible, in our opinion, was the maintenance, managing false positives, and probably the fear of breaking production. Capabilities have improved enough that a small model can decide inline where a signature used to, or write live patches while managing the version hell. You were and still are right about that gap, and we think the models can now close it.
Small cities, school districts and rural hospitals guard our most sensitive data with the fewest resources to defend it.
Today we’re launching Securing America’s Infrastructure Foundation, a new nonprofit, bringing AI-powered cyber defense to them at no cost.
SAIF starts with focused goals for on the ground impact and $25M in committed funding by a16z's founders. I’m joined by a founding board of Brian Barrios, @GoldsmithOnGov, Bob Lord, @vishalmisra and @philvenables.
If you lead a small city, rural school district or hospital system or want to join our team, please reach out: [email protected]
Read our full announcement here: https://t.co/Vb1WlpuQYV