No, what I was trying to say was, try to learn and generalize something from how a smart mind could violate what you thought were the underlying axioms of the computing system to get at you, if it happened before Rowhammer had been discovered (by humans).
SAFETY
-
Open Models Safety Through Iteration and Error Discovery
By
–
This is one of the best arguments for open models. In practice technologies are rarely if ever made completely safe from the start, but rather through finding errors and iterating. It’s best if those errors are relatively small.
-
Powerful Agents Optimize Against Lossy Approximations
By
–
They're lossy approximations, and we should expect more powerful agents to expend compute on avoiding the losses.
-
LLMs capabilities limits and security risks balance
By
–
YES!
One can believe that LLMs can do amazing things and are useful, *without* believing they are anywhere close to human-level intelligence (even if they are superior to humans in a few tasks). One can believe that LLMs will give new tools to people with bad intention *without* -
OpenAI’s Vision: Audio, Retrieval, Safety, Education, and GPT-5
By
–
I’m imagining things like:
– audio (90% of my inputs are via voice) – retrieval, search – safety techniques – education initiatives, funding – multimodal mobile – early GPT-5 demo
– mini models / easier few shot learning – multi person collab – chip dev’t
– new frontends -
Pandemics and AI risks are fundamentally different phenomena
By
–
Some people are assimilating the risks associated with pandemics with the risks associated with AI progress.
There is a fundamental difference. Pandemics are natural phenomena.
We can take preventive measures before they happen and plan for corrective measures in case they do. -
AI Defense Techniques Mirror Offensive Capabilities
By
–
That's pretty much true for any nefarious use of AI: if bad actors can use AI for something, the same AI techniques can be used to protect against them.
-
AI Safety Summit at Bletchley Park: International Cooperation Era
By
–
Check out the latest article in my newsletter: AI safety summit at Bletchley Park: A New Era of International Cooperation in AI Safety? https://
linkedin.com/pulse/ai-safet
y-summit-bletchley-park-new-era-international-fabio-moioli-ydief
… via @LinkedIn -
Governments Must Establish VR Safety Councils Urgently
By
–
governments should discuss VR safety councils before this gets out of hand
-
Regulation Risk Trade-offs in AI Policy
By
–
No, if advocating for regulation makes the risk greater, we shouldn't do it.