– I haven't engaged with you in years.
– I've always been quite vocal about the limitations of Auto-Regressive LLMs.
– I'm working on ways to lift these limitations, not merely complaining about them.
– Unlike you, I'm not claiming that AR-LLMs will harm society.
@ylecun
-
Beyond Auto-Regressive LLMs: Technical Solutions to Current Limitations
By
–
-
Current AI Systems Face Creativity Limitations
By
–
Another great piece by @Jake_Browning00 on the limitations of current AI systems.
They are boring https://
t.co/6RS6oeWFgt -
Partnership on AI publishes foundation models safety guidelines
By
–
On This Day 6 years ago: the first general meeting of the Partnership on AI took place in Berlin. PAI funds studies and publishes guidelines on questions of AI ethics and safety.
It just published a set of guidelines for the safe deployment of foundation models: -
Engineering Reliable Systems: From Turbojets to AI Innovation
By
–
Yes, like there is a risk that a turbojet will fail.
But we can engineer them to be extremely reliable (not an easy feat), and they couldn't be engineered to be reliable before they were actually invented and built. -
Scary Science Fiction versus Real AI Problem Solutions
By
–
Do you want to read scary science fiction?
Or do you want to find solutions to real problems? -
Constructive Criticism Without Insults Matters
By
–
Criticism is good, valid or not.
Insults combined with assumptions of dishonesty and ill intent are a complete waste of time for everyone. -
AI Dependency on Human Infrastructure and Maintenance
By
–
AI require electrical power, hardware maintenance, software updates, and many other things that they can't get without humans.
-
Designing AI Intrinsic Objectives to Prevent Sociopathic Behavior
By
–
You are right.
You can't control if a human is a sociopath.
But we can make sure an AI system won't be a sociopath because we get to design their intrinsic objectives. -
Designing AI Systems Without Sociopathic Traits
By
–
There are quite a few sociopaths in society.
Most of them aren't smart enough to take physics classes at MIT.
But history shows that the can get elected to the US presidency. We simply won't build AI sociopaths. Unlike with humans, we get to *design* AI systems intrinsic -
AI Alignment Easier Than Aligning Humans Through Design
By
–
No.
But I'm arguing that
a) we are familiar with the concept of aligning intelligent entities, whether they are children, adults, or corporations (which are super-human).
b) it is *much* easier to align AI than humans because we get to *design* the intrinsic motivations of AI