– Superintelligence is possible
– it won't happen next year
– it won't have all the drives of human nature
– humans suck at chess and at many other tasks at which computers excel
– until we have a working design for superintelligence (and we don't) banning it for safety reasons
@ylecun
-
Superintelligence possible but not imminent, lacking human drives, says LeCun
By
–
-
Yann LeCun clarifies his role: ran FAIR, not involved in Llama
By
–
I didn't "run Meta's AI initiatives."
I ran FAIR (the fundamental AI research lab) from 2014 to 2018.
Then, I became an IC, and returned to research. I was nobody's boss.
Iwas never involved in LLM research, and made zero technical contributions to Llama. I merely cheered from -
Demucs produced at FAIR-Paris by @honualx
By
–
Demucs was produced at FAIR-Paris by @honualx and collaborators.
-
Human imperfection compared to machine capabilities in various tasks
By
–
If we are so perfect, why are we so bad at arithmetics, computing integrals symbolically, playing chess, go, or poker?
Machines are already more "perfect" than us at these and many other tasks. (Also, why can we die from diseases if we are so perfect?) -
AI will have traits except spirituality; many moral humans aren’t spiritual
By
–
Indeed, AI today does not do or possess any of those things. But at some point in the future they will. Except perhaps for the spiritual part.
Many humans aren't spiritual either, yet have empathy and are highly moral. -
LeCun: Intelligence equals trained world model plus optimal control
By
–
Essentially. Or rather: Trained world model + optimal control = intelligence
-
Sparse Convolutional Autoencoders for Fine-Tuning
By
–
It does work. With stacks of pre-trained sparse convolutional autoencoders, we can achieve a strong starting point for fine-tuning on small labeled datasets (like Caltech 101, which had 30 training samples per category) and reach near-state-of-the-art performance.
-
Revisiting Sparse-Regularized Autoencoders
By
–
Back when we believed deep network pre-training could be achieved using stacked autoencoders trained with group sparsity regularization.
-
Researcher Claims Priority on World Models Concept Since 2016
By
–
I started using the concept in 2016 (e.g. in my NIPS 216 keynote, in which I called it a "world simulator"). I published papers on video prediction in 2016.
This was meant to be a key step to train world models.
Ha&Schmi appeared in 2018. The slide below is from a talk I gave -
Transformers History: From Attention Mechanisms to Modern AI
By
–
Transformers were published in 2017 (by Google), but was based on work on attention and associative memory by University of Montreal and FAIR 2 years earlier. And they use neural nets and backprop that were popularized in the late 1980s. Oh, and GPUs are based on work by Bill