It was Demis or Sundar.
They picked Sundar.
@ylecun
-
Sundar Pichai Selected Over Demis Hassabis for Leadership Role
By
–
-
Drama at OpenAI and Google versus FAIR’s smooth creation
By
–
What this NYT article tells us about the genesis of the new AI industry is that there was a lot of drama at Google, DeepMind, OpenAI, and Anthropic. Comparatively, the creation of FAIR was straightforward and drama free. https://
nytimes.com/2023/12/03/tec
hnology/ai-openai-musk-page-altman.html?smid=nytcore-android-share
… -
AI Leaders and the Battle for Control Over Technology
By
–
I'm on the cover of the weekly Le Point, as some sort of tiny éminence grise pulling the strings behind @elonmusk
, @sundarpichai
, @kaifulee and a larger-than-life Sam Altman. The title reads "AI, the battle to control our brains" I just want to say that I have no intention of -
Google Research and Meta GenAI Publishing Practices Differ
By
–
Yes. For example, large swaths of "Google Research" never publish.
And parts of Meta's GenAI group (which is a product group) do publish. -
Gradient-Based Optimization Superior to Gradient-Free Methods
By
–
No. That's just false.
Whenever you can use it, gradient-based optimization is incomparably more efficient than gradient-free, regardless of hardware.
There are situations where you have no choice but use gradient free, you should make that as rare as possible.
When you do, you -
Gradient-Based Learning Enables Distributed Representation Emergence
By
–
Gradient-based learning causes distributed representations to emerge.
-
Convex Loss Limitations in Machine Learning Theory
By
–
The people mocking "no proof" didn't realize that it was their insistence to have a particular type of theory that was holding them back. They would only work on "trivial" models for which the loss was convex, so they could use fancy semi-definite programming methods and prove
-
ConvNets computational efficiency compared to 2007 object recognition models
By
–
Not true. ConvNets were actually relatively cheap compared to the dominant models of object recognition circa 2007.
They were much smaller than current architectures, of course. -
Academic AI Course Ahead of Industry Trends
By
–
One case where an academic course was ahead of industry trends Now in my course, I talk about SSL, energy-based models, and JEPA….
-
Resources Required: From Invention to Recognition and Impact
By
–
You won't have the resources to make your idea real.
So you won't have a product, and you won't have publications.
The world will never hear about your invention. Without a working product or a peer-reviewed publication, the establishment (academia, industry, investors) will