We need more events like this. It is vital to debate and understand the impact of AI on Liberty. @Brendan_McCord is leading the way
Maybe next time at the Acropolis? Or Independence Hall
This image is sacred to me.
This weekend we held the 2nd Cosmos Feast at the chateau of Alexis de Tocqueville. This is the Norman aristocrat who wrote in the 1830s the greatest book on democracy.
Tocqueville feared we would lose the vigorous use of our capacities and slip into passivity and dependence. A great “tutelary power” would be raised above us, keeping us in a state of permanent childhood.
This is my greatest fear for AI.
I lectured on the future of freedom in front of the hearth where Tocqueville wrote his masterwork.
And as I started to leave the room to reunite with my wife and kids at its conclusion, one of Alexis’ descendants put his hand on my arm.
He said, “Alexis was watching you.” And then pointed up to the portrait in the corner of the room. On the train back to Paris, this popped up on WhatsApp.
Between Tocqueville’s gaze, the love in Adriane’s eyes from the front row, and the presence of 70 of the most thoughtful people… people who descended on Normandy from around the world to spend a weekend communing with one of my favorite thinkers…
I’ll treasure it forever.
Congratulations @andrewdfeldman , Sean , Andy , Natalia, and the rest of the Cerebras team on this well earned success . There have been less than 5 companies with an IPO of this scale in the last 2 decades, and it is no doubt due to Andrew’s steadfast leadership and conviction
Convolutions, as I have maintained, have significant benefits in data-efficiency, arbitrary resolution processing, generalization, in vision models, and have been overlooked in the scaling regime.
The addition of a few conv-layers can enable arbitrary resolution handling, removal of positional encoding, removal of ROPE, etc
Moreover, the masking problem has many solutions (eg with ConvMAE https://t.co/MnndThM6dG and as proposed by the paper below) and it's relatively easy to get high performance out of lightweight convs.
it's refreshing when two different hypotheses i've been excited about get validated in a single paper.
tl;dr: convolutional inductive biases in early stages of visual processing, and latent prediction of global semantic features from local spatial context, can both aid in achieving higher sample efficiency on visual tasks.
7 years ago, I was asked by someone from OpenAI not to release the checkpoint from our Transformer-XL OSS effort because the weights might be "too dangerous". Today's argument against open source is that "the Chinese might win" (https://t.co/hXyXtIuSV0)
Our greatest weakness is a failure to build for the long term
It’s difficult because it first requires an understanding of what the world Should be
Not slowness, but rapid motion towards necessary, if immediately unpalatable, objectives
We must build for the long term
I had some time to think recently about plans for the future and found myself working on woefully short term horizons. We've reprogrammed the world to think barely 5 feet in front of our face and i think there is increasingly a huge edge in taking the long term view.