Gavia Gray
@reflected.codes
Member of Technical Staff at Cerebras (views expressed my own) (also technically prof at Dalhousie University)
https://github.com/gngdb
https://reflected.codes/
https://github.com/gngdb
https://reflected.codes/
Reposted by Gavia Gray
With Google removing their Responsible AI Principles, they no longer state that they will *not* engage in "Technologies whose purpose contravenes widely accepted principles of international law and human rights".
Concerns about surveillance and injury are also erased.
ai.google/responsibili...
Concerns about surveillance and injury are also erased.
ai.google/responsibili...
February 4, 2025 at 7:58 PM
With Google removing their Responsible AI Principles, they no longer state that they will *not* engage in "Technologies whose purpose contravenes widely accepted principles of international law and human rights".
Concerns about surveillance and injury are also erased.
ai.google/responsibili...
Concerns about surveillance and injury are also erased.
ai.google/responsibili...
Reposted by Gavia Gray
Folks are mostly missing the forest for the trees re: the recent Meta announcement. Specifically, the focus on the move from fact checking to community notes. There two different changes that are a much bigger deal that have received less focus. 🧵 1/11
January 9, 2025 at 8:29 PM
Folks are mostly missing the forest for the trees re: the recent Meta announcement. Specifically, the focus on the move from fact checking to community notes. There two different changes that are a much bigger deal that have received less focus. 🧵 1/11
I got Claude to condense github.com/mattjj/autod... to a single file to make it more readable: gist.github.com/gngdb/a6bdf4...
It's the same design as the core Jax autograd engine as far as I know.
It's the same design as the core Jax autograd engine as far as I know.
Generated condensed version of https://github.com/mattjj/autodidact
Generated condensed version of https://github.com/mattjj/autodidact - autodidact.py
gist.github.com
December 20, 2024 at 4:49 PM
I got Claude to condense github.com/mattjj/autod... to a single file to make it more readable: gist.github.com/gngdb/a6bdf4...
It's the same design as the core Jax autograd engine as far as I know.
It's the same design as the core Jax autograd engine as far as I know.
Our poster is up in the compression workshop for Esha Singh's paper on bounds for scaling sparsity. We investigated better ways to compare sparsity inducing regularizers by grounding to compute efficient dense models. openreview.net/forum?id=qOn...
Empirical Upper Bounds for Unstructured Sparsity in...
Sparsity in deep neural networks promises two improvements in computational efficiency, fewer FLOPs spent both to train the network and to perform inference. We find that both may be quantified...
openreview.net
December 15, 2024 at 5:36 PM
Our poster is up in the compression workshop for Esha Singh's paper on bounds for scaling sparsity. We investigated better ways to compare sparsity inducing regularizers by grounding to compute efficient dense models. openreview.net/forum?id=qOn...
Reposted by Gavia Gray
How do a neural network's final parameters depend on its initial ones?
In this new paper, we answer this question by analyzing the training Jacobian, the matrix of derivatives of the final parameters with respect to the initial parameters.
https://arxiv.org/abs/2412.07003
In this new paper, we answer this question by analyzing the training Jacobian, the matrix of derivatives of the final parameters with respect to the initial parameters.
https://arxiv.org/abs/2412.07003
December 11, 2024 at 8:30 PM
How do a neural network's final parameters depend on its initial ones?
In this new paper, we answer this question by analyzing the training Jacobian, the matrix of derivatives of the final parameters with respect to the initial parameters.
https://arxiv.org/abs/2412.07003
In this new paper, we answer this question by analyzing the training Jacobian, the matrix of derivatives of the final parameters with respect to the initial parameters.
https://arxiv.org/abs/2412.07003
I'll be presenting our poster on efficient estimation of critical batch size today at 4:30pm at NeurIPS 2024 neurips.cc/virtual/2024...
NeurIPS Poster Normalization Layer Per-Example Gradients are Sufficient to Predict Gradient Noise Scale in TransformersNeurIPS 2024
neurips.cc
December 11, 2024 at 6:04 PM
I'll be presenting our poster on efficient estimation of critical batch size today at 4:30pm at NeurIPS 2024 neurips.cc/virtual/2024...
Reposted by Gavia Gray
I would put this even more strongly: open source AI is probably our only realistic chance to avoid a terrifying increase in concentration of power. I do not want to live in a world where the people with all the money also have all the intellectual power.
The most realistic reason to be pro open source AI is to reduce concentration of power.
"money has flowed to tech giants and others in their orbit... [and] raises an uncomfortable prospect: that this supposedly revolutionary technology might never deliver on its promise of broad economic transformation, but instead just concentrate more wealth" www.bloomberg.com/opinion/arti...
November 29, 2024 at 9:35 PM
I would put this even more strongly: open source AI is probably our only realistic chance to avoid a terrifying increase in concentration of power. I do not want to live in a world where the people with all the money also have all the intellectual power.
Reposted by Gavia Gray
A feature/bug of BlueSky is that you can back-date posts. People are using this to repost their tweets.
If you search for posts from the 1st century, there are jokes about the resurrection of Jesus. And, perhaps even funnier, some posts dated 24-11-20 instead of 2024-11-20.
bsky.app/search?q=unt...
If you search for posts from the 1st century, there are jokes about the resurrection of Jesus. And, perhaps even funnier, some posts dated 24-11-20 instead of 2024-11-20.
bsky.app/search?q=unt...
November 20, 2024 at 4:27 PM
A feature/bug of BlueSky is that you can back-date posts. People are using this to repost their tweets.
If you search for posts from the 1st century, there are jokes about the resurrection of Jesus. And, perhaps even funnier, some posts dated 24-11-20 instead of 2024-11-20.
bsky.app/search?q=unt...
If you search for posts from the 1st century, there are jokes about the resurrection of Jesus. And, perhaps even funnier, some posts dated 24-11-20 instead of 2024-11-20.
bsky.app/search?q=unt...
Reposted by Gavia Gray
when you try to convert your text into smaller pieces but all it gives you is Elvish, that’s a tolkienizer
November 20, 2024 at 5:51 PM
when you try to convert your text into smaller pieces but all it gives you is Elvish, that’s a tolkienizer
Reposted by Gavia Gray
One of the hardest,and most meaningful, projects I do in my job is monitor the epidemic of fatal violence against the trans and gender-expansive community.
Today, on Transgender Day of Remembrance (#TDOR), we honor those we’ve lost over the last year.
reports.hrc.org/an-epidemic-...
Today, on Transgender Day of Remembrance (#TDOR), we honor those we’ve lost over the last year.
reports.hrc.org/an-epidemic-...
The Epidemic of Violence Against the Transgender & Gender-Expansive…
The 2024 Report
reports.hrc.org
November 20, 2024 at 2:03 PM
One of the hardest,and most meaningful, projects I do in my job is monitor the epidemic of fatal violence against the trans and gender-expansive community.
Today, on Transgender Day of Remembrance (#TDOR), we honor those we’ve lost over the last year.
reports.hrc.org/an-epidemic-...
Today, on Transgender Day of Remembrance (#TDOR), we honor those we’ve lost over the last year.
reports.hrc.org/an-epidemic-...
Reposted by Gavia Gray
Good reporting here: www.thetrillium.ca/news/municip...
How a handful of Toronto businessmen got their way on bike lanes provincewide
'Non-partisan group' Balance on Bloor includes a former PC candidate, donors, and real estate industry players
www.thetrillium.ca
November 14, 2024 at 8:52 PM
Good reporting here: www.thetrillium.ca/news/municip...
Reposted by Gavia Gray
Neurips reviews are now publicly available.
Don't forget to check out Open RL benchmark, very useful when implementing algorithms or checking performance/impact of hyperparameters.
openreview.net/forum?id=ZDv...
Don't forget to check out Open RL benchmark, very useful when implementing algorithms or checking performance/impact of hyperparameters.
openreview.net/forum?id=ZDv...
Open RL Benchmark: Comprehensive Tracked Experiments for...
In many Reinforcement Learning (RL) papers, learning curves are useful indicators to measure the effectiveness of RL algorithms. However, the complete raw data of the learning curves are rarely...
openreview.net
November 14, 2024 at 11:01 AM
Neurips reviews are now publicly available.
Don't forget to check out Open RL benchmark, very useful when implementing algorithms or checking performance/impact of hyperparameters.
openreview.net/forum?id=ZDv...
Don't forget to check out Open RL benchmark, very useful when implementing algorithms or checking performance/impact of hyperparameters.
openreview.net/forum?id=ZDv...
Reposted by Gavia Gray
In summary, Ford's bike lane removals will, according to a new City Hall report:
- Cost $75 million in direct costs and sunk costs.
- Make traffic much, much worse for at least nine months during construction to remove lanes
- Have minimal long-term impact on car travel times.
- Cost $75 million in direct costs and sunk costs.
- Make traffic much, much worse for at least nine months during construction to remove lanes
- Have minimal long-term impact on car travel times.
November 13, 2024 at 10:11 PM
In summary, Ford's bike lane removals will, according to a new City Hall report:
- Cost $75 million in direct costs and sunk costs.
- Make traffic much, much worse for at least nine months during construction to remove lanes
- Have minimal long-term impact on car travel times.
- Cost $75 million in direct costs and sunk costs.
- Make traffic much, much worse for at least nine months during construction to remove lanes
- Have minimal long-term impact on car travel times.
Reposted by Gavia Gray
We're cooked. The economy's cooked. The planet is really cooked.
But trans people are especially cooked, here in the U.S.
Because no one's coming to save us:
But trans people are especially cooked, here in the U.S.
Because no one's coming to save us:
November 6, 2024 at 3:42 PM
We're cooked. The economy's cooked. The planet is really cooked.
But trans people are especially cooked, here in the U.S.
Because no one's coming to save us:
But trans people are especially cooked, here in the U.S.
Because no one's coming to save us:
I have just learned that when my dad said "it's like herding cats" he was probably referring to a 1999 superbowl ad www.youtube.com/watch?v=m_Ma...
EDS, an HP Company 'Cat Herders'
EDS, an HP Company 'Cat Herders'
www.youtube.com
June 14, 2024 at 3:45 PM
I have just learned that when my dad said "it's like herding cats" he was probably referring to a 1999 superbowl ad www.youtube.com/watch?v=m_Ma...
Reposted by Gavia Gray
RIP Lynn Conway, chip design pioneer and transgender-rights advocate
news.engin.umich.edu/2024/06/the-...
From wiki: IBM fired Conway in 1968 after she revealed her intention to transition. IBM apologized in 2020 (!).
en.m.wikipedia.org/wiki/Lynn_Co...
news.engin.umich.edu/2024/06/the-...
From wiki: IBM fired Conway in 1968 after she revealed her intention to transition. IBM apologized in 2020 (!).
en.m.wikipedia.org/wiki/Lynn_Co...
The legacy of Lynn Conway, chip design pioneer and transgender-rights advocate
Conway, professor emerita of electrical engineering and computer science, has died.
news.engin.umich.edu
June 12, 2024 at 12:23 AM
RIP Lynn Conway, chip design pioneer and transgender-rights advocate
news.engin.umich.edu/2024/06/the-...
From wiki: IBM fired Conway in 1968 after she revealed her intention to transition. IBM apologized in 2020 (!).
en.m.wikipedia.org/wiki/Lynn_Co...
news.engin.umich.edu/2024/06/the-...
From wiki: IBM fired Conway in 1968 after she revealed her intention to transition. IBM apologized in 2020 (!).
en.m.wikipedia.org/wiki/Lynn_Co...
To do machine learning research it's important to have a degree in maths so you can read the math in most papers, a degree in computer science so you can write replicable code and a degree in physics so you can talk to all the physicists here. Luckily, getting them will leave no time for research.
May 31, 2024 at 6:23 PM
To do machine learning research it's important to have a degree in maths so you can read the math in most papers, a degree in computer science so you can write replicable code and a degree in physics so you can talk to all the physicists here. Luckily, getting them will leave no time for research.
Years of plotting research yet no use for anything except calling `fig.tight_layout()` and hope.
May 21, 2024 at 7:20 PM
Years of plotting research yet no use for anything except calling `fig.tight_layout()` and hope.
When Karpathy said he can't watch any movie more than 20 years old he was actually using Cunningham's rule to get movie recommendations, obviously.
May 21, 2024 at 12:55 PM
When Karpathy said he can't watch any movie more than 20 years old he was actually using Cunningham's rule to get movie recommendations, obviously.
The fact about polar bear fur being technically transparent is dumb if you think that means you shouldn't say the fur is white but much more interesting if you know that the same transparent structure makes the bears mostly invisible in IR, so night vision would be useless.
May 17, 2024 at 1:37 PM
The fact about polar bear fur being technically transparent is dumb if you think that means you shouldn't say the fur is white but much more interesting if you know that the same transparent structure makes the bears mostly invisible in IR, so night vision would be useless.
Reposted by Gavia Gray
The geomagnetic storm seems to be winding down, so if you still need to communicate across time with your dead father via ham radio to solve your mother's murder, now's the time!
May 13, 2024 at 12:17 AM
The geomagnetic storm seems to be winding down, so if you still need to communicate across time with your dead father via ham radio to solve your mother's murder, now's the time!
Is "back to the old drawing board" just from looney tunes or were drawing boards a universal experience in the 40s?
May 10, 2024 at 1:00 PM
Is "back to the old drawing board" just from looney tunes or were drawing boards a universal experience in the 40s?
It's wild how much better Fiesta Farms is
Happy Loblaws Boycott month!
Remember kids, they own a LOT! You'll have to get clever and seek out independent grocers, and it may be inconvenient and stressful but it will be WORTH it
Remember kids, they own a LOT! You'll have to get clever and seek out independent grocers, and it may be inconvenient and stressful but it will be WORTH it
May 1, 2024 at 8:57 PM
It's wild how much better Fiesta Farms is
Weird side effect of trying to use copilot: when I know what I want to type, I try to type fast enough that it doesn't have time to suggest anything.
May 1, 2024 at 5:39 PM
Weird side effect of trying to use copilot: when I know what I want to type, I try to type fast enough that it doesn't have time to suggest anything.