Smerity (@smerity) 's Twitter Profile
Smerity

@smerity

gcc startup.c -o ./startup. Focused on machine learning & society. Previously @Salesforce Research via @MetaMindIO. @Harvard '14, @Sydney_Uni '11. ๐Ÿ‡ฆ๐Ÿ‡บ in SF.

ID: 15363432

linkhttps://state.smerity.com/ calendar_today09-07-2008 08:00:18

13,13K Tweet

32,32K Followers

2,2K Following

Patrick Walton (@pcwalton) 's Twitter Profile Photo

TIL .NET starts with a fast hash and switches to a DoS resistant one after too many collisions in the default hash table implementation. I wish we had thought of that for Rust; SipHash is a performance footgunโ€ฆ

Dr Kate Compton (@galaxykate) 's Twitter Profile Photo

Capacitive: Remember that when the iPhone came out, we were there at GDC talking about how important the accelerometer was, and maybe also tapping. Not swiping and gesture. That came *years* later. Multifinger, angle and pressure sensitivity come and go. Pen tablets matter!?

Capacitive: 

Remember that when the iPhone came out, we were there at GDC talking about how important the accelerometer was, and maybe also tapping.

Not swiping and gesture. That came *years* later.

Multifinger, angle and pressure sensitivity come and go.

Pen tablets matter!?
Alex Blechman (@alexblechman) 's Twitter Profile Photo

Programming is chaotic magic. There are no rules. You ask a game dev โ€œCan the player summon a giant demon that bursts from the ground in an explosion of lava?โ€ and theyโ€™ll say โ€œsure, thatโ€™s easyโ€ and then youโ€™ll ask โ€œcan the player wear a scarf?โ€ and theyโ€™ll go โ€œoofโ€

Smerity (@smerity) 's Twitter Profile Photo

I made a mistake. I commented on Reddit. Someone immediately accused me of working for Google as I thought TPUs were good ... ๐Ÿ˜… I then made a second mistake. I replied to the person who replied to my comment on Reddit. Third mistake: I then posted to Twitter..? ๐Ÿค”

Mitchell Wortsman (@mitchnw) 's Twitter Profile Photo

Sharing our project on 1) accelerating and 2) stabilizing training for large language-vision models 1) Towards accelerating training, we introduce SwitchBack, a linear layer for int8 quantized training which matches bfloat16 within 0.1 for CLIP ViT-Huge arxiv.org/abs/2304.13013

Sharing our project on 1) accelerating and 2) stabilizing training for large language-vision models

1) Towards accelerating training, we introduce SwitchBack, a linear layer for int8 quantized training which matches bfloat16 within 0.1 for CLIP ViT-Huge

arxiv.org/abs/2304.13013
Tim Dettmers (@tim_dettmers) 's Twitter Profile Photo

Our work on loss spikes and stable 8-bit CLIP training is the largest Int8 training to date (1B). We introduce the SwitchBack layers and StableAdamW to ensure stability at these scales. Work with the awesome Mitchell Wortsman Paper: arxiv.org/abs/2304.13013 Colab: github.com/mlfoundations/โ€ฆ

Our work on loss spikes and stable 8-bit CLIP training is the largest Int8 training to date (1B). We introduce the SwitchBack layers and StableAdamW to ensure stability at these scales. Work with the awesome <a href="/Mitchnw/">Mitchell Wortsman</a>

Paper: arxiv.org/abs/2304.13013
Colab:  github.com/mlfoundations/โ€ฆ
Elissa Harrington (@eharringtontv) 's Twitter Profile Photo

An autonomous vehicle just drove through police tape at a crime scene in San Francisco. ๐Ÿ˜ฌ Now a bunch of cars are driving through the area that was supposed to be sealed off.

An autonomous vehicle just drove through police tape at a crime scene in San Francisco. ๐Ÿ˜ฌ Now a bunch of cars are driving through the area that was supposed to be sealed off.
Smerity (@smerity) 's Twitter Profile Photo

Joshua Achiam โš—๏ธ The field is so full of wonder and beauty that the lies are a straight up detractor and destroy our understanding. Separating the facts of the present from the near yet unfulfilled promises of the future is how we make progress and use our existing wonders correctly.

Moxie Marlinspike (@moxie) 's Twitter Profile Photo

As a kid, the magic of software was that I could sit down and make something with no license, degree, or ~money. Gotta say, publishing an iOS app from scratch today is a verrry diff vibe. I wonder how many young people's ideas we've lost at "and now register for a DUNs number."

Jeremy Howard (@jeremyphoward) 's Twitter Profile Photo

There's a new bill, SB-1047 "Safe and Secure Innovation for Frontier Artificial Intelligence Models Act". I think it could do a great deal of harm to startups, American innovation, open source, and safety. So I've written a response to the authors: ๐Ÿงต answer.ai/posts/2024-04-โ€ฆ

Smerity (@smerity) 's Twitter Profile Photo

We're going to find out how long it takes for a person with a USB to reach every Windows blinkenlights box holding up a corporation's house of cards

Smerity (@smerity) 's Twitter Profile Photo

Just started reading Andy Grove's "Only the Paranoid Survive". The replay from 1994 (the eroding brand trust due to the Pentium floating point bug, half a billion spent replacing CPU stock) in the present (Intel 13th / 14th Gen CPU instability issues) is eery in all the bad ways.

Smerity (@smerity) 's Twitter Profile Photo

On searching for Gaudรญ's use of inverted weights for architecture I discovered the List of Physical Visualizations (List of Physical Visualizations) with such wonders as: - 1915 โ€“ Wire Models of Factory Worker Movements - 1935 โ€“ 3D visualization of a year of power consumption dataphys.org/list/

On searching for Gaudรญ's use of inverted weights for architecture I discovered the List of Physical Visualizations (<a href="/dataphys/">List of Physical Visualizations</a>) with such wonders as:
- 1915 โ€“ Wire Models of Factory Worker Movements
- 1935 โ€“ 3D visualization of a year of power consumption
dataphys.org/list/