@Nithum
Joined April 2009
  • Tweets31
  • Following289
  • Followers236
Check out our most recent Explorable "Can Large Language Models Explain Their Internal Mechanisms?" pair.withgoogle.com/explorab…

Can Large Language Models Explain Their Internal Mechanisms?

An interactive introduction to Patchscopes, an inspection framework for explaining the hidden representations of large language models, with large language models.

pair.withgoogle.com
This Post is from an account that no longer exists. Learn more
7
188
Can large language models (LLMs) explain their internal mechanisms? Check out the latest AI Explorable on Patchscopes, an inspection framework that uses LLMs to explain the hidden representations of LLMs. Learn more → goo.gle/patchscopes
A visual walkthrough of the patching process for explaining hidden representations of LLMs.

ALT A visual walkthrough of the patching process for explaining hidden representations of LLMs.

12
146
8
559
49,633
While large language models appear to have a rich understanding of the world, how do we know they’re not simply regurgitating from training data? Check out the latest AI Explorable on a phenomenon called grokking to learn more about how models learn. → goo.gle/45ohnQh
35
459
40
1,754
318,898
Do Machine Learning Models Memorize or Generalize? pair.withgoogle.com/explorab… An interactive introduction to grokking and mechanistic interpretability w/ @ghandeharioun, @nadamused_, @Nithum, @wattenberg and @iislucas
20
243
29
1,135
256,782
Some of my thoughts on generative AI... and a reboot of the PAIR blog... medium.com/people-ai-researc… #responsibleai #hci #machinelearning #GenerativeAI

Meet the new co-leads of PAIR: Lucas Dixon and Michael Terry

Back in 2017, we announced the launch of PAIR by stating, “We believe AI can go much further — and be more useful to all of us — if we…

medium.com
1
1
13
1,457
Confidently Incorrect Models to Humble Ensembles by @Nithum, @balajiln and Jasper Snoek pair.withgoogle.com/explorab…
1
7
2
30
3,936
ML models sometimes make confidently incorrect predictions when they encounter out of distribution data. Ensembles of models can make better predictions by averaging away mistakes. pair.withgoogle.com/explorab…

From Confidently Incorrect Models to Humble Ensembles

ML models sometimes make confidently incorrect predictions when they encounter out of distribution data. Ensembles of models can make better predictions by averaging away mistakes.

pair.withgoogle.com
2
79
In partnership with @GoogleMagenta, we invited 13 professional writers to use Wordcraft, our experimental LaMDA-powered AI writing tool. We've published all of the stories written with the tool, along with a discussion on the future of AI and creativity. g.co/research/wordcraft

Wordcraft Writers Workshop

The Wordcraft Writers Workshop is a collaboration between Google's PAIR and Magenta teams, and 13 professional writers. Together we explore the limits of co-writing with AI.

wordcraft-writers-workshop.appspot.com
3
16
4
49
Most machine learning models are trained by collecting vast amounts of data on a central server. @nicki_mitch and I looked at how federated learning makes it possible to train models without any user's raw data leaving their device. pair.withgoogle.com/explorab…
26
7
63
🤔 We've come a long way with #NLP, but what have language models actually learned? Watch Senior Software Engineer at Google PAIR, Nithum Thain, discuss AI language model learnings → goo.gle/3HVtolv
15
58
Check out our new explorable on machine learning calibration: Machine learning models express their uncertainty as model scores, but through calibration we can transform these scores into probabilities for more effective decision making. pair.withgoogle.com/explorab…

Are Model Predictions Probabilities?

Machine learning models express their uncertainty as model scores, but through calibration we can transform these scores into probabilities for more effective decision making.

pair.withgoogle.com
32
6
117
Beautiful "RNN with attention" tutorial from one of the authors of Google's troll-fighting AI @Nithum. github.com/conversationai/co…. We presented this toxic comment detection model together in the "Tensorflow and modern RNNs without a PhD" talk. Excuse our French 🤬!
14
35
Replying to @martin_gorner @Devoxx
My co-speaker for this session will be @Nithum from Google @JigsawTeam. He fights bad behavior online with neural networks.
1
1
4
"Tensorflow and deep learning without a PhD" continues @Devoxx on Monday 9:30. Deep learning novices welcome, fresh neurons required :-)
4
56
2
103
Awesome work by @adamscj and @nithum! wired.com/2017/02/googles-tr… via @WIRED

Now Anyone Can Deploy Google’s Troll-Fighting AI

Google subsidiary Jigsaw is now offering developers access to an API for its AI-based detector for abusive comments.

wired.com
2
1
Nithum retweeted
Jigsaw
@Jigsaw23 Feb 2017
Introducing Perspective, using machine learning to improve discussions online. bit.ly/2lIZEjS – at Manhattan, NY
5
81
33
149
Nithum retweeted
Jigsaw
@Jigsaw7 Feb 2017
We collected and labeled over 1 million @Wikimedia page edits to determine where personal attacks were made. bit.ly/2lgWfcB – at Manhattan, NY
3
17
1
24
Nithum retweeted
Jigsaw
@Jigsaw19 Oct 2016
How can we keep extremists from using technology to cause harm? @POTUS and @WIRED asked our very own @yasmind. bit.ly/2eCOxa8
1
24
2
34
Wikipedia building n-gram models to detect personal attacks and harassment: meta.m.wikimedia.org/wiki/Re… xitter.cf/wikiresearch/status/75… – at San Mateo, CA

Research:Detox - Meta-Wiki

meta.wikimedia.org
WikiResearch@WikiResearch20 Jul 2016
wikidetox.appspot.com: a demo of algorithmic classification of personal attacks on Wikipedia talk pages
1
3
4
Detecting personal attacks on Wikipedia: some context from the 2015 survey meta.wikimedia.org/wiki/Rese…
3
1
2
Load more