Jason Lin

ML PhD, language models @GoogleDeepMind. NLP @Stanford. deep learning @Lyft @GoogleX @Spotify @Apple. 30 hackathons🏅

I train mixture of experts

[Jason Lin @realJLin]

@SlackHQ @kschzt can this feature (also send as DM) be enabled by default yet?

[Jason Lin @realJLin]

@Lakers BALR - bron, AR, Luka, Rui

[Doodle Dad @WickeyandLeia]

@elonmusk Always Know who Controls Who….

[AK @_akhaliq]

Explore, Establish, Exploit: Red Teaming Language Models from Scratch

paper page: https://t.co/aRR7HMcYH7

Deploying Large language models (LLMs) can pose hazards from harmful outputs such as toxic or dishonest speech. Prior work has introduced tools that elicit harmful outputs in order to identify and mitigate these risks. While this is a valuable step toward securing language models, these approaches typically rely on a pre-existing classifier for undesired outputs. This limits their application to situations where the type of harmful behavior is known with precision beforehand. However, this skips a central challenge of red teaming: developing a contextual understanding of the behaviors that a model can exhibit. Furthermore, when such a classifier already exists, red teaming has limited marginal value because the classifier could simply be used to filter training data or model outputs. In this work, we consider red teaming under the assumption that the adversary is working from a high-level, abstract specification of undesired behavior. The red team is expected to refine/extend this specification and identify methods to elicit this behavior from the model. Our red teaming framework consists of three steps: 1) Exploring the model's behavior in the desired context; 2) Establishing a measurement of undesired behavior (e.g., a classifier trained to reflect human evaluations); and 3) Exploiting the model's flaws using this measure and an established red teaming methodology. We apply this approach to red team GPT-2 and GPT-3 models to systematically discover classes of prompts that elicit toxic and dishonest statements. In doing so, we also construct and release the CommonClaim dataset of 20,000 statements that have been labeled by human subjects as common-knowledge-true, common-knowledge-false, or neither.

[Rowan Cheung @rowancheung]

AI-generated QR codes are the next big thing in marketing.

And it's completely free.

Here's my full tutorial on how to do it in 5 simple steps:

(Try scanning the image below)

[Jason Lin @realJLin]

Pixel 7a + tablet + watch + Nest < Pixel Fold 😂

[Jason Lin @realJLin]

@arjie @paulg Apple has been doing right by the customer in reverting product changes unpopular with the public i.e. butterfly keyboard. Admittedly slower at times, they do not speak it but use action to show commitment to user feedback.

[Jason Lin @realJLin]

Saw ChatGPT plugins coming in 2015. Next step? Own your vision

[Mark Tenenholtz @marktenenholtz]

I've spent 1000's of hours building ML models over the last couple of years.

Here are some tips that would have made me work 10x faster (that you can read in 2 minutes):

[Jason Lin @realJLin]

@elonmusk not at this point. basic ideas inherited from google papers, incredible engineering though. until you do science research, I admire the spirit

[Jason Lin @realJLin]

just finished Tyranny of Merit & picked up How not to be Wrong. will get back to biweekly publishing once time permits, but a preview:
* Raising capital is just the first step; deploying capital is hard
* Op-ed: America’s Olympic disappointment & the hubris of college athletics

[Jason Lin @realJLin]

second article on blockchain, bitcoin and why decentralization deserves your attention.

P.S. will try to get a (bi)weekly publication going, follow my medium to stay tuned!

https://t.co/Aw2KjVSIU5

[Jason Lin @realJLin]

First dabble in publishing & lessons learned running an early-stage startup. Stay tuned for more! #founderwellness @justinkan⁩
https://t.co/lu26OqcrCo

[Jason Lin @realJLin]

@gkimbwala @she_256 @megdeemeh So great to hear your impact with D&I in crypto! (imo it really needs more representation)

[Jason Lin @realJLin]

How to represent part-whole hierarchies in a neural network: peeking into the future of transformer-based vision systems https://t.co/Aw2KQJtyXb by @geoffreyhinton

[Jason Lin @realJLin]

“Consider instead what would be possible if these brands connected the data they have from the production process all…” https://t.co/iz1T7pNovf

[Andrej Karpathy @karpathy]

The equivalent of open source in Software 2.0 land are open datasets. But while plenty of former exists little of high quality latter does.

[Andrej Karpathy @karpathy]

Impressive and surprising https://t.co/9QpeWbhQFU I use those words sparingly

[Jason Lin @realJLin]

Honestly, unless there’s a vaccine or cure, how are we going to have multiple “waves” of coronavirus. Unless a society is completely immobile & closed off from rest of the world, there’s always a non-zero chance a carrier imports the virus, so we’re forever in the big first wave.

[Jason Lin @realJLin]

Does World Bank's latest paper (https://t.co/IB5Qp0qdwx) on China’s productivity slowdown hint a need to curtail its foreign debt load rather than deepening capital towards non-productive GDP growth?

[Jason Lin @realJLin]

Stoked for what the future holds in #QuantumBlueprint Launch to the Future: Quantum Internet — The Department of Energy & UChicago https://t.co/uUw0kC5XD9