Best of Twitter - Week of May 23, 2022
Large Language Models are Zero-Shot Reasoners Simply adding “Let’s think step by step” before each answer increases the accuracy on MultiArith from 17.7% to 78.7% and GSM8K from 10.4% to 40.7% with GPT-3. arxiv.org/abs/2205.11916
@0xmaddie_ @arankomatsuzaki Future math proof sketches generation be like: "Prove X. Chanelling inner Ramanujan", "Compute Y the way Terry Tao would do", "what thoughts would Grothendieck have on this?"
❗Light-activated mitochondria extend lifespan in C. elegans. This project has been a long time coming, I'm so excited to share 🧵👇 biorxiv.org/content/10.110…
biorxiv.orgOptogenetic rejuvenation of mitochondrial membrane potential extends C. elegans lifespanDespite longstanding scientific interest in the centrality of mitochondria to biological aging, directly controlling mitochondrial function to test causality has eluded researchers. We show that specifically boosting mitochondrial membrane potential through a light-activated proton pump reversed age…
Gato🐈a scalable generalist agent that uses a single transformer with exactly the same weights to play Atari, follow text instructions, caption images, chat with people, control a real robot arm, and more: dpmd.ai/Gato Paper: dpmd.ai/Gato-paper 1/
Indeed great collaboration. We found that ppl diagnosed with schizophrenia have 50-fold increased rate of T>G and CpG transversions in TFBS in early development. biorxiv.org/content/10.110…
Glad to share the latest preprint from my PhD w/ my co-first authors @jones_attila @seplyarskiy! We find enrichment of somatic variants at transcription factor binding sites in schizophrenia w/ specific mutational patterns. In collaboration w/@ChrisAWalsh1 @AndyChessNYC labs🧵1/n https://t.co/DCLQHwr6rU
Eduardo Maury @emauryg
This is shocking. DuckDuckGo has a search deal with Microsoft which prevents them from blocking MS trackers. And they can't talk about it! This is why privacy products that are beholden to giant corporations can never deliver true privacy; the business model just doesn't work.
^ note that first and second columns are “walking” and “cycling”
1/Paxlovid rebound: with 2/2 in my family with Paxlovid rebound, stories rife on Twitter, and 2 completely unscientific polls showing 46-47% rebound, I am thinking about Paxlovid efficacy in terms of relative risk, absolute risk, and Number Needed to Treat (NNT)
An important new study is out. Baseline findings from the NIH's longitudinal, intramural Long COVID study — perhaps the most detailed, controlled, comprehensive investigation of multiple health metrics in this setting thus far conducted — were just published in @AnnalsIM A🧵 1/
The study enrolled 189 mostly non-hospitalized adults post-Covid & 120 controls. Post-acute sequelae of COVID (PASC) were defined as subgroup w/ worsening symptoms/conditions after COVID-19 All participants went to the NIH for an extensive battery of tests, including: 2/
To summarize, there was a large burden of symptoms & worse mental health post-COVID (particularly among those with PASC) but no difference in objective neurocognition, & no evidence of "ongoing systemic inflammation or immune activation" or any organ damage. Some caveats: 13/