Best of Twitter - September 26- October 8, 2022

Davis Kedrosky @dkedrosky

Are autocracies better at delivering on growth promises? Not really—a new paper soon out in the JPE argues that the GDP stats are very extremely fake.

Davis Kedrosky @dkedrosky

Raw GDP figures suggest that autocracies grew by 76%, not 147%, over 2002-2021, much more in line with the democracies, who barely lied at all. Only 1 of the top 10 fastest-growing countries on raw data was "free", versus 5 of the top 10 when adjusted for satellite imagery.

Stilicho @StilichoReads

"Russia losing in Ukraine, protestors overthrowing the government of Iran . . . did your faith waver, anon?"

TradingView @tradingview

Currency strength over the last year:

fabians.eth @fabianstelzer

Synth, Don't Search this simple gpt3( ) google sheets function generator has saved me HOURS since I got it up and running last week compare below approach with the process of googling an answer 🤯 I'll eat my macbook if all this doesn't measurably increase global productivity

Flo Crivello @Altimor

Language models have become so good that our benchmarks can’t keep up. It’s like you were looking to measure the rapidly rising temperature in a room, but your themormeters break faster than you can improve them.

Tanya Goyal @tanyaagoyal

✨New preprint✨ Zero-shot GPT-3 does *better* at news summarization than any of our fine-tuned models. Humans like these summaries better. But all of our metrics think they’re MUCH worse. Work/ w/ @jessyjli, @gregd_nlp. Check it out here: https://t.co/vmGsMH1joJ [1/6] https://t.co/I8dY7pPkPc

nearcyan @nearcyan

self-programming AI: "a code-generating language model with the ability to modify its own source code is created and then queried to generate new source code for itself; a new iteration of the model is initialized and trained, and the cycle continues" openreview.net/forum?id=SKat5…

nearcyan @nearcyan

It's always nice to look at the benchmarks used in these papers e.g. GSM8K is a dataset of math questions that require 2-8 steps to solve: github.com/openai/grade-s… GPT-3 got 10% of them correct, then 41% correct with arxiv.org/pdf/2205.11916… we're now at 82% with the above model!

Alexey Guzey @alexeyguzey

been using gpt-3+copilot heavily recently--both surprisingly useful at amplifying my thought+action. thinking broadly now about ways ai could make humans think&ideate better+faster, make better decisions, empower us in general. lmk if you're thinking/tinkering around this stuff

James Cham @jamescham

“Why did Wikipedia, superficially similar and a relatively late entrant, attract a community of millions and build the largest and most comprehensive compendium of human knowledge in history?” @makoshark mako.cc/academic/hill-…