🚨ATTN Natural Stories users!🚨
We found a misalignment in the self-paced reading times. Everything is off by one position. In the released dataset, the SPR RTs for the word at index t are actually for index t+1. If you are using the dataset, please use the realigned data.
Once, we ran a study on Prolific and a participant wrote on Reddit that the study “Felt like I was losing the will to live.” I went on the Prolific Subreddit (24k members!) and asked what matters. Here is what they told me. A thread on happier participants and better studies 1/9
@EvelinaLeivada@LanguageMIT@lexfridman I know the best models now are all chat focused rlhf etc. but at least for those trained on a pure language modelling objective, wouldn’t a better way to access their “theory of what’s English” be to compare probabilities of input strings?
Now that the eclipse is within the short range, global ensembles add less value while more high resolution models become available.
I switched my eclipse page to use the High Resolution Ensemble Forecast (HREF), including a cloud layer breakdown:
https://t.co/ygXzAsTEhf
It was an amazing experience to film our music video at the beautiful historic @est8ofmind Oakhurst Manor back in November. Very sadly, I learned that they suffered a devastating fire there on March 1st. Thankfully and most importantly no one was injured. But they are - (1/2)
The video for Dancing With A Shadow is out now.
Starring: Katherine Bickford & Jacob Louis Hoover @postylem
It was an incredible experience working with a great cast and crew. Stephanie Houten (Director) and Dara Nicole Capley (Choregrapher), thank you.
https://t.co/4HxodATsaj
Need a reminder about how to report effect size or confidence intervals for your frequentist stats model? This looks very promising resource: https://t.co/xXeaNyjlkF
Now out in @PNASNews! Large-scale reading evidence that next-word predictability effects in humans are driven by *inference* (logarithmic in predictability) rather than preactivation (linear in predictability). https://t.co/qZLNXp7FQh
1/10
Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.
@SNEWxCenter is that temperature colorscheme reversed? or better for celsius at least…. got me confused for a bit expecting red to be warmer and blue colder.
@miniapeur Incompleteness (Gödel's 2nd). It's both amazing and surprising, and frustrating, and fascinating again, and then just perfect that it can even be proven. And after that roller coaster it feels just so right.