Twitxy Saved

Sol Messing

Verified account This account is verified. Learn more

@SolomonMg

AI eval science & AI-augmented work @GoogleDeepMind, @NYU @CSMAP_NYU. Past: @Twitter, Acronym, @Facebook, @PewResearch, @Stanford.
2019 solomonmg.github.io Joined September 2009
1.9K Following 7.1K Followers 5.4K Posts

Latest posts 5

Sol Messing Verified account @SolomonMg
Very nice Pangram validation. Tho in fairness pre-2020 may mean in the training data unless you can find text that wasn’t on the internet.
Emu Brigadier Verified account @emubrigadier 3d
Right, mates.

I really tried to go for this Pangram bounty.

I spent $180 checking all kinds of pre-2020 sources: blogs, recipes, essays, sermons, news articles, short stories, textbooks.

I seriously considered the criticism from African influencers that Pangram penalises "African English" : out of my 599 texts, 285 texts came from Africa, from 14 countries, and aslo pan-African culture and literary sites.

The false positive rate was literally zero.

ZERO.

As in: not only NONE of the texts scored more than 75% AI, none of them scored more then 0% AI.

🤷‍♂️

P.S. This was not part of the experiment, but the only pre-2020 texts Pangram scores as AI are the archived outputs of GPT-2 (a precursor of ChatGPT and modern AI models), which is fair enough.
View quoted post on X
1 replies 0 reposts 1 likes 1.5K views View on X