I trained a 900M parameter model expecting better results.
It performed worse than my 434M model.
Here's what I learned winning the Data4Good AI Challenge (99.11%, 1st place out of 300+ teams):
🧵
#MachineLearning#AI#NLP
@JHUCarey@teppercmu@BerkeleyHaas@SASsoftware@INFORMS Three counterintuitive lessons:
1. Bigger ≠ Better (434M beat 900M)
2. More models ≠ Better (1 beat 4)
3. LLMs as arbitrators, not classifiers
Full analysis here (upvotes appreciated!):
👉 https://t.co/ERMoSkQP9f
What's surprised you about model performance?
@jriddlecult@beatmastermatt is now seeming to be right winged. In the words of Reagan, I didn't leave the Democratic Party, the Democratic Party left me.
@jriddlecult@beatmastermatt radically right, the Republican Party still advocates for smaller less activist law and order abiding government and has been since Nixon. So, the right has not shifted so much as the left has pushed so far to the left on social axis (some good) that someone previously centrist
@jriddlecult@beatmastermatt Yeah, and Obamacare is still the law of the land when it comes to healthcare. Mexican border is as porous as it was before Trump took office. Iraq war led to ISIS and Afghanistan is under the Taliban rule (just as we found it). So, neither did Republican rhetoric materialize. WYP