Zoom Logo

MIT GOV/LAB: Evaluating 20,000+ World Bank Development Projects - Shared screen with speaker view
Haihao
49:34
Wouldn’t intuition suggest better performance with at review than at approval?
Haihao
56:18
What about Switch Transformers?
Alisa Zomer
01:08:41
Please add any questions in the chat. Thanks!
Justin Shenk
01:09:40
Switch transformer model weights aren't yet released by Google, but would be cool!
Anum Mustafa
01:11:09
Is it fair to interpret your results as saying that development agencies can focus more on the more important features you find (e.g. abstracts and reviews) to determine project success? Assuming the people game the system problem doesn’t exist. If this is an oversimplification, how is it so?
Julia Seifert_REDEFINE
01:11:17
Potentially multiple WB projects are implemented in the same sector in a single country at the same time - how may have this influenced the outcome variable?
Justin Shenk
01:12:05
i cant unmute
Haihao
01:12:11
Me neither
Haihao
01:15:19
Is there any intuition WHY the larger models are actually more interpretable? Would’ve thought more parameters/embedding dimensions might make it more difficult.
Laura Ahearn
01:16:23
I’m also a big fan of “Don’t Build It” — thanks for that! I’m wondering if emphasis placed in that piece (as I recall) on local ownership and designing with users, etc., could be captured in this analysis. In other words, can you say anything about whether there’s a correlation between locally led development and positive outcomes?
Haihao
01:17:00
And could you share a few practicalities on your current implementation, e.g. training time, inference speed, GPU power/memory, maybe cloud TPUs?
Alisa Zomer
01:18:33
(FYI, here’s a link to Luke’s Don’t Built It guide in case you want to check it out later: https://mitgovlab.org/resources/dont-build-it-a-guide-for-practitioners-in-civic-tech/)
Julia Seifert_REDEFINE
01:23:53
One more question - since documentation appears paramount. Might some regions be disadvantaged due to poorer documentation
Julia Seifert_REDEFINE
01:24:31
Thank you Lily
Haihao
01:25:33
Thank you all!