Ehud Karavani
@ehudkar
Followers
153
Following
89
Media
74
Statuses
319
Machine Learning ๐ค๐. Causal Inference ๐ดโ๐ โ๐ก. Creator of ๐ฒ๐๐๐๐๐๐๐๐. Was active in a pseudonym account, Now at https://t.co/ZJDLboXmUn
Joined July 2016
I'll be speaking tomorrow at @PyData TLV about causal inference - providing a conceptual overview and briefly touring causallib. Last few tickets remaining. grab yours at https://t.co/jJAVnJywCZ
0
0
3
Knowledge comes from falsifying hypotheses, constraining on what cannot be true (and limiting the flexibility of imagination). Yet deep learning engineers think just pouring more data "and letting the model figure it out" gonna solve it. Are we surprised LLMs hallucinate?
0
0
1
This causal inference podcast episode is really good in how @ehudkar talks about how science, machine learning and causal inference interact.
podcasts.apple.com
Podcast Episode ยท Causal Bandits Podcast ยท 06/17/2024 ยท 54m
1
1
1
To conclude, LLMs' scientific value is over-hyped but that are still useful if you understand their limitations and think critically. Namely, they are a new, natural, and convenient interface to query the Internet, which will only improve as factuality will be improved.
0
0
0
LLMs might be inefficient query engines, but so is asking Siri; natural language itself is quite inefficient - but it's natural. And conversing with the Internet is a convenient way for humans to surf the Internet.
1
0
0
Yep. Sorry you've put billions in engineering to finally make me value chatbots. But they are. See how diverse are the people interacting with LLMs. Put aside many lack the literacy to critic the correspondence, the fact they can pull it off technically is no less than amazing.
1
0
0
This eventually brings me to where LLMs shine the best - they are an extremely intuitive user interface.
1
0
0
In the episode I said LLMs could be great for evidence triangulation, but it is clear to me know that I treated them as knowledge bases, which they are not (yet). https://t.co/VNCvlPh53h
1
0
0
And you can extract knowledge from it - if you're careful. Namely, LLMs are good to combat the blank page block, be it writing a paper or cooking up a DAG. They cannot be fully trusted with a final output, but it's a great conversation starter ๐ https://t.co/SdtWq2UCgG
haha, using ChatGPT to bootstrap an initial causal graph to later be critiqued and refined by an expert is just the manifestation of Cunningham's Law: "The best way to get the right answer on the Internet is not to ask a question; it's to post the wrong answer"
1
0
0
But that doesn't mean LLMs aren't useful, which I think is clearly evident to anyone with functioning senses. Overfitting the entire Internet can be useful.
1
0
0
*(call it "hallucinations", but LLMs just extrapolate badly where support is sparse, do you remember early Vision models were "dreaming"?)
1
0
0
They are not grounded with facts* so they can't make valid arguments. They have no formal logic, so they can't deduce sound arguments. And they generalize poorly, which is of utmost importance for true scientific discovery.
1
0
0
What an excellent opportunity to revise my answer and introduce some more nuance to it. First, admittedly my prior is often hype-averse so it is fair to say I'm a LLM skeptic at the moment, and I don't believe they currently hold much *scientific* value.
1
0
1
I'm humbled by Alex asking me to interview and thankful for him persisting (I had to cancel our first try due to child-induced Covid ๐
). Super excited about this Monday!
0
1
4
Everyone is very excited for their new bedtime story ๐ with its colorful drawings and excellent content, all by the wonderful @AleksanderMolak
0
0
1
haha, using ChatGPT to bootstrap an initial causal graph to later be critiqued and refined by an expert is just the manifestation of Cunningham's Law: "The best way to get the right answer on the Internet is not to ask a question; it's to post the wrong answer"
0
0
0
All in all, I think that's a weak criticism of observational studies. And that there are better ones out there. A personal favorite is this one: https://t.co/KvmzWdcSY8
@yudapearl @robertwplatt ...and sometimes those eligibility requirements which make representativeness difficult in RCT's are in fact quite critical to ensuring validity (in a causal sense) of the result - representing a narrower target population than "everyone" to be sure - but we do this for a reason!
0
0
0
I cannot disagree with an assessment saying it is easier to produce a garbage observational study than a garbage RCT. But this is exactly why target trials should be encouraged as a framework for observational studies. https://t.co/hK0DLCnIvL
2
0
0
And I believe Frank is too harsh on observational studies, and being so completely dismissive of them is counterproductive by discouraging those who actually try to make them better. https://t.co/fJI44TpeOy
@RetractionWatch Very nice. It may be making the assumption that all journals care equally about quality. Journals publishing mainly observational studies may not care as much as journals mainly publishing experimental work.
2
0
0