jeremyrmanning.bsky.social
Context Lab (@contextlab.bsky.social) director, Dartmouth prof, memory & 🧠 network modeler, data scientist, dad x2, husband, tree hugger 🌲, & 🧁+🍪 baker
https://www.context-lab.com
154 posts
2,336 followers
659 following
Prolific Poster
Conversation Starter
comment in response to
post
Very cool, congrats and looking forward to reading this!
comment in response to
post
Congrats, Commander Sulu!! 🖖🌈📖
comment in response to
post
I agree that LLMs can often do data science stuff well, but they also frequently make subtle errors that change functionality, or end up doing really dumb things that they get stubborn about.
You at least need to know how to critically evaluate and correct the outputs!
comment in response to
post
I'm not sure there's a secret hack to it. Just take the core courses to learn the fundamentals, and then continue to practice!
comment in response to
post
In addition, we have repeatedly seen over the past year that simply "throwing compute at the problem" does not work. Huge models like GPT-4.5 have way under-performed expectations.
So: attempts to stop training people to code (or hiring coders), are astoundingly short sighted. Keep on coding!
comment in response to
post
I've noticed this too! A related behavior comes up with coding tasks. When sanity checks fail, Claude often assumes that it's the *tests* that are the problem, or that we're in some special case where that check doesn't matter, and then it forges enthusiastically ahead with the broken approach.
comment in response to
post
Does he understand what CDC & NIH do? Their success stories (including COVID vaccines)? Are decisions being made through scientific consensus or political appointees with extreme opinions. What is long-term vision? What are the anticipated effects of indirect cost cuts on entire health care system?
comment in response to
post
I'd like him to address why NIH has been making massive changes to funding, cancelling grants, not paying invoices and not making any attempt to work with scientific community on reforms. Implications of denying access to COVID vaccines. No preparation for next pandemic.
comment in response to
post
Thank you for writing this! Very well said.
Re: brainstorming, the things on my mind have been:
- Infrastructure: what can be virtualized and decentralized?
- Funding: can we do more "go fund me" style science ventures?
In general, what parts of science can *scientists* control more fully?
comment in response to
post
Well it takes them *time* for an intern to run your manuscript draft through an LLM to correct it by introducing the optimal number of typos. And ChatGPT subscriptions* don't just buy themselves, you know! So that must be at least a few billion (well spent) dollars right there!
*And company execs
comment in response to
post
Seems like a good idea for a study!
comment in response to
post
...or...I think I can? One thing that has totally fascinated me is wondering about how *other* people process faces. Apparently it's different from me, but of course faces look "normal" to me from my perspective since I've never known anything else!
comment in response to
post
I don't know that I have deep insights into it-- I wasn't even aware that I processed faces differently until ~a year ago when I participated in a face perception study! I guess maybe I'm worse than average at reading people's expressions, but I can def tell when people are smiling, frowning, etc.
comment in response to
post
As a prosopagnosic I suspect I'll nearly always be in favor of less reliance on faces, in any domain 😂
comment in response to
post
@esfinn.bsky.social does awesome work on individual variability if you want some neat papers in that area! One of my faves: www.sciencedirect.com/science/arti...
comment in response to
post
Also, though, just to emphasize-- I'm 100% not an emotion expert! I think it's a neat field with unique challenges, and I find the computational/modeling parts especially interesting. But I don't know that literature well and am mostly picking things up through conversations and random talks I see 😁
comment in response to
post
In that paper above, we approached it by assuming that everyone had (very) roughly the same range of emotions, but not necessarily at the same times or in response to the same things. It let us make progress without needing a complete history of each person...but our approach has many issues too!
comment in response to
post
I.e., one could expect to make accurate(ish) predictions of V1 given a person's retinotopic map and the stimulus. Even semantic representations are somewhat similar across people. But emotions have an idiosyncratic element that makes them appear unstable and tough to accurately model/predict.
comment in response to
post
I'm a newbie in this area too-- but one of the interesting and somewhat unique challenges with emotion (unlike, say, perception) is that in many settings different emotions are experienced at different times for different people in response to the same stimulus. E.g., www.science.org/doi/10.1126/...