Joel Chan | 🦣: [email protected]
@JoelChan86
Followers
4K
Following
34K
Media
838
Statuses
6K
Professing #hci #creativity #toolsforthought #metascience @INFOCollegeUMD & @hcil_umd | building @discoursegraphs 🦣: [email protected] 🦋: joelchan86
epistemological hinterlands
Joined October 2011
I think this is me!!
you can just decent enough at stuff and then be a switchboard for other people’s excellence and that can be a tremendous contribution
0
0
16
What a beautiful metaphor for creative knowledge work! I was able to track down the quote, and share it here so others may savor it too 👇 Source: https://t.co/tPZ8jbVNCs
I am not an algebraic geometer but by many accounts the great genius of Grothendieck was partly in his talent at definitions. With the right definitions the problems start to just dissolve. He used to tell a parable about opening a nut by dissolving it rather than cracking.
1
1
3
Absolutely fascinating. It’s precisely this kind of “superhuman search” that will challenge commonly-held assumptions about our capacity to discover new findings in the exponentially growing literature (and therefore in what formats, venues, and indices such findings must
My posts last week created a lot of unnecessary confusion*, so today I would like to do a deep dive on one example to explain why I was so excited. In short, it’s not about AIs discovering new results on their own, but rather how tools like GPT-5 can help researchers navigate,
1
6
43
NUANCE IS BACK, BABY
It's pretty funny that we are converging on the reality that will upset the maximum amount of people: LLMs are pretty good. They're not transformative, and also not useless. There are some tasks where LLMs will save you a lot of time. And that's it
3
9
231
Deeply wrong message to draw this as the defining line. Peer reviewed journal publications have done incalculable harm due to claiming validation through peer review which doesn’t happen by opinion the moment science is dropped but by continuing to test and build upon the work
Preprints are powerful, but not yet validated. They accelerate discovery and debate, but they are not peer-reviewed and should never be used to shape public policy, especially when it directly impacts the health and safety of women and children.
6
9
73
Many this morning came early to attend Nicholas Carlini's keynote at #COLM2025. He spoke up to bring awareness on a range of problems with LLM use, summarized by the question: are they worth it? Here I report his main points and why they deserve attention. 🧵
4
23
181
Some scientists aren’t waiting for journals to catch up. They’re showing us what’s next. The Beyond the Journal awards honor those breaking the mold in how science is shared. Nominate or self-nominate below 👇
3
10
47
Andy is on one of the most under-discussed and fascinating beats out there, making a compelling argument that America is still pretty affordable for people willing to live outside major metro areas:
I know of a cabin for sale on a half-acre lease bordering state forest for $20,000. And I know of a bank that would give a loan for it with 20% down at ~7% interest with a credit score of 650+, no appraisal or inspection required. That's $4k down, mortgage of $120/mo, plus
23
4
244
Copilot in Excel is a global financial crisis waiting to happen.
468
4K
61K
👀👀👀👀
🚨 Just published! All frontier AI models have failed “Radiology’s Last Exam” - the toughest benchmark in radiology launched today! ✅ Board-certified radiologists scored 83%, trainees 45%, but the best performing AI from frontier labs, GPT-5, managed only 30%. ❌ These results
0
0
2
The error analysis in this paper is 👨🍳 💋 The narrative that AI has “solved” certain disciplines is a flawed way of thinking that serves only the AI suppliers. The best AI models only perform well on subcategories of problems in a discipline and that too with high variance.
🚨 Just published! All frontier AI models have failed “Radiology’s Last Exam” - the toughest benchmark in radiology launched today! ✅ Board-certified radiologists scored 83%, trainees 45%, but the best performing AI from frontier labs, GPT-5, managed only 30%. ❌ These results
2
2
24
(Systematic) literature reviews should not be published as papers. Instead, they should be living repositories with releases and changelogs, with key conclusions as modules. People should then be able to "import" modules to inform their ongoing work.
0
2
30
AI contributions to community notes sighted! I... don't hate this idea?
0
0
0
New Commoncog homepage is up, completing our 2025 repositioning. Finally!
4
3
41
Thankfully....
0
0
3
TIL that paperswithcode was sunsetted, while trying to look up eval data for a model I was considering, and got redirected to https://t.co/Z1C34oDbNe via google... :/ apparently killed by spam? ughhh
BREAKING: we've partnered with @metaai and @paperswithcode to build a successor to Papers with Code (which was sunsetted yesterday) PWC, founded by @rbstojnic and @rosstaylor90 has been an invaluable resource for AI scientists and engineers over the years (and an inspiration
1
0
2
Here's a way of visualizing the maps *defined by* an autoencoder. The encoder f maps high-dimensional data x to low-dimensional latents z. The decoder tries to map z back to x. We *always* learn a k-dimensional submanifold M, which is reliable only where we have many samples z.
3
11
157
Did you know your MacBook has a sensor that knows the exact angle of the screen hinge? It’s not exposed as a public API, but I figured out a way to read it and make it sound like an old wooden door.
610
3K
54K
Helpful reframing. Picture to help: From @3blue1brown: https://t.co/fjRXDTFjHX
Amusing how 99% of people trying to explain LLMs forget that they don't generate the next token, they generate a probability distribution over the entire vocabulary space that the end application is free to sample from You are very often not presented with the Most Likely Token
0
0
3
