Tessera
Folgen

Brendan Bartanen

@brendanbartanen.bsky.social

Associate professor of education and public policy at the University of Virginia. Stumbling my way through the AI revolution at https://brendanbartanen.substack.com/

37 Posts2320 Follower11009 Folgt
Brendan Bartanen@brendanbartanen.bsky.social · 2mo ago

Re-upping my post on AI interpretability from a few months back in light of the government's decision to ban access to Mythos/Fable for foreign nationals. For now, AI "guardrails" are just a high-stakes game of whack-a-mole. Read more: brendanbartanen.substack.com/p/we-dont-re...

We don’t really know how AI models work

“Interpretability” and the high-stakes race to figure out how to understand and control powerful AI

brendanbartanen.substack.com

Brendan Bartanen@brendanbartanen.bsky.social · 2mo ago

An easy next step if you want to know more: direct your AI model of choice to examine the public Gitbhub repo and explain to you everything Claude did to create its products: github.com/brendanbarta...

GitHub - brendanbartanen-svg/naep-achievement-decline: Verifying the post-2013 US student achievement decline (NAEP) and testing eight explanations

Verifying the post-2013 US student achievement decline (NAEP) and testing eight explanations - brendanbartanen-svg/naep-achievement-decline

github.com

Brendan Bartanen@brendanbartanen.bsky.social · 2mo ago

It's time to move on from this outdated understanding of what AI models do. Agentic AI is not just fancy autocomplete. The models can pull data, write code to analyze it, verify claims, create audit trails, etc. It's not error-proof and vigilance is essential, but it's not the ChatGPT of 2024-2025.

Brendan Bartanen@brendanbartanen.bsky.social · 2mo ago

It's time to move on from this outdated understanding of what AI models do. Agentic AI is not just fancy autocomplete. The models can pull data, write code to analyze it, verify claims, create audit trails, etc. It's not error-proof and vigilance is essential, but it's not the ChatGPT of 2024-2025.

Bruce D. Baker@schoolfinance101.bsky.social · 2mo ago

The problem with using an LLM to write social science is that all it can do is scrape/weight/summarize reorganize text that's been used in relation to the topic - even if that training set is rife with mis & dis-information. It's quality neutral regurgitation. Hence the cell phone conclusion.

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

Did you wake up this morning thinking "I wish someone would explain 'sparse autoencoders' in an approachable way"? Do you wish you knew about 'Golden Gate Claude', perhaps the best AI model ever created? It's your lucky day--all of that and more below!

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

AI models are very different. We don’t actually know how they work. Despite more capital investment than for any technological endeavor in history (roughly 50 Manhattan Projects after adjusting for inflation), we are, in some sense, flying blind. open.substack.com/pub/brendanb...

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

AI models are very different. We don’t actually know how they work. Despite more capital investment than for any technological endeavor in history (roughly 50 Manhattan Projects after adjusting for inflation), we are, in some sense, flying blind. open.substack.com/pub/brendanb...

We don’t really know how AI models work

“Interpretability” and the high-stakes race to figure out how to understand and control powerful AI

open.substack.com

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

AI models are very different. We don’t actually know how they work. Despite more capital investment than for any technological endeavor in history (roughly 50 Manhattan Projects after adjusting for inflation), we are, in some sense, flying blind. open.substack.com/pub/brendanb...

We don’t really know how AI models work

“Interpretability” and the high-stakes race to figure out how to understand and control powerful AI

open.substack.com

Matt Blackwell@mattblackwell.bsky.social · 5mo ago

LLMs will usher in a democratization of intellect. Now you too can have a junior (AI) scholar that you can blame for errors and fraud. Previously, only the top scholars had access to that kind of power.

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

Assuming you're using agentic AI tools, my default work mode is now: every single task is an opportunity to figure out how to get Claude to make that task easier/better for me. Anytime it works well, save it to claude.md so that it just does it automatically next time.

Brendan Bartanen@brendanbartanen.bsky.social · 5mo ago

Fascinating study and a demonstration of one of the deep challenges with LLM-based AI models: because they are trained on human-generated text, they can exhibit patterns resembling human cognitive biases (e.g., anchoring) even in clinical contexts where the stakes are high.

Scott McGrath@smcgrath.phd · 5mo ago

A new Nature Medicine study on ChatGPT Health isn't encouraging. The system under-triaged 52% of emergencies, telling folks with DKA to wait 24-48 hours. It does fine on routine stuff, but fails at clinical extremes. I'll take a deeper look 🧵 #MedSky #ChatGPTHealth