@johnsteill.bsky.social
I have personally found chatbots to be *amazingly* effective tutors. I feed it a pdf and prompt it to ask first easy questions, then progressively harder ones, tying in other concepts from other chats in the project. Domains: Statistics, software languages, hugging face AI papers.
The evidence AI tutors work "astonishing well" is incredibly thin. We've seen these same promises made for virtually every technological "innovation" of the past 100 years. (Carl's essay is more nuanced than this post describing it.)
Latest post looks at the complicated picture emerging on AI tutoring: when AI tutors work astonishingly well and when they quietly make students worse. ⬇️
November 9, 2025 at 5:44 PM
I have personally found chatbots to be *amazingly* effective tutors. I feed it a pdf and prompt it to ask first easy questions, then progressively harder ones, tying in other concepts from other chats in the project. Domains: Statistics, software languages, hugging face AI papers.
I just wasted a very surreal, absurd hour with OpenAI's new Atlas browser. It insisted it had capabilities, like examining files in a directory or accessing my calendar, which it definitely did not.
"Well, I guess those haven't been released yet."
Vaporware.
"Well, I guess those haven't been released yet."
Vaporware.
October 25, 2025 at 5:46 PM
I just wasted a very surreal, absurd hour with OpenAI's new Atlas browser. It insisted it had capabilities, like examining files in a directory or accessing my calendar, which it definitely did not.
"Well, I guess those haven't been released yet."
Vaporware.
"Well, I guess those haven't been released yet."
Vaporware.
This week I read this “old” GenAI paper: ReAct: Synergizing Reasoning and Acting in LLMs (Yao et al. ‘23). It pioneered interleaving planning and tool use (calculator calls, web searches, etc.) mid-trajectory. 🧵
September 8, 2025 at 9:05 PM
This week I read this “old” GenAI paper: ReAct: Synergizing Reasoning and Acting in LLMs (Yao et al. ‘23). It pioneered interleaving planning and tool use (calculator calls, web searches, etc.) mid-trajectory. 🧵
Just finished reading Agentic Reinforced Policy Optimization (ARPO) by Dong, Dou, et al. Dense but fascinating. It’s all about how LLMs decide when to call tools — and how uncertainty (entropy) can guide smarter, more efficient reasoning. 🧵
August 22, 2025 at 2:41 PM
Just finished reading Agentic Reinforced Policy Optimization (ARPO) by Dong, Dou, et al. Dense but fascinating. It’s all about how LLMs decide when to call tools — and how uncertainty (entropy) can guide smarter, more efficient reasoning. 🧵
Reposted
