SAIL: Students, Life Coach, Agents
April 20, 2025
Welcome to Sensemaking, AI, and Learning (SAIL). I focus on AI’s impact on higher education.
AI can do many discrete knowledge tasks well - tasks that are the basis for much of the modern education system. Since we can’t out-compute AI, humans will need to perform higher level or higher order tasks. That means that the intention of the university system moves from teaching learners certain types of things and teaching learners to be certain types of ways.
Essentially this is a type of cognitive escalation where what an individual used to learn in the past such as the discrete distinct tactics of how to code or discrete specific processes that are required in some kind of a business analysis can now be somewhat automated and handed off such as a report that can be quickly generated now through any growing number of tools.
I have yet to see a satisfying list that gets at these future knowledge/skills/ways of being needs. I would be challenged ethically to argue the existing higher education system is the best way for most people to prepare for the emerging world.
AI and Education
Take a few moments to reflect on We Need a New Kind of Insurance for AI Job Losses: “Unlike traditional job losses that allow for sector-based rebounds AI-driven unemployment presents unique challenges to existing social insurance systems. If AI automation renders entire industries obsolete displaced workers won’t be able to simply wait for market conditions to improve or transition into similar roles elsewhere.“
Being Human in 2035. The majority of the 301 experts feel AI’s impact will be considerable or dramatic. Unfortunately, many of the negative impacts will be on core human connection, our ability to think deeply, and our mental health.
Google is the natural king of AI as they have the data, the internal expertise, and the computation. They’re now starting to flex that profile by offering free access to their top AI tools to higher education students.
Welcome to the Era of Experience. New sources of data enable new models and improved performance. Fei Fei Li is doing this with world models. The authors of this paper suggest that: “Ultimately experiential data will eclipse the scale and quality of human generated data.This paradigm shift accompanied by algorithmic advancements in RL will unlock in many domains new capabilities that surpass those possessed by any human."
Compound systems. The LLM MOOC from 2024 (a new advanced MOOC on the same topic is currently being offered) had a talk on compound systems, which they define as “modular programs that use LMs as specialized components”. In education, this will be the norm. We aren’t only focused on a chat or agent interactions. Context and student data matter in the overall performance of the system. Unlike a call center or a customer service chatbot, education has so much nuance in guiding and directing and assessing learners that any system that succeeds will fit into the profile of these compound systems.
Agentic AI: Stanford. Great lecture that reviews basics of LLMs and moves on to how agents can be developed and deployed.
General AI
It’s all agents these days. And OpenAI has their guide on how to build them.
Agents are central to how AI will advance and impact society. Various use cases are emerging, even though education hasn’t moved beyond basic tutoring. The tools/platforms and infrastructure for developing agents will be key to longer term vendor lock in.
Here’s a nice summary of a current battle, favoring Anthropic’s agent framing from last year over OpenAI’s (i.e. bullet point above). Ultimately it’s about the distinction between agents and workflows and knowing when to use which. I find myself nodding with the approach discussed here which focuses more on hybrid, but with the realization that eventually, AI takes over more.
Excellent reflections here from the founder of langchain on the OpenAI Agents paper: “The hard part of building reliable agentic systems is making sure the LLM has the appropriate context at each step.”. And of course, when using multiple agents, the orchestration between agents.
This agent framework comparison chart is worth a skim as well.
OpenAI released several models, including one that essentially kills the need for the $200/mo o1. This post has been getting attention (helped by Altman retweeting) in stating that o3 is near genius level intelligence.
If you’re wondering how OpenAI manages to release new products and features weekly, listen to the CPO discuss how they’re developing and rolling out their products.