SAIL: Agents (still), Are we getting realistic?
June 20, 2025
Welcome to Sensemaking, AI, and Learning (SAIL). I look at trends that impact higher education.
AI and Learning
Software is Changing (Again). I was on a call with a colleague today and I said “hey, you need to watch this”. Apparently I was the fourth person to advise this, so it’s clear that the talk is making rounds. It’s not only about software. It’s the most grounded and fair assessment of what LLMs can do (and not do) that I’ve seen. The vision presented is realistic and gives agency to the individual. If you only watch one thing this week…
The State of AI in EdTech. Summary image on Twitter. It’s interesting to see how much of the assessment of AI in education (this talk mainly focuses on K-12) is being driven by venture firms and consulting firms. Education does move slow. And change and disruption benefits VCs. So watch with the dichotomy of “we need to be better as a system in innovating” with “the voices driving the conversation are seeking and economic reward”. Perhaps they’re more compatible than they appear on the surface. However, the 23:30min comment about “blow it up and have them all learn by an AI tutor” suggests that AI innovation will be violent in education.
There are all kinds of ethical dynamics here, but the use of celebrities (AI of course) to teach learners random content (see Onlocklearning). Drake, Sydney Sweeney, and others make appearances. What’s the relationship between entertainment and learning? I don’t think this is it… but maybe with Lamar.
Also not sure what to make of this - will the AI use by political affiliation become the marker of in-group status?
Higher education most urgently needs leadership to step up and prepare for the risks, opportunities, and ethics of AI. This article presents a framing for that (not only focused on education): Leadership in the Age of AI. “The rise of AI will usher in a new era where leadership is not defined by traditional notions of vision, authority, and influence but by the symbiotic partnership between man and machine, sparking a radical evolution in trust and ethical stewardship. In an AI-driven future, leadership will be fluid and algorithmic, providing oversight over humans and redirecting their imperfections, forever altering human agency and the notion of leadership itself.”
Multi-agent systems are key to making sense of AI in education. We won’t see one model to rule them all (at least not until significant memory and context size challenges are solved). We’ll see agents that perform discrete pedagogical tasks. This week, two articles were released that, on the surface at least (and in their title, I guess), suggest the path forward is contested. Both, however, focus on context engineering as a key solution to ensuring the learner experience is reasonably holistic and authentic.
Don’t build multi-agent systems. “No single approach to building agents has become the standard yet, besides some of the absolute basics…If you’re an agent-builder, ensure your agent’s every action is informed by the context of all relevant decisions made by other parts of the system.”
How we built our multi-agent research system. “A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together…Systems with multiple agents introduce new challenges in agent coordination, evaluation, and reliability.”
AI Technology
The big news this week was Meta stepping into the AI game with gusto. Their Llama model, which has languished, has been a substantive contribution. But less than stellar performance, or more accurately, somewhat gamed performance, and delays has resulted in Meta playing some catchup. They announced a new super intelligence AI lab. But not prior to considering purchasing Perplexity. And Ilya’s SSI (who doesn’t have a spare $32b sitting around). Ilya declined but Meta may be hiring several key co-founders. Altman says Meta is offering $100m signing bonuses. I haven’t seen any other sources that support this, so it’s possible that it’s a bit of propaganda to shame people at OpenAI to not make it look like they’re leaving for money rather than mission.
Attention in AI is turning to multi-media models. Google launched Veo 3 last week. Midjourney (who for the longest time had a Discord only option for creating images) dropped it’s own video model this week. A Chinese model, Hailuo, is a smaller, more affordable, model. And sits at #2 on video leaderboards, ahead of Veo 3. What does this mean? Learning designers will have a growing range of multimedia tools to work with. And at a broader level, Hollywood will move from actors to coders as the key talent.
Verification of Frontier AI Models. It feels like AI safety, security, and alignment are the equivalent of an eye dropper of water on a raging forest fire. Yoshua Bengio, however, continues the fight, leading this UN Brief.
Cluey upsets Very Serious People because it reflects a mindset that rewards cheating (the founder got famous with Leetcode cheating). they are continuing to raise funding. And are now an app that “operates discreetly on users’ desktops, intelligently interpreting live audio and on-screen context to deliver proactive insights.” Remember. Nothing is real. Everything is flux and change. Today’s heroes are tomorrows villains. And vice versa, of course. Reality is what is trending. Remember nothing.
Microsoft and OpenAI are fighting. It has the potential to be one of the most significant breakups in technology, given the enormous resources at risk.
OpenAI, and others, have changed their anti-military stance from several years ago. OpenAI was awarded a $200m grant to assist what seems to be a slight variation in language between them and DoD. The DoD says OpenAI “develop prototype frontier AI capabilities to address critical national security challenges in both warfighting and enterprise domains”. Open AI says that the project will focus on “how frontier AI can transform its [DoD’s] administrative operations, from improving how service members and their families get health care, to streamlining how they look at program and acquisition data, to supporting proactive cyber defense”
Speaking of military and AI: The DoD has enlisted a group of AI execs on a special project where “they will work on targeted projects to help guide rapid and scalable tech solutions to complex problems”
OpenAI Files. OpenAI is one of the most consequential technology companies of our era. These files offer a fascinating look into leadership and Sam Altman as CEO. It feels chaotic.