SAIL: AI Products, America still winning, Call for Papers, Agents
March 27, 2025
Welcome to Sensemaking, AI, and Learning (SAIL). I look at how AI impacts higher education.
Building capacity and skill in developing AI products is a key university requirement. For those wanting to get started, there are a growing number of services and infrastructures available that make this process much easier than even six months ago. OpenAI, for example, just launched MCP support for their agent SDK, joining a large list of existing organizations that support the standard (specification?). This makes it easier for developers to interact with LLMs and takes friction out of working with different models. This is one of those examples where need and popularity, not W3C committees, drives rapid adoption of promising ways to lower barriers for developing agents.
A challenge that universities will face, though, is determining their infrastructure. This is like selecting an LMS, but with broader institutional impact. Most universities will likely follow the AI offerings of their primary cloud provider (currently between Azure, AWS, and GCP in USA). The AI support layer from Anthropic, Google, and OpenAI keep advancing and it looks, currently, like models can be swapped in and out without too much loss of quality. AWS is late to the game, but they understand infrastructure and developers better than anyone. In a meeting yesterday, one developer on our Matter and Space team said “why can’t we just do this all on AWS. We’d immediately see a 32% [oddly precise] in costs”. Things aren’t quite “click button” easy, but for talented technical people, the barriers to building are dropping daily.
Which leads to a question that I think every university department and faculty should be asking: “What is our AI product building strategy?” I used to want universities to have lofty goals and visions (well, I still want that). For today, I’ll settle for a dean who says “we’re building support agents for students who need additional tutoring using langchain and we’re revising prompts with langfuse so we can make sure we’re nailing our evals. We have faculty, support staff, and leadership involved in testing. We’ll be ready for a June pilot.” We won’t be able to reason and argue our way through this - we need things our faculty and students can play with.
AI & Learning
Ok. So let's say you’re on board and want to advance a small build team on campus. What do you do? What can you learn from others? This article by Hamel Hussain is gold: A Field Guide to Rapidly Improving AI Products “With new tools and frameworks emerging weekly, it’s natural to focus on tangible things we can control – which vector database to use, which LLM provider to choose, which agent framework to adopt. But after helping 30+ companies build AI products, I’ve discovered the teams who succeed barely talk about tools at all. Instead, they obsess over measurement and iteration.” (my emphasis). It’s really the only thing you need to read this week.
The opportunity of technology to restructure systems of learning to better serve learners has been a core interest for roughly my entire career. We have a special issue of Computers & Education:AI planned that addresses exactly this need. We’re asking researchers to help us think through the last mile of AI in higher education by detailing anticipated systems impact. Call details: Generative AI and new systems of learning for higher education. “The advancement of AI technology is unusually rapid and its pace vastly exceeds typical progress and adoption curves from previous technologies. It is for this reason that we are taking an atypical approach to calling for papers that place greater emphasis on impact than is typically expected in research publications.”
When will we have AGI? Colleague and intermittent source of irritation, Pete Smith sent this article: Some things to know about achieving artificial general intelligence. The article is a bit broad and at times going back from intelligence measuring through to the Chollet’s measure of intelligence. It then ends a bit unsatisfying by simply arguing for…a new theory of general intelligence.
This is K-12-focused, but has been making rounds: ‘AI tutor’ rockets student test scores to top 2% in the country. I’m less interested in those results than I am in the nature of the school curriculum because a comparable shift needs to happen in higher education (flipped classroom 2.0 is coming): “kids crush academics in 2 hours, build life skills through workshops, and thrive beyond the classroom”. That’s the epistemology to ontology shift Paul LeBlanc and I keep talking about. We are in the middle of a golden age of rethinking learning design where almost everything we know needs to be rethought. It’s more about design for AI than for learners and giving space to develop human connectedness with self, others, and nature. And the literature that needs updating and rethinking makes AI and learning design a fantastic field of study. But significant rethinking is required by existing practitioners since they risk not seeing the opportunity in service of existing knowledge and practices.
AI & Technology
With ongoing advancement in AI from China, there is growing dialogue about whether USA is falling behind. Apparently not as every major category of AI (reasoning, coding, video, image, OCR) is from a USA-based organization.
Deepseek releases V3 stating that it “outperforms other open-source models and achieves performance comparable to leading closed-source models.” What’s unique is the reduced training and operating costs compared with other models.
Tracing thoughts from a large language model. This is cool. And a bit confusing. But as they say at the start of the short video, LLMs are trained, not programmed. “During that training process, they learn their own strategies to solve problems. These strategies are encoded in the billions of computations a model performs for every word it writes. They arrive inscrutable to us, the model’s developers. This means that we don’t understand how models do most of the things they do.”
It may not feel like it from media, but Google is winning the AI race. They have the compute, infrastructure, data, and best performing models. (though benchmarks are starting to be rather meaningless). The vocal developers are on OpenAI and Claude. The quieter ones are on Google :). This week, they dropped Gemini 2.5 “a thinking model, designed to tackle increasingly complex problems. Our first 2.5 model, Gemini 2.5 Pro Experimental, leads common benchmarks by meaningful margins and showcases strong reasoning and code capabilities.” Unfortunately for Google, OpenAI dropped an image generation model and that’s all everyone talked about. Though there are significant copyright and ethical questions to work through. With AI, you’re lucky to get a day or two in the sun (or benchmarks).
Have you heard about agents? Apparently, they’re the big focus this year :). This is a nice visualization, from a downloadable book) on agentic architectures.