SAIL: How to use LLMs, Robots, Voice, MCP, Vibe Coding
March 15, 2025
Welcome to Sensemaking, AI, and Learning (SAIL). I look at the trends in AI that have implications for higher education.
This week, I’m focusing on learning opportunities for educators to get started in the domain of emerging AI concepts. AI is best learned by doing and playing. The Karpathy video in particular is highly recommended.
AI and Learning:
The best educator in AI is Andrej Karpathy. He dropped another two hour video on how he uses LLMs. I appreciated the nuanced review of when to trust LLMs, how to get a feel for which LLMs succeed in which tasks, and general “hacks” to get more value from an LLM (such as with math problems). Some suggestions are simple - get the LLM to explain what it’s going to do or to create a data table first before it starts processing).
How close are we to AGI? A number of labs seem to think “very soon”. The discussion is becoming more public. I’m reminded of an interview last year with Daniel Schmactenberger where he argues we already have a type of AGI in the form of modern companies, governments, and social structures. These are systems that hold more intellectual capability than individuals. This lines of reasoning fits in with thinking on connectivism that Stephen Downes and I have pursued over the years.
Vibe coding is all the rage and a short search will generate many results. I like this short tutorial from DeepLearning on WindSurf. I would expect most educators to get comfortable with these tools. But as important is to have an understanding of what’s happening “under the hood”. Here’s an accessible intro on “a non-technical guide to how code works”.
Anthropic released model context protocol (MCP) last year. They describe it as a standards protocol for AI applications. Here’s a short visual explanation. MCP can help universities build and develop apps (help with literature reviews, data processing, integrate course materials, create personal learning plans, etc) by leveraging Claude (well, any frontier LLM) as part of an overall AI system. More info here. (Anthropic calls it the “USB-C port for AI applications”
AI Technology
China is starting to worry big tech, with OpenAI and others calling for government supported moats. Why? Well, on the one hand you have Manus, a new “agentic” offering from China (still need to sign up for an invite code). Here’s an example of a “big tech stock performance I requested. The Tesla stock seems mismatched in performance compared with recent price moves. It’s hard to fully track how it produced some of the analysis that it did.
Baidu, speaking of Chinese companies, launched it’s own 4.5-worthy LLM, Ernie. More in this thread. Biggest difference between Ernie 4.5 and GPT-4.5? Ernie is 1% of the cost.
AI agents are this year’s big trend. OpenAI just released supporting resources for agent builders. The use of single/multi-agent workflows will be useful for academics who want to solve a range of simple to complex problems (think of being able to use multiple agents to engage a learner in assessment, learning, self-testing). Tools like langchain have offered an agentic workflow (or, more in line with their marketing cognitive architectures). OpenAI launched there agent resources with this video. They define agents as “a systems that act independently to do tasks on your behalf”. This targets developers (SDK and APIs figure prominently, so it’s about OpenAI offering the technical infrastructure for organizations to develop approaches to solving complex problems (i.e. multi-agent architecture)) and university IT departments should be watching this closely.
This is the year of AI agents. Next year is going to be robotics (not sure when glasses will get their due, but that’s on the horizon as well as general wearables). Deepmind dropped Gemini Robotics - using the Gemini 2.0 model for physical and spatial tasks. Nice overview at the end of the article that details the adaptive and responsive nature of their robotics. The application of reasoning and LLM and “real world” models to robotics provides flexibility that current robotics don’t posses. Here’s (yet) another overview of robotics in 2025.
There is significant progress with AI/voice. Sesame has released state of the art voice - a bit reminiscent of the move Her. Spend some time with it. Unnerving. They have also released the model on Huggingface, so it’s open-ish.