SAIL: Data consortium, All of Life, Context Windows, GPT-4o, AI and Deception
Sensemaking, AI, and Learning (SAIL)
Hi all,
There’s an important conversation happening in AI that was reflected in a short exchange I had on a panel last week. One of the panelists was the proverbial futurist, pitching the prospect that even significant problems today will be resolved through AI. I’ve assumed this role of “pitching the future to solve real problems today” in many public sessions myself. Sometimes, the point is to provoke an audience to question their assumptions. Regardless, this individual felt that LLMs are everything and all problems will be solved with large context windows. I, in contrast, think we’re going to have to wrap significant support around LLMs to make them useful and increasing context window alone isn’t the answer. Background: A context window is the number of tokens that an LLM can use when generating an output. The larger the window, the better. This enables the LLM’s responses to include more information (and retain more history if the interaction is ongoing), though placement in the context window can determine the accuracy of the response.
LLMs do some wonky things and researchers have used a range of emerging techniques to make them more accurate and useful. Or to make them more relevant for a constrained data set such as institutional data. These include RAG, DPO, GraphRAG, etc.
The big question is: will large, even infinite, context windows and improved memory solve all things LLM? My view - no. As great as LLMs are, they don’t exist in isolation and they need a support infrastructure (like an OS?).
AI and Education
Higher Education is not engaged in AI at the level it should be. We propose a data consortium as an on ramp to building with AI. I’d love to hear feedback on this concept paper.
We don’t like it when AI gets too close to core human spaces, notably emotions and connection. The soundly (and appropriately) ridiculed Apple ad reflects that when tools of human creativity are destroyed. Or Bumble AIs conversing with many AIs to find the right relationship fit in advance.
Writing effective prompts might best be done by…AI. Anthropic has a solution.
AI and LLMs are starting to make a big appearance in learning analytics and educational data mining literature. Here’s an interesting recent contribution that focuses on traditional tutors (i.e. CMU lineage) and AI: Combining Dialog Acts and Skill Modeling: What Chat Interactions Enhance Learning Rates During AI-Supported Peer Tutoring?
AI in General
OpenAI released GPT4o (Omni) model yesterday. One of the more interesting features is the voice assistant. Short overview of announcements: free access (but with rate limits), less latency, cheaper, faster, with vision (photos/videos). Promising opportunities as a tutor. What’s left for humans and human teachers?
A few responses to GPT-4o:
Unbridled passion: “The 1st biggest day in AI was on 11/30/2022 (ChatGPT’s debut). The 2nd was on 3/14/2023 (ChatGPT-4’s release). Yesterday was the 3rd, marking the launch of ChatGPT-4o.”
Confused meh “The speech synthesis is terrific, reminds me of Google Duplex (which never took off). but If OpenAI had GPT-5, they have would shown it.They don’t have GPT-5 after 14 months of trying.
Impressed “First, it's multi-modal across text, images and audio as well. The audio demos from this morning's launch were extremely impressive.”
It will accelerate existing trends in agentic workflows. Universities in particular should care about this.
AI and Deception: “This paper argues that a range of current AI systems have learned how to deceive humans. We define deception as the systematic inducement of false beliefs in the pursuit of some outcome other than the truth.”
Let’s not forget Google. They are still the AI heavyweight. And they’re starting to deliver. OpenAI’s runway is short. Which is likely why they want to nail down an iPhone deal.
The AI powered internet. “The AI-powered internet is here, and not many people have realized it yet.” Good list of implications follows.
LLMs are public facing AI that we can all marvel at and remain in states of existential worry. There is more to AI than language. We need to keep focused on the other solutions that might impact, oh, all of life: “By accurately predicting the structure of proteins, DNA, RNA, ligands and more, and how they interact, we hope it will transform our understanding of the biological world and drug discovery.”