What Google's LearnLM reminds to educators
A call for revival
As a potential Google LearnLM partner, I closely follow the company’s moves that affect EdTech.
Last May, Google released LearnLM - a Gemini spinoff targeted to take upon classroom and other forms of education.
The Driving factor behind LearnLM:
Even before AI, Google’s presence in education sector is quite sizable. During pandemic, it was bullish on selling to governments.
Besides the mere presence of a business objective, though, there was a technical gap. Gemini was merely a chatbot - far from sufficient for teaching purposes.
Without a compelling LLM dedicated to mainstream education, the market would be soon filled with EdTech GPT (or Gemini/Claude)-wrappers.
Prompt-based GenAI is a wild-wild west where globally unregulated EdTech reigns by selling piecemeal and incoherent content.
Folks at Google realized that a data model fine-tuned on pedagogical data and classroom conversations would serve the world better.
Without going into details, lets define a foundational tanet:
In Large Language models, when it comes to the efficacy of the output:
\(Fine Tuned Model > Prompt Engineering\)
The task of fine tuning Gemini was an ML engineering challenge. It was challenging because classroom interaction datasets are extremely scarce (only 4 are publicly available). Whatever is available has questionable quality.
Tests have shown that LearnLM's evaluations of student excercises were quite close to those done by human teachers.
Is this the Khan Academy moment of the AI era?
Maybe, this is that moment.
Not to mention, Khan Academy is also busy developing its own Phi-3 model, fine-tuned for education.
But what’s worth pondering is: When Big tech begins to deliver anything to aid/supplement/drive education, what are the tanets it adhers to? What went into the research driving LearnLM's creation?
Google researchers went back to the roots of what makes teaching valuable and learning truly worthwhile.
They found 5 pillars of effective teaching upon which LearnLM could stand:
🏋️♀️ Manage cognitive load - Transmit content in an organized, easy-to-understand manner
🛝 Encourage active learning - Engage students to find answers and build things
🪞 Deepen metacognition - Ensure a deep understanding of what they are doing
🙋🏻 Motivate and stimulate curiosity - Encourage them to keep searching beyond the classroom
🧽 Adapt to the learners’ goals and needs - One-size-fits-all never works.
The GenAI race will keep flipping winners and losers. It's just getting started.
The important question is:
How near are our mainstream classrooms and EdTech to the founding principles of effective pedagogy?
Have we ever been capable of covering all 5 intellectual necessities of learning? Let’s keep finding out.