From: redpointai
AI is rapidly transforming education for students, teachers, and school systems [00:00:44]. Schools might be among the first places to see mainstream adoption of AI for productivity [00:00:05].
Khanmigo: An AI-Powered Tutoring Assistant
Salman Khan, the founder and CEO of Khan Academy, one of the most influential education platforms globally, has been at the forefront of implementing AI in education [00:00:26]. Khan Academy serves over 150 million learners across 190 countries [00:00:30].
Their main AI initiative is Khanmigo, an AI-powered tutoring assistant [00:00:36]. Khanmigo has already been deployed to over 1.4 million students and teachers [00:00:40].
The core mission of Khan Academy has always been to approximate or replicate the personalization provided by a good tutor [00:05:10]. This vision draws inspiration from concepts like “The Young Lady’s Illustrated Primer” from Neal Stephenson’s Diamond Age [00:05:58]. The latest generation of models, particularly GPT-4, which Khan Academy accessed months before ChatGPT’s public release, showed the potential to approximate tutoring and teaching assistance [00:06:08].
Khanmigo’s Features and Development
Khanmigo was launched as an AI tutor and teaching assistant with key features:
- Guardrails: Ensures teachers can monitor student activity [00:06:34].
- Anti-cheating: Designed to prevent cheating [00:06:38].
- Safety and Privacy: Prioritizes student data protection [00:06:39].
- Socratic Method: Leans into good pedagogy by using a Socratic approach [00:06:44].
Initial projections estimated 100,000 pilot users by 2025, but Khanmigo quickly surpassed this, reaching 1.3 to 1.4 million users [00:06:55]. Districts pay approximately $15 per student per year for the service, covering compute costs, support, and training [00:07:11].
Evolution to Proactive AI
Khan Academy is transitioning Khanmigo to be more proactive [00:07:33]. Currently, like a human tutor waiting in the back of the room, only about 10-15% of students proactively engage with the AI [00:07:44]. The next version, “Khan Academy Classroom,” aims for a more concierge-like AI that actively welcomes students, reminds them of tasks, and offers assistance [00:07:57].
The “blank screen problem” (users not knowing what to prompt) is addressed by suggesting actions and making the AI more proactive [00:08:18].
Addressing AI Limitations and Efficacy
While many are developing AI tutors, the AI is not yet ready to drive learning entirely on its own for most people [00:09:12]. While curious individuals can learn a lot from tools like ChatGPT by spending significant time prompting it, this is not typical for most students [00:09:21].
Current AIs are still not perfect at creating high-quality questions without errors [00:09:35]. Efficacy observed in deployments like Newark comes from combining traditional Khan Academy practice with AI as a support tool to drive engagement [00:09:48]. The main challenge is engagement; if students engage with a reasonably healthy tool, it’s likely to be efficacious [00:10:09]. AI can help human systems facilitate engagement and accountability [00:10:44].
Unexpected Uses and Accuracy
Students have found surprising ways to use Khanmigo. For example, a student had a lengthy conversation with an AI simulation of Jay Gatsby, which drove the conversation by asking about her own life [00:11:52]. Students are also using it to explore advanced math concepts [00:12:28].
Khanmigo’s error rate when anchored to Khan Academy content is about 2% [00:12:57]. This 2% is split evenly between math errors (1%) and evaluation errors (1%), where the AI might incorrectly assess a student’s close answer [00:13:07]. While aiming for 0% error, this rate is already considered better than many human tutors [00:13:26].
Teacher Adoption and Impact
Teachers are leveraging AI to streamline lesson planning and delivery:
- Lesson Planning: AI helps create more entertaining and appropriately sized lesson plans [00:14:18].
- In-class Gaming: A partnership with Blooket allows Khanmigo to generate game questions, reducing a teacher’s prep time from half an hour to two minutes [00:14:26].
- Insights: Teachers gain insights from student performance to inform future planning [00:14:56].
- AI Simulations: Teachers use AI to open classes with engaging simulations of historical or literary figures [00:15:27].
- Writing Coach: Khan Academy’s “Writing Coach” tool addresses cheating fears by having the AI act as an ethical coach for assignments. Teachers receive the entire writing process, not just the final output, and the AI flags copied content [00:15:39]. This undermines all forms of cheating, not just AI cheating [00:16:09].
The ideal usage involves teachers regularly integrating AI into their routines, creating habits around using it for assignments, data analysis, and student accountability [00:15:13].
Policy and District-Level Adoption
School districts are rapidly adopting AI, seeing it as a primary driver for productivity and learning [00:19:11]. Teachers have already embraced it due to its ability to streamline planning and grading [00:19:30].
AI offers a significantly cheaper alternative to previous interventions. For example, after the pandemic, 25-10-$15 per year and offer more “dosage” [00:20:18].
Districts using Khanmigo report saving teachers at least 5 hours per week [00:21:03]. It’s also being used as a recruiting and retention tool for teachers [00:21:06].
Necessary Enhancements Beyond Core Models
Building effective AI for classrooms requires substantial development beyond just integrating large language models (LLMs):
- Safety and Moderation: Implementing robust safety measures and moderation tools, initially overly conservative, but now more refined [00:21:48].
- Math Accuracy: Significant work is needed to reduce math errors, especially in complex evaluation scenarios where students are “hammering” the system or switching contexts [00:22:07].
- User Interface: Creating a natural and intuitive user interface [00:22:35].
- Proactive AI Architecture: Re-engineering the front end for a proactive, AI-first approach, moving beyond a simple chatbot [00:22:41].
- Integrated Tools: Integrating various AI functionalities like brainstorming for thesis statements, outlining tools, and drafting assistance directly into the platform [00:23:02].
- Prompt Chaining: Developing robust systems for “prompt chaining” behind the scenes, allowing a single app to perform multiple AI actions [00:23:37].
- Robust Memory: Enhancing the AI’s memory to give models more context, with options to reset or manage specific aspects of memory [00:23:58].
- Advanced Voice Integration: Integrating advanced voice capabilities for more natural interaction [00:24:26].
- Higher Quality Question Generation: Improving the models’ ability to create higher-quality questions [00:24:58].
- Visual Work Recognition: Enabling the AI to see and provide feedback on students’ handwritten work, especially on tablet devices [00:25:32]. This could make AI tutoring indistinguishable from human tutoring [00:26:07].
Model Evaluation
Khan Academy uses a rigorous model evaluation framework:
- Tough Test Cases: A series of hundreds of “tough test cases” where models initially struggled (e.g., fractional representations, distributive property) [00:29:16]. Initially, models failed on 70% of these hard cases; now, they fail on less than 10% [00:29:46].
- Machine Labeling: Using AI to analyze AI interactions and identify potential errors [00:30:13].
- Human Labeling: Human reviewers analyze sample conversations (e.g., 2,000 conversations were reviewed six months ago) to confirm errors and assess the “productivity” of conversations (e.g., student engagement vs. disengagement) [00:30:27].
Future Classroom Vision (20 years out)
A great classroom in 20 years will still prioritize active engagement, interaction, and problem-solving, but with AI enabling teachers to have “superpowers” [00:02:14]. AI will free up teachers’ time from grading and lesson planning [02:57:00]. It will provide better insights into student progress and offer more interactivity [00:03:31].
The future classroom will feature ambient AI observing and supporting learning without requiring constant screen interaction [00:03:50]. Additionally, mainstream virtual and augmented reality, combined with generative AI, will allow for immersive simulations, like a “magic school bus ride” to ancient Rome [00:04:10].
The ideal is for AI to prompt the teacher when a student is struggling, allowing the teacher to assign a specific AI tutoring session [00:34:44]. This “teacher in the loop” approach ensures accountability and engagement [00:34:45].
Skills for the Future Workforce
Core skills like critical thinking, writing, reading, math, and general knowledge remain essential [00:42:59]. However, the ability to act as an “entrepreneur” — taking existing resources and combining them in new permutations to create value — will become increasingly crucial [00:43:26]. This means individuals will need to continually experiment with new tools and integrate AI into their workflows to accelerate productivity [00:43:55].
For example, using AI to transcribe verbal thoughts and generate a first draft of a speech can save hours, but still requires human refinement [00:46:16]. Similarly, AI can generate entire apps, but human input and understanding are still needed [00:46:33]. This highlights the need for a hybrid approach where individuals learn both traditional skills and how to leverage AI tools.
The Broader AI in Education Market
The AI in education technology market is currently noisy, with many startups offering thin “prompting layers” over existing AI models [00:38:23]. While AI can accelerate app development, the low switching costs and barriers to entry make it a challenging environment [00:38:46].
Non-profits like Khan Academy have an advantage due to their ability to take a longer-term view on development and build trust within the education community regarding pedagogy, efficacy, and safety [00:39:06].
Other areas ripe for AI innovation include:
- Interviewing and Assessment: Addressing inefficiencies in the hiring process, which often involves extensive, non-standardized, and expensive assessments [00:40:33]. AI could streamline initial screenings and provide standardized certifications [00:41:13].
- Corporate Training: Making mandatory corporate training (e.g., cyber security, sexual harassment) more engaging through AI simulations [00:42:11].
Conclusion
The rapid advancement of AI has led to a significant shift in Khan Academy’s strategy, with the organization now viewing itself as AI-first [00:48:21]. The education community has also moved faster than expected in adopting AI [00:48:36]. Initial concerns about errors, hallucinations, bias, and cheating are being addressed by transforming these risks into opportunities [00:49:30].
For those interested in exploring Khan Academy’s AI work, users can try Khanmigo and the Writing Coach at Khan Academy [00:49:57]. Additionally, Schoolhouse.world, a sister non-profit, offers free tutoring through volunteers and is launching “dialogues initiatives” for conversations on tough topics [00:50:07]. Khan Academy also aims to offer high school and eventually college credits and diplomas, making quality education more accessible globally [00:37:39].