GPT-4 is here. Only a few months after the launch of ChatGPT, we’re living in a time where we can turn a napkin sketch for an app idea into an actual app and convert pictures of food into recipes. GPT-4 has sparked the "magic" of tech in a way that the personal computer did in the late 70s and the smartphone did in the late 2000s.
This shock and awe is changing how we collaborate, brainstorm, and explore. From coding to language learning to storytelling - AI is becoming intertwined with our daily interactions. OpenAI has become the “front door” for hundreds of millions of people to experience the power of AI and fundamentally transforming how we learn. While AI like GPT-4 can generate seriously impressive results (even if it's often totally wrong!), it's tempting to think we can learn everything by chatting with an AI. But should we?
In an age where AI can imitate humans and natural language processing is the future, we're at an inflection point in human-computer interaction. There’s a lot at stake here beyond homework.
The pandemic has been devastating for students and teachers alike, with reports showing that students experienced up to a year of learning loss. While technology has helped fill in gaps, staring at screens for hours of lectures and video calls is taxing and ineffective. A degree from “Zoom University” will only get you so far.
As AI systems become increasingly sophisticated, it's time for a paradigm shift in education.
What Makes GPT-4 Different?
Unlike earlier AI systems, GPT-4 can understand complex ideas and have nuanced discussions on a wide range of topics. It accepts various inputs like text, images and diagrams, enabling it to explain concepts visually.
GPT-4 builds on the transformer architecture of previous models but was trained using reinforcement learning and engineered prompts to become more capable. While OpenAI hasn’t shared the exact details of its architecture or training, GPT-4 can process 32,000 tokens at once, allowing it to handle longer, more sophisticated conversations.
A new input type called “system messages” allows users to specify the model’s tone and style for a given discussion. For example, you can direct GPT-4 to respond as the ancient philosopher Socrates would, encouraging critical thinking through questioning.
This feature can help students learn and think for themselves by adjusting questions to the knowledge of the student and break down a problem into simpler parts until it’s at just the right level for the student.
GPT-4 achieves state-of-the-art results on multiple AI benchmarks and even surpasses human performance on some tests. It demonstrates strong results in question answering, common sense reasoning, coding, reading comprehension and more — across 24 languages. The model also excels at understanding and discussing visual content like images, charts, diagrams and infographics.
In simulated tests like the LSAT, SAT and AP exams, GPT-4 scored between 80% to 100%. While not perfect, GPT-4 stays within guardrails and avoids inappropriate responses over 95% of the time, a major improvement over previous versions.
OpenAI candidly acknowledges GPT-4's potential downsides: "While less capable than humans in many real-world scenarios...GPT-4's capabilities and limitations create significant and novel safety challenges." Though GPT-4 outperformed predecessors in accuracy tests, it still generates incorrect facts and biased or illogical responses, and expresses unfounded confidence. It also lacks knowledge of events after September 2021, when its training data was finalized. OpenAI outlines these concerns in a report here.
As language models become more advanced, they also become more useful. OpenAI seems to consider GPT-4 ready for commercial use immediately—the first time it has unveiled a new model with simultaneous product launches utilizing it. While we have yet to fully understand the implications (and ethical considerations) of this technology, the adoption is clear.
What This Means For Education
Education has been known to be a laggard when it comes to tech adoption and implementation. But in this AI wave, students have become power users, spreading rapid usage and stirring cries of cheating among teachers. Educators have experimented with AI to create syllabi and grading rubrics and champion its creative potential. The idea of an AI tutor that’s just as good as a human? Suddenly it seems closer within reach.
ELIZA is considered to be the first chatbot in the history of Computer Science, developed by Joseph Weizenbaum at MIT in 1966. Today, we interact with chatbots for FAQ’s, customer service, and our food orders - why would we think about education any differently?
We've seen firsthand how groundbreaking developments in AI, coupled with thoughtful design and implementation, can enable and scale impactful educational experiences. Tools like TeachFX (voice AI for teacher feedback), WriteLab (writing feedback; acquired by Chegg), Mainstay (student support) and Gradescope (grading assistant; acquired by Turnitin) are now category leaders that have reached millions of students and teachers.
Students learn best by doing, asking questions, and articulating their thinking—not passively listening.
Conversational learning is a methodology that focuses on the use of conversation to facilitate the learning process. It is based on the premise that conversation is a natural and effective way to learn, and that by engaging in meaningful conversations, learners can deepen their understanding of a topic and improve their skills.
Georgia Institute of Technology used a chatbot named Jill Watson — created by IBM’s Watson — as one of the nine TAs for a 300-student classroom. “Ms. Watson” was able to answer approximately 10,000 messages a semester with a 97% success rate. “Ms. Watson” was so human-like that it took months before one student figured out, through a small technicality, that his TA was actually a bot rather than a human.
Georgia State University introduced an AI chatbot, Pounce, that prevented 324 students who claim they would have dropped out had it not been for the chatbot’s guidance. Pounce was able to answer questions round-the-clock, through text, on topics regarding financial aid, scholarship applications’ deadlines, and course enrollment inquiries.
Dialogue with professors and experts accelerates learning far more than reading alone. The Greeks knew this: tutors like Aristotle produced remarkable students like Alexander the Great. Today, there’s only one method that reliably improves performance by two standard deviations: 1:1 tutoring with experts. Conversational AI can replicate this at scale.
Emerging Learner-Facing Use Cases
Poe is a platform that lets people ask questions, get instant answers and have a back-and-forth dialogue with AI chatbots. Short for “Platform for Open Exploration,” Poe is designed to be a place where people can easily interact with a number of different AI agents, including GPT-4.
For example, imagine you’re in an intro to computer science course and you need to better understand the concept of API’s. Your instructor isn’t slowing down for questions and you’re not entirely sure what to do.
**Delphi **lets users tap into the knowledge and experiences of successful people through AI-simulated conversations. Students could ask historical figures like Steve Jobs or even past US Presidents how they might handle challenging situations, enabling discovery through debate and experimentation.
Summative assessment and lectures are becoming obsolete. AI assistants like Duolingo’s new Duolingo Max and Khan Academy’s Khanmigo act as personalized coaches, guiding learners through lessons and providing instant feedback. Khanmigo mimics a writing coach by giving prompts and suggestions to move students forward as they write, debate, and collaborate.
Duolingo Max helps users practice conversational roleplay and allows learners to practice real-world conversation skills with world characters in the app. Khanmigo offers AI-guided lesson planning feedback, mimics a writing coach by giving prompts and suggestions to move students forward as they write, and can even help students improve their coding skills.
The essence of education is conversation and experience, not lectures. Instead of passive lectures, students can have dynamic conversations with AI and debate with peers using AI as a mediator. They'll explore topics in a personalized, adaptive way. While the pandemic exposed huge vulnerabilities in our education system, it has also created an opportunity to build it back even better.
One-to-one learning is incredibly effective, but it's not right for every situation. Not all AI tutors are made equal. Research shows the most effective tutors build rapport, ask more questions than they give directions, and create opportunities for reflection. If you're building an AI tutor, how will you design interactions to inspire students and push them out of their comfort zone in a good way?
Human interaction and relationships are crucial for learning. AI tools need to scale these interactions, not remove them entirely. The future of education will be conversational—but it'll be a conversation between humans and AI, not humans or AI.
Thoughtfully designed AI can take learning to the next level, but it needs to enhance human relationships, not replace them. The most cutting-edge classrooms will combine the personalization of AI with the human touch. Lectures are over—the future is all about co-learning with AI. MOOCs are fading into obsolescence, learners desire productive struggle, debate, and real-world scenarios.
We have to realize the limitations of AI, algorithmic decision-making, and blind trust in technology. GPT-4 is not a silver bullet for the challenges in education. It's crucial we ensure this technology aligns with best practices and ethics. Learners must be active participants, not passive recipients, and we must address biases in AI that could worsen inequality.
Still, progress in AI will enable customized "tutors for life" to guide us through evolving interests and careers. Blended with human judgment and empathy, conversational learning with GPT-4 represents the future of education: interactive, data-driven, and tailored for each learner. The only limit will be our own curiosity.