Monday, May 13, 2024

Is teaching becoming obsolete with GPT4o? Do we now have a UNIVERSAL TEACHER?

I have written and talked about the idea of a UNIVERSAL TEACHER for a long time, notably in my new book, where I go into detail about the learning theory behind 'dialogue' in learning and the key role of chatbots in teaching and learning. This is what AI promised to deliver. A free teacher who speaks, listens, remembers, tutors, using all media types, can read handwriting, provide personalised feedback, on any subject, anytime, anywhere, in any language.

What I never imagined was that it would come so fast. Yet OpenAI has delivered on what I thought was this utopian idea. In their demo they showed this in action. A frictionless, fast and sophisticated tutor. 

Open AI has become the Apple of AI. They understand, like Steve Jobs, that the user experience is all. This is especially true in teaching and training. Of all the applications, teaching and learning is the one that has most to gain from GPT4o. They may have out-Appled Apple, as Siri, Google assistant and Alexa are nowhere near as good as this. At this rate, teaching is rapidly becoming obsolete.

Realtime teacher dialogue

You can chat with it in realtime as it has realtime speech and, in the live demo, understands your voice, your emotions as expressed through your voice, even your facial expressions. You can have dialogue just as you would with a real teacher, you can interrupt it and it responds fast, as fast as a real teacher. It can generate a teacher’s voice in many styles – funny, friendly, serious, academic… whatever. The teacher’s voice is extraordinarily realistic.


Real maths problem taughtThere was lots of pre-launch chat, stimulated by Altman, about the movie 'Her'. We now see why. The system is very 'chatty'. Indeed the fact that you can define the character by asking it to be someone, is astounding. This is a shift in branding away from Google's serious sounding assistant towards a real, emotional definition of a chatbot, so that it can play any defined role. The possibilities from different type of teachers, tutors, trainers and instructors to role playing with patients, customers and employees, even therapists, are endless.They showed a maths problem, with the learner doing it on paper, the handwritten problem, then shown to his smartphone. 

He shows her a linear equation:  3x+1=4. 

The tutor suggests he get all numbers on one side, giving a hint.

He subtracts 1 from both sides to give 3x=3. 

How does this look? He asks.

She congratulates him.

He asks for a hint for the next step. 

What undoes multiplication? She asks. 

He suggests subtraction but she says, think of the opposite of multiplication. 

He says division? 

Go ahead and divide both sides by 3 she instructs.

He does this and the solution is x=1

Well done you’ve solved it, she replies.

He then asks for real world applications and she gives him several. The tutor is endlessly patient, friendly, gives relevant feedback and can read his written steps in moving towards a solution.

n another teaching problem, by Salman Khan, you ask it to be a tutor and it talks you through a maths problem around the sides and angles in a right-angled triangle. There’s great back and forth dialogue between the student and AI, with hyper-personalised feedback and reinforcement. It does everything at the pace of the learner, behaves just like a patient tutor and corrects any errors the learner makes. All by simple showing the problem and the learner’s efforts to his smartphone. Although one has to be careful with staged examples, as it isn't doing much 'hard reasoning' here. That may be there already, it will certainly come. There is also the issue of having to show it an image, the next step is surely to draw and write on the screen as an option.

I wrote about Khan almost a decade ago, saying he was an important figure, way beyond Robinson, Mitra and many others. OpenAI have been wise in teaming up with Khan on education.

Some great features in this teaching video. For examplewhen the tutor is talking, if you start speaking, it stops and waits for your response. Focus on the voice of the teacher - it's very neat. The intonation is also interesting, teacher-tone. and it can be adjusted to suit any individual or audience. Its endless patience removes the frustration that every teacher and parent feels when teaching, as it defuses the stress.

One thought I did have, which is almost existential - if it can teach and do all of this maths, why would we teach it. Roger Schank used to go on about this a lot - why teach skills that can be automated, especially algebra and geometry? In any case the focus is often on maths as that is an area of catastrophic failure for many kids and adults. Maybe AI will solve the problem by solving the maths itself, not teaching millions to do what can be done in a millisecond by voice and AI.

Multimodal teaching

As a teacher it can read text, recognise images and video, just like a real teacher. In another example, the voice app on your desktop helps the learner solve a code problem, understands the written code on the screen, also the voice of the learner. She explains what the ‘foo’ function is, step by step. More than this you can then ask it to see the data visualisation that the code produces. Ask it questions about the graph and itinterprets the graph for you. So it can teach you maths, code and data analysis.


It also does real time translation, so you can teach in one language, give feedback in another, great for people learning in a second language. The possibilities in language learning are also mind blowing.


The accessibility featured of GenAI often overlooked. I bang on abut this all the time and has just been boosted by GPT4o. Text2speech, speech2speech, now highly personalised dialogue. 'Bemyeyes' with GPT4o is amazing. The fact that it is free is also a huge boon for access by the poor and all who are excluded due to cost.


There was already an administrative function at enterprise level for GPTs, already used by some global companies, such as Moderna. This will be extended. It is at this organisational level that they will make money.


GPT4o is better than GPT4 and hammers other models. It is also faster and smarter across text, vision and audio, truly multimodal. What have they done and how? Behind the scenes the optimisation needed to deliver low-latency audio to audio in real time is massively impressive. This is not trivial as dialoghue overlaps, interrupts and is difficult to map This is a huge leap on ease of use in dialogue and intelligence with low latency, necessary for the smooth dialogue needed in tutoring. It reasons across text, voice and vision, making the teaching experience seem like a real human teacher. This could revolutionise teaching, accelerate learning, even accelerate home-schooling, maybe the end of the personal tutoring business. I feel that this is a game changer for parents as well as teachers. As a first step this is astonishing, as it is a globally scalable solution to a problem that has plagued education, where teacher shortages and costs are a problem. This is a great leveller.

The fact that the branding is GPT4o is tantalising, as if they have something else up their sleeves - GPT5? But what they have done is redefined AI as something that becomes more human using dialogue. This shift in our relationship with computers is fundamental.

Of course, this still needs testing across a range of examples but this is an astounding start. There is no stopping of progress here, the UNIVERSAL TEACHER will happen, and soon. The future is now.


Noted they were using iPhones and a deal was stuck yesterday with Apple - something very big is brewing there.

No comments: