Proffer Edu: Decentralized Education on the Blockchain
App #2 of Proffer’s 5 apps in 5 days series
Proffer Edu is a peer-to-peer question/answer based education app on the blockchain — one of 5 apps my teammate Sinchan and I built for Coinbase’s Toshi/Token Hackathon in June ’17 to explore use cases of social search on the blockchain. These 5 apps helped us iterate on the protocol design for Proffer, and collectively won the grand prize in the hackathon.
The goal of this article is to discuss why decentralized 1-on-1 education is a valuable pursuit, and how it can be implemented easily with Proffer, the foundational protocol for search through crowd intelligence on the blockchain.
Note: If you’re curious to learn more about Proffer before going through this article, read the tech spec here and a higher level walkthrough of a social search on Proffer here.
TL;DR — one minute summary
- 1-on-1 education is awesome (2 sigma effect on student performance in the classroom) but doesn’t scale.
- AI-lead instruction can help with scale but won’t replace human instruction anytime soon. Need to look for ways to do 1-on-1 education using a human network.
- Most attempts to scale 1-on-1 education with human networks have been unprofitable due to difficult market dynamics and unsustainable unit economics. Supply utilization (productivity of the tutors you’re paying to teach) and LTV/CAC (lifetime value of customer / cost of acquiring that customer) are two important success metrics that all but the earliest market entrants have struggled with. Stiff competition in tutoring marketplaces, lack of customer loyalty, and high variance in student needs and tutor capabilities doesn’t make the task any easier.
- A peer-to-peer education app on the blockchain can outperform present-day approaches with the help of smart contracts that allow for easy creation of incentives (e.g. frictionless p2p payments) and trust.
- Peer to peer information exchange can be mapped to the problem of social search using expert networks where a Seeker (student) poses a question to the global, distributed social search engine named “Proffer” and receives in response an Answer Space of responses from Responders (tutors). Through a process of crowd-sourced peer review, these answers are sorted and filtered down, providing speed of answer delivery and correctness (an otherwise difficult task for subjective answers) to the student end-user.
Author’s past experiences in ed-tech
My focus over the past few years has been on building tech for education, so I’m especially excited about this application of the Proffer protocol. 6 years ago, as a computer science Masters student, I was part of the MIT team developing AppInventor, a visual programming tool now used by 6M worldwide to build mobile apps without prior coding experience; around the same time, I founded Camp K12 in India to bring fun, hands-on coding camps, curriculum, and competitions to k-12 students across the nation; and more recently, I launched and co-direct OpenEd.ai — a non-profit initiative promoting and developing open-source AI to solve problems in education. In Summer ’17, OpenEd.ai organized a global AI for Education HackWeek in partnership with IBM, Omidyar Network, Google Developer Groups, and Amazon Web Services, awarding $19K in prizes to the top 5 out of 1,400+ participants from around the world.
I believe that there is ample opportunity to remove friction from the process by which we acquire knowledge so that we may learn better, faster, and cheaper; and there are an ensemble of innovations — technological, pedagogical, and operational — that will enable us to do so.
The blockchain is one such innovation. The model described here for p2p education via Q&A on the blockchain is part of an evolving vision, one that will be brought to life and put to the test with Proffer. Would appreciate your feedback and reactions as you read.
1. 1-on-1 Education is awesome but doesn’t scale
Thirty years ago, educational psychologist Benjamin Bloom published his work on The Search for Methods of Group Instruction as Effective as One-on-One Tutoring. In it, he describes that across a number of classroom subjects and student groups in his study, students receiving 1-on-1 or small group tutoring perform 2 standard deviations above their peers.
More recently, experiments have shown that adding 1-on-1 mentorship to e-learning platforms (e.g. MOOCs) provides much needed interactivity and improves completion rates and learning outcomes. Case in point: Udacity saw a 3x improvement in course completion in the presence of 1-on-1 human coaches accessible via chatbot. Codecademy (via 24/7 live chat advisors, see image below), the FlatIronSchool (via learn.co), and many other learning platforms have added 1-on-1 mentorship for this very reason. It’s great to be able to get help right when you need it, from an expert who can offer you not only an answer, but explanations and understanding.
The obvious challenge with human-provided 1-on-1 mentorship is cost and scale.
For offline learning, it’s unrealistic to assume that there will be 1 teacher available for 1 student. Bloom himself points out in his publication that individual tutoring for every student is “too costly for most societies to bear on a large scale” and urges the education community to “search for group methods of instruction that might be as effective as one-to-one tutoring.”
Even online, providing 1-on-1 tutors requires volunteers or paid tutors “on-call” to be available on short-notice. Meeting an ever-increasing demand of questions to be answered with a supply of paid labor can get quite expensive. More on this in point #3 below.
2. AI helps with scale in narrow domains; won’t replace human instruction anytime soon
Recent developments in machine learning models, decreasing costs of large-scale computation, and increasing availability of large-scale data sets to be used for the ‘learning’, make it possible to build Artificially Intelligent (AI) tutors that can answer student questions with little or no human intervention, for a small fixed cost and minimal / zero ongoing cost, anytime / anywhere.
This is one of the many problems the team at OpenEd.ai has been working on, applying cutting edge research in machine comprehension and information retrieval to building Q&A bots capable of fetching answers from a database of past answers (as demonstrated by “Jill Watson” a bot created by Prof. Goel at Georgia Tech) or finding phrases in some provided input text that best answer a provided question, demonstrated below:
Looks promising right? In narrow domains where large data sets are available (e.g. “6th grade US History” with 1 Million q&a pairs for just this subject), AI tutors can perform quite well given that for any new incoming question, there is likely already a semantically similar past-question with a corresponding past-answer that can be retrieved.
However, in open domains (e.g. general intelligence that can answer any question on any topic), or in domains where data isn’t readily available, AI tutors are useless.
The AI discussion so far has been on AI tutors that do q&a — building AI tutors that provide “education” rather than simple “answers” is even harder. Education typically entails an extended two-way dialogue and word-by-word composition of answers rather than retrieval (fetching a past answer).
In a 2016 interview, Andrew Ng had the following to say about technologies like “AI tutors”:
Most of the value of deep learning today is in narrow domains where you can get a lot of data. Here’s one example of something it cannot do: have a meaningful conversation. There are demos, and if you cherry-pick the conversation, it looks like it’s having a meaningful conversation, but if you actually try it yourself, it quickly goes off the rails.
The conclusion: if we wish to provide 1-on-1 learning experiences at scale, present-day AI will only get us a part of the way there. We’ll need to look at human solutions instead — centralized (employer hires tutors, tutors answer student questions in real-time) and decentralized (marketplace where tutors and students can enter and exit the market freely and transact with each other directly).
Further reading: excellent post by Denny Britz, ex-Google Brain, on Q&A chatbots: http://www.wildml.com/2016/04/deep-learning-for-chatbots-part-1-introduction/
3. Centralized approaches to 1-on-1 education have proven unprofitable
Fully centralized (high quality, high cost, hard to scale):
Imagine a call-center, with a full-time staff of trained teachers, ready to jump on the phone, video chat, or a chat thread to offer answers and knowledge the moment a student posts a question.
This method ensures high quality of education as teachers are pre-trained and carefully vetted, but is difficult to scale and suffers from economic inefficiency as there is down-time while hired teachers are on the job but have no students to teach. A company must spend money to hire up-front and is not guaranteed to recover its costs or earn a profit.
Semi-centralized (medium quality, medium cost, still hard to scale):
A different, semi-centralized approach is one where employees are hired part-time and scheduled to work only during peak demand. They are paid on an hourly basis, sometimes with a required minimum # of hours per week, sometimes without any restriction.
Over the past five years, this asset-light pay-as-you-go model has gained popularity in the education space. E.g. Yup (a 24/7 homework live-help service for k-12 students), Zeal (a live math-coach service for in-class use by teachers), Codementor.io, SimpliLearn, Vedantu, and SnapAsk are a sampling of live learning platforms from around the world that try to scale their labor force up and down to match demand from students.
Some of these, e.g. SimpliLearn which offers live courses for Big Data/Analytics, have found a content niche that attracts working professionals with a willingness to spend. Others play in crowded spaces e.g. math/science, and compete aggressively through marketing and ad spend to win customers that have little loyalty for their education apps. The result is high cost of customer acquisition (CAC) relative to customer lifetime value (LTV), an unsustainable situation. Andrew Ackerman describes the LTV/CAC challenge well in this article on “Hackneyed ideas that make ed-tech investors cringe”.
4. Peer-to-peer education on the blockchain could get around these challenges
In this section, we envision a decentralized education app where teachers and students are identified and matched for live online tutoring sessions on the fly without a company at the center accruing revenues or incurring labor costs and ad expenses. There are two questions to be addressed on why an app of this sort would solve the aforementioned problems. The first is ‘why p2p education in the first place’, and the second is ‘why on the blockchain’.
Why peer-to-peer?
In a peer to peer app, each user would be able to play the role of student or teacher depending on question topic. A number of cases that previously would have required a trained, expensive instructor to be dispatched to a student can now be answered by classmates either free of cost, for social rewards (e.g. reputation points), or for fees smaller than before.
A peer to peer app is inherently scalable since supply of teachers grows 1-to-1 with student demand and the network effects are strong, resulting in significantly lower user acquisition costs. Quality of these peer-to-peer interactions may be lower than the formal teacher-to-student interaction so a system could be designed where the easier questions are answered by peers and the harder questions are sent to trained instructors.
Why on the blockchain?
a) Easy p2p payments to reward answers and peer-reviews of answers:
Financial incentives will play a key role in the success of a p2p education app. While some peer learning platforms have managed to scale in the absence of financial exchange between teachers and students (e.g. Brainly, a social learning network for k-12 students, has 80M users across 35 countries), they have been restricted to simple q&a. For tougher questions that require the teacher to put in time and effort, financial rewards would be necessary to compensate for the teacher’s opportunity costs. Even for simple questions, the prospect of financial reward might bring additional teachers onto the platform and encourage them to contribute an answer.
Direct payments from student to teacher are easy to do on the blockchain, regardless of location, currency, and denomination, as demonstrated in the screenshot below from ProfferEdu:
Such financial rewards could also be paid out to users to encourage them to evaluate and upvote/downvote each others’ responses — a process known as peer review — in order to sort and narrow down the set of answers that the student receives for each query.
b) Novel economic model that doesn’t require entity in the middle extracting profit:
Traditionally, a peer-to-peer app would need to charge transaction fees on payments between students and teachers, or some form of membership/subscription fee from users in order to make enough money to cover tutor wages, customer acquisition costs, and other company costs.
Contrarily, a decentralized organization building a similar p2p solution on the blockchain would facilitate the matching of student demand with teacher supply through code running on a smart contract without needing to take a cut of the transactions, leaving more money on the table for teachers to make, or less money for students to pay. Developers and other members of the decentralized org could instead be compensated in tokens whose value grows with the usage of the platform, as Fred Ehrsam describes in “The Dawn of the Decentralized Business Model”.
There may still exist a need for a decentralized app of this sort to charge fees to users to cover costs not covered through appreciation of token value in the token business model, yet those fees would be smaller than their counterparts in the traditional approach.
5. Implementing P2P Education on the blockchain with Proffer
Proffer is a search protocol that finds answers to questions through crowdsourced human intelligence. At a high level, it works by posting a question to a network of responders with subject-matter expertise, and then sorting and filtering the responses it gets from experts and presenting them back to the seeker — the user who initially asked the question.
The exact mechanism by which this works can be summed up as crowdsourced peer review using expert networks, and is described in detail here:
Here’s how the Proffer protocol would handle a student’s query / request for help in a peer-to-peer education app:
- Student opens up his p2p education app, e.g. Proffer Edu, on his phone and posts a question “Why is the sky blue?”, with a seeker stake of $1.00 (this is the amount the seeker is willing to pay to responders that offer a correct response). This value is optional, but providing a non-zero seeker stake incentivizes more responders to attempt a response.
- Proffer Edu, the higher level decentralized app (dApp), then invokes the Proffer protocol by calling Proffer.search() and passing to this function a number of protocol configuration parameters described here.
- Proffer queries its global expertise bank to find subject-matter experts in “Physics”, “Optical Phenomena”, “Light Phenomena”. It then sends the question “Why is the sky blue?” to each of them.
- Each of these experts can either upvote/downvote answers that have already been contributed by others, or add a new answer to the answer space.
- Proffer keeps track of the expertise of each responder that upvotes, downvotes, and contributes an answer, and can therefore compute the net skill (sum of the skill levels of the various responders) supporting or opposing each answer. This allows the protocol to weigh the opinions of more skillful experts more than that of novices.
- Based on the net skill backing each answer, Proffer sorts the answer space and filters out the responses it thinks are likely incorrect given the number of downvotes they received and the skill of the downvoters.
- Proffer rewards responders who contributed ‘correct’ answers and penalizes responders who contributed ‘incorrect’ answers. Rewards and penalties consist of both money (tokens) as well as skill, such that correct responder continue to accrue skill over time and incorrect ones continue to lose it. At Proffer, we call this dynamic expertise network a self-optimizing expert network; it is the foundation of our crowdsourced peer review protocol.
- Sorted and filtered answer space is presented to the student (the seeker) who originally posted a question on Proffer Edu.
Proffer Edu (@profferedu) is live on Toshi, the ethereum browser / messaging app published by Coinbase. You can also view a demo video of it at https://www.youtube.com/watch?v=f4OuRMFdi6g. Check out our other articles explaining the Proffer protocol and its various use cases at https://blog.proffer.network
