news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author: Shidao AI Group

Editor: Shiraishi

In the early morning of July 17, Andrej Karpathy, former co-founder of OpenAI and former director of Tesla AI, officially announced his startup Eureka Labs, an AI+education company. Its first product will be an undergraduate course LLM101n that guides students to train AI models, similar to a scaled-down AI Agent.

He described a completely new learning experience: subject experts are scarce, and subject experts who are proficient in foreign languages ​​are even scarcer, just as Feynman could not personally guide the world to learn physics. However, all of this will be rewritten by GenAI - helping teachers design courses, helping students complete courses, and allowing 8 billion people around the world to innovate their learning experience.

Karpathy pointed out that both product development at Tesla and AGI research at OpenAI were side jobs, and now Eureka Labs is what he really wants.

There are many tech giants who are keen on the education industry, such as Andrew Ng, who co-founded the online education platform Coursera with Daphne Koller.

Another is Bill Gates, who loves to read textbooks. Last week, he updated his blog post My trip to the frontier of AI education, telling about his recent "fantasy trip" to First Avenue Elementary School in Newark, New Jersey.

Imagine that in a chemistry class, students no longer stare at boring formulas, but use daily necessities such as marshmallows, water bottles, and balloons to do experiments by themselves, observing the volume changes of marshmallows under different air pressures. Chemistry teacher Melissa Higgason adopted the advice of AI and turned the abstract Boyle's law into a fun experiment.

In Chinese classes, students do not need to rack their brains to figure out the author's psychology, but can have a few rounds of dialogue with the "AI version of Gatsby" to explore more details about "why he looked at the green light in the distance" and complete a book review. After class, teachers do not have to correct 180 homework assignments word by word, because AI is capable of processing students' open-ended answers - marking out the basis for scoring according to given standards, making it easier for teachers to give final scores.

With the help of AI tools, many scenes of future education are already unfolding.

This AI education application that Bill Gates praised is Khanmigo, developed by Khan Academy. As a "universal science master", it can not only explain complex concepts and create relevant mathematical problems, but also encourage students to solve problems independently through the Socratic method. As an "all-round teaching assistant", it can not only help teachers write lesson plans and plan courses, but also correct homework and track the learning progress of each student, deeply integrating AI into every aspect of teaching work.

As we all know, education is a perfect match for the current wave of AI and has become an excellent field for GenAI to land. According to Techcrunch, the most popular AI tutoring apps in the United States are all made in China - the "big factory" faction: Zuoyebang Question AI, Byte Gauth; the "non-big factory" faction: Answer AI.

It may seem like the old "take a photo and search for questions", but its foundation is ChatGPT and other excellent large models, achieving the perfect combination of "smooth experience + powerful capabilities". With an annual fee of less than $100, these "overseas kings" are snatching jobs from traditional education and training giants such as Kumon, and have generated tens of millions of dollars in annual revenue.

But if we look beyond the current situation, we can find that although chatbots have achieved initial success in the field of education, they are definitely not the ultimate form of AI education. So, what are the definite trends in AI education and what are the unnoticed opportunities? Shidao will combine Bill Gates' blog, a16z insights, and the AI ​​education map drawn by Laurence Holt, the founder of K-12 institutions, to sort out the relevant content.

Established: The big framework for educational reform

At the beginning of last year, when ChatGPT caused global excitement and panic, a16z was the first to conclude that education will not be replaced by AI. Because human beings' desire to learn comes from self-motivation and is an instinct. On the contrary, people will engage in more in-depth and extensive learning because of the easy access to resources, such as Wikipedia.

According to a16z's predictions, there will be five major changes in the education industry, some of which have already been realized.

1. 1v1 will become the mainstream

Scientists have found that students who receive 1v1 teaching perform better than children in traditional classrooms. In response to this, Numerade's AI tutor Ace can generate personalized learning plans and curate appropriate content based on students' skill levels.

AI can also become an expert and academic master. For industries that value "master-apprentice inheritance", it is equivalent to completely passing on skills to ordinary people. Imagine what it would be like to chat with an "AI version of Bill Gates" or an "AI version of Plato"? This is exactly what the startup Delphi is trying to do. We also saw similar innovations in Khanmigo mentioned above.

2. Personalized learning becomes a reality

With AI, everything can be personalized, from learning styles, needs (visual vs. text vs. audio), content types (introducing favorite characters or favorite genres), to courses. For example, Cameo launched a children's product featuring Blippi, Spider-Man, and other anime IPs. One mother even asked "Spider-Man" to encourage her child to learn and train, which was very effective.

Moreover, AI can handle different types of learners with ease: whether they are subject geniuses, underachievers, or sensitive children who don’t like to raise their hands, etc. ADHD is ecstatic.

3. The rise of customized AI tools

a16z believes that university is an ideal place because it is a densely populated environment with a large number of people and organizations, and popular products can be quickly rolled out through student organizations and club activities to gather word of mouth.

Students and teachers are the first adopters of many productivity software. For example, students at the university where the founder of Canva works use the design platform to create the school yearbook; a marketing professor at Northwestern University uses Qualtrics to collect data for MBA and doctoral students.

In addition, many teachers are easily overwhelmed by trivial matters, with high workloads but low salaries. AI can reduce teachers' workload by creating plans and draft syllabi. By freeing up time, teachers can focus on other tasks, such as giving students more personalized attention.

Students also like to find creative methods, a typical example of which is "taking a photo to search for questions". a16z cited the example of Chegg and predicted that the opportunity for AI teaching aid applications has arrived.

This has been verified. Chegg has recently launched a new automatic answering system and developed an AI platform for education, including 26 large language models that are vertically customized for the education field. In addition, there are applications such as Photomath and Numerade that focus on solving math problems, as well as the "overseas kings" just listed - Question AI, Gauth, and Answer AI.

4. The old standards are no longer applicable

Is it cheating to use AI to write homework? a16z’s view is that we cannot use the old-era yardsticks to grade homework in the AI ​​era, and new assessment tools will be developed.

The next generation of tools could help schools better assess student learning outcomes and award certificates, and they could also become levers for applying artificial intelligence to reduce the burden on teachers and students.

Sal Khan, founder of Khan Academy, devotes a chapter to cheating in his new book Brave New Words. He points out that cheating was rampant before AI. If anything, AI has only made the problem more visible, or "equalized" cheating. That is, everyone can cheat equally...

However, there is a "middle ground" worth thinking about. For example, on Khanmigo, you can write homework with AI, but when AI submits, the teacher receives not only the final output, but also the communication process between the student and AI. AI will say: I wrote with Adam for four and a half hours and found that she had difficulty in presenting the argument. Look, this is the complete conversation record. "

5. Fact-checking is crucial

A study by the University of Washington published in the WSJ showed that 72% of the audience believed that news written by AI was credible, even if the facts were inaccurate.

In the age of AI, people will blindly trust the people, brands, and “experts” they deem worthy of attention and respect. We may create a generation that is capable but doesn’t understand the basic details, which may lead to a marginal crisis when detailed knowledge of the underlying details becomes important. Take web development as an example - we move away from low-level hardware, infrastructure, and backends and enter a world of GitHub Copilot where front-end engineers rarely need to touch databases or backends. There are even no-code solutions for ordinary users. What happens when a critical error occurs on the backend and no one understands how to fix it?

Opportunity: A Panoramic "Treasure Map" of AI Education Applications

Bill Gates wrote in the preface to Sal Khan's new book Brave New Words: Every time there is a change and innovation, there are people who worry that machines will take away jobs. But when it comes to education, AI tools and tutors can never and should not replace teachers. What AI can do is support and enhance the ability of teachers.

This view coincides with NFX's "continuous innovation". That is, after the introduction of AI, some industries will expand in size infinitely and give rise to some niche markets. This is where startups win. The key is to recognize what kind of changes will happen in the industry and what variants will appear in the process. Opportunities are hidden in the changes. At present, AI education applications are in full bloom, and "photo search" seems to be the leader. However, in a huge market, it is impossible for only a single tool to exist around a single link. So, under the big framework predicted by a16z, what kind of hexagonal warrior will be born?

Recently, Laurence Holt, founder of K-12 institution, and Jacob Klein, product manager of TeachFX, jointly updated A Map of Generative AI for Education. The full text is 20,000 words long and covers all aspects of AI education - discussing in detail 70 sub-scenarios of AI tools in five major links, including teacher helpers, classroom materials, evaluation and feedback, and student helpers.

It is no exaggeration to say that in the education industry, every AI transformation opportunity that is happening or about to happen is reflected in this "treasure map".

1. Teachers are a good helper

1. Course Generation Tool

Teacher-directed – e.g., x minutes of group work, y minutes of class discussion, etc.

Integrate into the curriculum and build coherent knowledge—for example, being able to remember and understand what was taught two weeks ago, as well as what will be taught next.

Produce courses with high quality standards – for example, produce courses based on the teaching principles proposed by Rosenshine or other similar assessment standards.

2. Teaching and tutoring tools

Tools allow teachers to record lessons and generate automatic analysis and feedback.

For example, TeachFX and Edthena provide analysis after the lesson. In the future, tools may provide real-time coaching to help teachers guide on data-based frameworks (such as Marzano or Danielson) to support school-wide implementation.

Audio quality remains an unresolved issue.

3. Teaching consultant

An elementary school teacher found that some students did not understand negative numbers and suspected that there was something wrong with his teaching method. Therefore, he used an AI consultant to evaluate the teaching plan and students' classroom feedback. The AI ​​consultant helped him connect with like-minded teachers, recommended articles and papers, and facilitated a discussion meeting including AI subject experts. In the end, the teacher found a new teaching method.

Tools such as EduGPT have emerged to play the role of teaching consultant, as well as multiple consultants for different subjects and topics.

If pedagogy is fine-tuned, AI tools can act as a coach to teachers — advising on specific approaches to teaching, suggesting alternatives, and diagnosing students’ strengths and misconceptions.

For example, tools like TeachingLab.ai and Coteach.ai can understand existing lessons and suggest improvements to teachers — including changing the pace, adding content, activities to check for understanding, providing support to ensure students master foundational skills, and making connections to concepts students have already learned.

AI-driven collaboration tools can breathe new life into online professional learning communities, which otherwise often fail to attract enough educators. AI can learn each teacher’s priorities and challenges, match them with similarly situated educators, and inject relevant research and blog posts into the conversation.

4. Classroom Management Simulator

Simulators are widely used to train pilots, surgeons and even CEOs. Why not teachers? The Teacher Development Trust is looking into the issue.

Late at night, a teacher who suffered a "bad setback" during the day wants to hone himself by simulating classroom scenes. AI can play the role of students, not only recreating the "misery" of the day, but also showing the students' true personalities. Teachers will get feasible expert advice while trying different methods. Future tools can use AI agents to play students with specific personalities and then interact with human teachers in real time.

5. Student data analysis

An elementary school teacher uses data to analyze the growth of students who received intervention: Is growth accelerated? Who responded to the intervention and who didn't? As a group, what letter combinations did they miss most often? What skills did they no longer need to learn? The teacher uses this data to optimize instruction and reassign students to intervention groups that are better suited for them.

Although schools collect a large amount of data, only a small portion is available for use.

AI tools can extract student data (e.g., from SMS, LMS, or proprietary applications in CSV format or via API) and analyze it to suggest optimal student groupings, focus areas, and generate insights.

AI tools that can perform analysis and generate data visualizations — including OpenAI Advanced Data Analysis and Fluent. Tools specifically designed to understand educational data, such as Doowii and Strived.io, are emerging, for example.

AI tools allow for multiple levels of analysis—for example, you can upload NWEA MAP data for several classes and ask the tool to identify areas of fastest and slowest growth; ask higher-level questions, including analyzing possible academic regression over the summer and whether tutoring programs had a positive impact; and combine data from multiple sources to build a detailed, cross-disciplinary picture of the class.

2. A good teacher for students

1. 24/7 1V1 counselor

AI is not just a practice app, it is also expected to be a tutor in your pocket, just like interacting with a real tutor. Currently, AI tools are most prominent in language - writing tutors (Quill, StoryBird.ai, Caktus, StorySeed); foreign language tutors (Duolingo, LangoTalk, Iago, Supernova). But in programming, its performance can only be described as a helper - programming tools (CodeSignal Learn, Replit). Strangely, AI does not use its natural advantages in learning mathematics.

There are many problems to be solved.

Khanmigo is expensive to use, though it’s sure to drop, but CK-12’s Flexi is similar, but it’s free to use.

The user experience with so-called AI counselors is often more like a treadmill than building trust like with a human counselor. For example, they use a text chat interface, in part because the text-to-speech conversion rate is still too slow to feel natural.

There are existing tools for "homework tutoring" such as Brainly, CourseHero, Project Chiron, Studdy, CheggMate, Symbolab, etc.; practice apps such as edia; test preparation apps such as r.test and Archer. These tools provide step-by-step solutions, but they cannot be considered a true tutoring experience because they deprive learners of the opportunity to find a self-study path.

There are currently no tools designed for the three-weekly intensive sessions that are most effective in human tutoring. Over time, this problem will be solved. AI tutors may go a step further and provide immersive scenarios (e.g., through VR) and narrative-based learning scenarios (e.g., through videos, EngageAI Institute) that are highly engaging and better reflect the real world.

The chatbot lacks the ability to respond to student sketches. This is likely to change in 2024 as multimodality is a core focus of AI model innovation. (This paragraph was written in 2023 and has been improved.)

2. Children’s reading coach

An elementary school teacher gave students a tool to create their own extracurricular reading based on the Greek mythology unit they were studying. The AI ​​tool generated a mini book for each student based on their favorite mythological character, with matching illustrations. The book included comprehension checks embedded in the text.

Young readers need lots of reading material to practice. But curating a mix of narrative and nonfiction books based on students' reading levels and interests can be challenging.

In response, AI tools such as LitLab, Project Read, and Storywizard.ai can find appropriate texts in the classroom library or generate new ones, determine appropriate teaching strategies based on students’ voice performance, and include comprehension checks in the teaching process to ensure that students understand what they are learning. They can also ensure that vocabulary is reinforced in the text rather than appearing only once, which would increase the difficulty of learning.

As speech-to-text capabilities have improved dramatically, tools such as Microsoft Reading Coach, Ello, Edsoma, and Amira have incorporated the ability to listen to students read, give real-time feedback, and build learning paths based on reading science.

Kids can customize characters, put themselves in the story, choose how the plot unfolds (learning about story structure), and even change the illustration style.

3. Curiosity Coach

A student had just finished a radio transmission class. She was curious about why antennas could carry signals, but wires in a DC circuit couldn’t. So she asked the AI ​​tool and got an illustrated concept map showing her exploration path and other exploration suggestions. The student then went through a gamma ray introduction and into the topic of black holes, connecting to the exploration she had ended two weeks earlier. In the end, she realized she needed to learn more about black holes.

In the lives of K-12 students, there is little time for independent exploration. AI tools are expected to play a role in this regard. They can retain a large amount of information, more than any human. Therefore, they can provide more and broader knowledge resources to help students conduct independent exploration and learning.

HelloWonder is a safe browser with a chatbot designed to satisfy the curiosity of young children. Curio is similar to Moxie, but uses a voice robot. Portola is also suitable for young children, but focuses more on creativity.

SocratiQ, like Miro + Wikipedia. This kind of tool is like walking through a large library, or like the Young Lady's Illustrated Primer in Neal Stephenson's novel The Diamond Age.

These tools can also serve as social tools, connecting kids with like-minded peers and sparking interest in further research or new projects. They can also provide information snippets to parents, teachers, and experts so they can guide children’s curiosity and enter the world of expertise in a field.

4. Students of students

The key to the Feynman learning method: pretend to teach what you have learned to a child.

AI tools can play the role of "child" to teach students a given topic; act as a Socratic questioner to deepen students' understanding; act as an NPC in an educational game to complete relevant tasks, acting as a student of the students.

5. Good news for "double A" learners

Suppose you are a high school student with ADHD, your AI assistant can log in and browse the school's LMS, help you break down your homework into manageable tasks, give DDLs and reminders. The tool can track your behavior, provide suggestions for the best working time periods, and more. Moreover, it also adds game features to reward concentration and task completion.

AI tools can be particularly effective in providing additional support to students with disabilities throughout their education.

AI tools can provide assistive writing support (such as startup prompts), alternative input methods (such as speech recognition), executive function support (such as planning tools), visual support, and real-time classroom support (such as text-to-speech and speech-to-text).

AI personalized learning tools can adjust the teaching pace, provide alternative explanations, and provide support for specific skills. For example, Goblin.tools uses AI to help students with Asperger's complete tasks such as writing a book report and break them down into a series of simpler steps.

etc.