Introduction & Brief Overview

At Paideia, we believe AI can elevate human capabilities. Our mission is to build revolutionary educational products that harness artificial intelligence in order to improve human intelligence.

We are currently developing an AI product designed to nurture critical thinking skills. Our AI is multifaceted, acting either as tutor, interviewer, or both, depending on the task at hand.

We will offer the tutoring capabilities for free, to help propogate critical thinking skills for all users that choose to engage with our project, and in turn provide us with data to refine our model and enhance the evaluation process.

The interviewing capabilities, on the other hand, are the core of our business case. Through dialoguing with users, our AI assesses their abilities, and these evaluations can be marketed to university admissions or corporate human resources offices as a means for screening applicants.

Our product is unique on both the frontend and backend. For users, it engages in proactive Socratic dialogue, personalizing instruction and adapting in real time during every interaction. Internally, it is built with a complex cognitive architecture that mirrors the nuanced thought processes of a human educator or interviewer.

The Paideia project grew out of seeds planted during Zuzalu crosspollination. When Yihan and Veronica met during the innovative Zuzalu experiment in early 2023, they discussed the future of education in the era of artificial intelligence for hours, finding alignment both in broad visions for the future and the specific desire to harness emerging LLM technologies to help students build critical thinking skills. Paideia was born and has been growing ever since, with an initial demo launching this month.

Defining Characteristics

Far from being a passive tool, our AI Tutor engages its users in a Socratic dialogue, initiating the conversation rather than merely responding to queries. This is a distinct departure from current models, such as the AI utilised by Khan Academy, which patiently waits for a question before it springs into action. Instead, our AI takes a proactive approach, as any educator should when seeking to build a student's skills.

The second defining characteristic of our AI Tutor is its personalization capabilities. Personalization is a two-tiered concept in this context. The first level relates to adaptive learning. The AI understands the progress of each individual learner, identifying their strengths and weaknesses and tailoring the instruction accordingly. The second level involves adapting the teaching to align with the learner's interests. For instance, if a high school student has a keen interest in international politics, the AI Tutor could use debates on the Russia-Ukraine war to illustrate the principles of critical thinking.

Taking personalisation a step further, the AI Tutor can even draw from the learner's personal life to make the learning experience engaging and relevant. Therefore, this AI Tutor is much more than a simple teaching tool. It is also an interviewer and an evaluator, capable of providing insightful feedback and assessing the learner's abilities and progress in a nuanced manner.

Why is a critical thinking evaluation necessary? What about essays?

Now, let's take a step back and ponder the value proposition of this product. During the process of studying abroad, AI can directly aid in the essay writing part of the application. AI-assisted essay writing, even to the point of revising or ghostwriting essays, is technically feasible. But why don't we pursue this path? The answer is rooted in the fundamental purpose of the essay - a tool for universities to differentiate and select students.

If AI could enable everyone to write a standout essay, would universities continue to rely on this tool? It's likely they wouldn't. Therefore, we believe the value of the essay as a selection tool is a short-term proposition. It's not just AI that could erode its value. The essay selection method itself is a low-cost compromise, even in a time before AI. If essays were genuinely useful, wouldn't companies like Google and McKinsey use them in their hiring processes? Companies care deeply about hiring because every new hire represents an investment and potential for productivity. If the hire performs poorly, the company suffers losses. Rather, companies are known for asking applicants to solve challenging problems and justify their answers as a means for demonstrating their cognitive capabilities.

In the context of university recruitments, the bar for a perfect fit isn't as rigid as in the corporate world. A university can afford to admit a candidate who may not be the textbook perfect fit, provided a certain proportion of the admitted students are suitable. For instance, if 20% of the students a university admits end up as benefactors to the institution, it's likely considered a success. Even if a university errs in its recruitment process, and a student turns out not to be a good fit, the institution doesn't suffer a significant loss since it still collects tuition fees.

Universities with greater resources are already known to rely on interviews rather than mere essays, such as in the case of Oxford and Cambridge, which have professors conduct interviews with each applicant. In this context, the use of essays for university admission seems to be a mere compromise made due to limited resources, rather than a particularly crucial method for assessing personal character.

Consequently, we believe universities recognize that interviews are a superior method for student recruitment, but they're costly and difficult to standardize. Inconsistencies can easily arise from haphazard interviews. Standardizing interviews is, in itself, a high-cost process, in addition to the human resources involved.

What exactly are universities evaluating?

Over the years, due to our background in overseas study consultancy, we've had the chance to observe student applications and the feedback they receive from admissions officers. We've come to realize that, beyond objective academic achievements, universities place great importance on a student's thinking abilities, including creativity, depth of thought, and clarity of expression. One of these abilities, often referred to directly by universities, is critical thinking.

Why haven't universities been explicitly assessing this ability, even though they deem it important? It's challenging to evaluate this ability, especially on the scale of thousands of applicants per year. The term 'critical thinking' itself is somewhat vague, overlapping with rationality, argumentation, and debate. Defining it and figuring out how to measure it is difficult.

For instance, university applications often emphasize a person's debate skills, public speaking skills, essay writing skills, and their ability to express independent thought in essays rather than blind adherence to the norm. Even the SAT, GRE, and other tests used for undergraduate and graduate applications contain elements that point towards critical thinking. These are indirect assessments of critical thinking.

Critical Thinking in the AI Era

Critical thinking, despite its ambiguities, is widely valued, and quite dear to us at Paideia. In the era of AI, we believe critical thinking is all the more important, as AI tools increasingly saturate every aspect of society. We suspect that within a few years, all workplaces and every individual will be working alongside large language models or chatbots.

Information will come not from your colleagues, but from artificial intelligence, and these AI systems will likely have a more extensive knowledge base than you do, with the ability to articulate seemingly rational arguments. When you're working with such language models, you'll inevitably need to judge the credibility of their statements and know how to understand and handle what they say. In an era where AI systems seem stronger than humans, yet can certainly still make mistakes, everyone will need improved critical thinking capabilities to avoid making potentially problematic decisions.

Traditional Testing Methods vs. AI

Traditional critical thinking testing methods have a number of issues that we believe AI technologies can overcome. Traditional tests can be divided into two categories: problem-solving and self-expression or self-reporting.

Self-reporting tests, such as personality tests and essays, are typical examples of the self-reporting category, where you tell others about your qualities and capabilities. This category often appears in interviews but is also prevalent in examinations. As in any subjective claim, self-reporting is prone to fabrication, making it inefficient and therefore unnecessarily expensive from a screening perspective.

The other category is problem-solving tests, which often appear in exams but can also feature in interviews, such as the famous Google interview challenges. Though these tests have their benefits, they're prone to being gamed if the test format remains consistent, which can undermine the accuracy of the test.

This is not necessarily what you, as a screener, want to see. It's reminiscent of Goodhart's law: when a measure becomes a target, it ceases to be a good measure. People will prepare for the target.

An AI-based interview can avoid the above issues. For instance, it doesn't have to rely on multiple-choice questions. It can be more interactive, asking students about their thought process behind answering a question, yet not performing its own evaluation rather than taking their self-reporting at face value.

With such an interview process, people would of course still prepare for the target. But in this case, we welcome it! Preparing for our AI-powered critical thinking interview by dialoguing with our AI tutor will enable students to build their critical thinking skills, which is our ultimate goal. And as the evaluation capabilities are central to the tutoring capabilities, the measure remains accurate.

Our Technology

Our product boasts a complex internal cognitive architecture that replicates the intricate thought process a teacher or interview goes through when interacting with a student. This system continues to evolve and we already see immense promise in how our intricate network of prompt relationships results in sophisticated outcomes far beyond a single prompt call to a large language model.

To illustrate this, consider a large language model like GPT-4 as a compiler and prompts as function definitions in a program. Just as a program is composed of a sophisticated combination of functions, with intricate calling relationships, logical flows, and even recursive relationships, the use of GPT-4 can be enhanced by a complex combination of prompts, rather than a single prompt.

This concept serves as the foundation for our AI interviewer and tutor.

AI Alignment

As a new AI company built by avid members of the rationality community, AI alignment is a key topic of discussion.

In building our critical thinking AI tutor, we've been grappling with a crucial alignment issue. Our current endeavor isn't solely about escalating the AI's capabilities. We provide context and additional knowledge through instructional materials, tools, and outlines. The crux of our mission, however, resides in shaping the AI's thought process and responses to resemble those of an effective and empathetic human teacher, while utilizing the superhuman capabilities of an AI. This approach, particularly in tutoring and interviewing scenarios, is pivotal.

Let's consider a less consequential setting, like playing Go. It doesn't matter how the AI plays the game, as long as it wins. Its adherence to human norms or philosophies about the game is irrelevant. It can even invent novel strategies. However, this laissez-faire attitude doesn't hold up when we transition to education and interviewing, where the AI is engaged in human-centric tasks.

Applying a potentially more efficient but inexplicable method to judge a person won't garner acceptance. It's akin to the well-known psychological effect where people tend to distrust mechanical processes that they don't understand. Hence, it's imperative for AI in education and assessment to mimic the thought process of the best human experts in the field. We're currently working with top critical thinking mentors, dissecting their dialogues, exploring their subtle decisions and tools in conversations, all with proper permissions and confidentiality, of course.

These mentors employ a diverse array of logical tools in their dialogues. Our aim is to distil these elements and feed them to the AI as prompts. Although this process may be slow at the moment, given that we have to use the AI multiple times, we're steadily working towards improving the model's speed and complexity. We anticipate being able to merge prompts and allow the AI to update the student and mentor statuses simultaneously, thereby enhancing efficiency. We might also employ a renowned technique from alignment research, Iterative distillation Amplification (IDA), to train another model using data generated by a prompt-based structure built on GPT-4.

As for our current progress, we've implemented an initial prompt framework with two statuses and rudimentary capabilities for tutoring and assessment. The results are still a bit rough around the edges, but it already displays better performance than the original GPT-4.

Our first version was designed to teach about cognitive biases and logic, but we felt these subjects were too broad, and would require greater complexity in the cognitive architecture. Therefore, we have currently narrowed our focus to teach argumentation, a more specific task that enables us to better test, iterate, and improve ourAI's tutoring and evaluation capabilities.

Gitcoin

Our ongoing work, while bearing commercial intentions, holds significant societal value. In addition to our commercial (assessment) and non-commercial (tutoring) aspects, we plan on open-sourcing parts of our project to enable others to build complex and effective AI projects.

While the assessment component is clearly commercial, the tutoring component is non-profit. The open-source component refers to the complex interrelationships between prompts we’re currently using. The development and testing of these are indeed laborious, for which we have built specific tools. We're considering open-sourcing these tools so others don't need to reinvent the wheel when building with AI.

Our current fundraising on GitCoin is primarily for the non-commercial aspects of our project. We’ll publicly share our progress on the open-source tools, team salaries, and the extent of developments in the education component.

We're eager for more collaboration and welcome contributions to the open-source and tutoring components of our project. If you're interested in the commercial aspect, we'd also love to discuss our project with you, receive guidance, and if you could refer potential users or investors, we would greatly appreciate it. As we tread this exciting path, every bit of support and collaboration brings us one step closer to our vision.

Zuzalu

Zuzalu is on a mission to foster a global network of communities to advance humanity by creating playgrounds at the intersection of free and open technology, health, science and social innovation. Co-founders Yihan Zhou and Veronica Schrenk met at the inaugural Zuzalu playground, bonding over shared fascinations for the intersection of artificial intelligence and education, as well as shared values of rationality and critical thinking.

Conversing about the future of education as artificial intelligence advances and spreads, we discussed the immense opportunities AI brings to revolutionize education and learning for students all over the world, and discovered we had individually been ideating on the same idea: to build a tool built on a language model that helps humans develop critical thinking skills. We have been working together to build this project ever since, with a small team of aligned builders equally excited about making our vision a reality.

As Zuzaluans, we also seek to advance humanity with our project, developing a free tutor that can help build some of the most important skills of our time and open source frameworks for building with and "educating" artificial intelligence tools to be aligned with human advancement.

Grand Narrative

There are several grand narratives I’d like to explore in the context of our AI tutor/interviewer project.

The first is the role of the AI tutor as an excellent coach. The ability to teach and enhance critical thinking also implies the capability to help deconstruct and analyze real-world predicaments. Imagine a tutor that you've interacted with since high school. It's not just your academic confidante, but also a guide you can turn to with life's dilemmas. This AI tutor could be the entry point for many teenagers to understand and interact with AI. Although game characters could serve a similar function, you're unlikely to consult a game character about life's crucial decisions, whereas you might consult your teacher. This is an exciting opportunity to reimagine the possibilities of AI.

The second narrative revolves around how our project could assist in human resource allocation and responsibility distribution, reducing costs and friction in the process. In a way, it could significantly uplift humanity's overall productivity. Our project doesn't merely aim to create a powerful AI that can do everything; instead, it aims to make human communication more efficient. You'll be communicating and exchanging ideas with someone ideally suited to understand and help you. Although our commercial scenarios mostly involve job selections, university choices, and social interactions, the AI's role in facilitating human communication and improving organizational effectiveness is a classic case of AI aiding not just individual efficiency, but societal productivity as a whole.

The third narrative concerns alignment, an area our team is deeply interested in. Our work on AI is more about alignment than capability enhancement. We want the AI to act according to our intentions, rather than just making it smarter. We're using aligned AI to align humans, which could potentially initiate an interesting feedback loop in AI alignment, where AI and humans are aligning each other. This involves both inner alignment and outer alignment, and we believe our project could significantly contribute to important research in AI alignment.

Lastly, our project addresses the issue of educational inequity. Many initiatives are addressing subject-specific disparities. However, disparities in soft skills such as critical thinking and social-emotional abilities are perhaps more severe. They lack sufficient institutions or even online resources for improvement. If your personal environment isn't conducive to developing these skills, you might always lack them, especially if you come from an economically disadvantaged background. This is why we see high social value in our AI tutor – it could potentially bridge this gap, making quality education accessible to all.