夜雨聆风I'm excited to share a few spicy thoughts on artificial intelligence (AI). But first, let's get philosophical (ˌfɪləˈsɑːfɪkəl, 哲学的) with this quote by Voltaire伏尔泰, an 18th century Enlightenment (ɪnˈlaɪtənmənt, 启蒙运动) philosopher, who said, "Common sense is not so common." Turns out事实证明 this quote couldn't be more relevant (ˈrɛləvənt, 相关的) to AI today.
Despite that, AI is an undeniably (ˌʌndɪˈnaɪəbli, 无可否认地) powerful tool, beating the world-class "Go" champion, acing college admission (ədˈmɪʃən, 录取) tests and even passing the bar exam. I’m a computer scientist of 20 years, and I work on AI. I am here to demystify (diːˈmɪstɪfaɪ, 揭开神秘面纱) AI. So AI today is like a Goliath (ɡəˈlaɪəθ, 巨人). It is literally (ˈlɪtərəli, 确实地) very, very large. It is speculated (ˈspɛkjəˌleɪtɪd, 推测) that the recent ones are trained on tens of thousands of GPUs and a trillion (ˈtrɪljən, 万亿) words.
Such extreme-scale AI models, often referred to as "large language models," appear to demonstrate (ˈdɛmənˌstreɪt, 展示) sparks of AGI, artificial general intelligence (通用人工智能). Except when it makes small, silly mistakes, which it often does. Many believe that whatever mistakes AI makes today can be easily fixed with brute force蛮力, bigger scale and more resources. What possibly could go wrong?
There are three immediate (ɪˈmiːdiət, 直接的) challenges we face already at the societal (səˈsaɪɪtəl, 社会的) level. First, extreme-scale AI models are so expensive to train, and only a few tech companies can afford to do so. So we already see the concentration (ˌkɒnsənˈtreɪʃən, 集中) of power. But what's worse for AI safety, we are now at the mercy of those few tech companies because researchers in the larger community (kəˈmjuːnɪti, 社区) do not have the means to truly inspect (ɪnˈspɛkt, 检查) and dissect (dɪˈsɛkt, 剖析) these models. Let's not forget their massive carbon footprint (ˈkɑːbən ˈfʊtprɪnt, 碳足迹) and the environmental (ɪnˌvaɪrənˈmɛntəl, 环境的) impact.
Then there are additional intellectual (ˌɪntɪˈlɛktʃuəl, 智力的) questions. Can AI, without robust, common sense, be truly safe for humanity (hjuːˈmænɪti, 人类)? And is brute-force scale really the only way and even the correct way to teach AI?
I’m often asked these days whether it's even feasible (ˈfiːzəbəl, 可行的) to do any meaningful research without extreme-scale compute (kəmˈpjuːt, 计算能力). I work at a university and nonprofit (nɒnˈprɒfɪt, 非营利的) research institute (ˈɪnstɪtjuːt, 研究所), so I cannot afford a massive GPU farm to create enormous (ɪˈnɔːməs, 巨大的) language models. Nevertheless (ˌnɛvəðəˈlɛs, 然而), I believe that there's so much we need to do and can do to make AI sustainable (səˈsteɪnəbəl, 可持续的) and humanistic (ˌhjuːməˈnɪstɪk, 人文主义的).
Know your enemy. We need to make AI smaller, to democratize it. And we need to make AI safer by teaching human norms (nɔːmz, 规范) and values. Perhaps we can draw an analogy (əˈnælədʒi, 类比) from "David and Goliath," here, Goliath being the extreme-scale language models, and seek inspiration (ˌɪnspəˈreɪʃən, 灵感) from an old-time classic, "The Art of War," which tells us, in my interpretation (ɪnˌtɜːprɪˈteɪʃən, 解释), know your enemy, choose your battles, and innovate (ˈɪnəveɪt, 创新) your weapons.
Let's start with the first, know your enemy, which means we need to evaluate (ɪˈvæljueɪt, 评估) AI with scrutiny (ˈskruːtɪni, 仔细审查). AI is passing the bar exam. Does that mean that AI is robust at common sense? You might assume (əˈsjuːm, 假设) so, but you never know. Suppose I left five clothes to dry out in the sun, and it took them five hours to dry completely. How long would it take to dry 30 clothes? GPT-4, the newest, greatest AI system says 30 hours. Not good. Here's a different one: I have a 12-liter jug and a six-liter jug, and I want to measure six liters. How do I do it? Just use the six-liter jug, right? GPT-4 spits out some very elaborate (ɪˈlæbərət, 复杂的) nonsense. Step one, fill the six-liter jug, step two, pour the water from the six to the 12-liter jug, step three, fill the six-liter jug again, step four, very carefully pour the water from the six to the 12-liter jug. And finally you have six liters of water in the six-liter jug that should be empty by now.
Okay, one more. Would I get a flat tire by bicycling over a bridge that is suspended (səˈspɛndɪd, 悬挂的) over nails, screws and broken glass? Yes, highly likely, GPT-4 says, presumably (prɪˈzjuːməbli, 大概) because it cannot correctly reason that if a bridge is suspended over broken nails and broken glass, then the surface of the bridge doesn't touch the sharp objects directly.
So how would you feel about an AI lawyer that aced the bar exam yet randomly fails at such basic common sense? AI today is unbelievably (ˌʌnbɪˈliːvəbli, 难以置信地) intelligent and then shockingly (ˈʃɒkɪŋli, 惊人地) stupid. It is an unavoidable (ˌʌnəˈvɔɪdəbəl, 不可避免的) side effect of teaching AI through brute-force scale. Some scale optimists (ˈɒptɪmɪsts, 乐观主义者) might say, “Don’t worry about this. All of these can be easily fixed by adding similar examples as yet more training data for AI." But the real question is this: why should we even do that? You are able to get the correct answers right away without having to train yourself with similar examples.
Choose your battles. Children do not even read a trillion words to acquire (əˈkwaɪər, 获得) such a basic level of common sense. This observation (ˌɒbzəˈveɪʃən, 观察) leads us to the next wisdom (ˈwɪzdəm, 智慧), choose your battles. So what fundamental (ˌfʌndəˈmɛntəl, 基本的) questions should we ask right now and tackle (ˈtækəl, 处理) today in order to overcome (ˌəʊvəˈkʌm, 克服) this status quo (ˌsteɪtəs ˈkwəʊ, 现状) with extreme-scale AI? I'll say common sense is among the top priorities (praɪˈɒrɪtiz, 优先事项).
Common sense has been a long-standing challenge in AI. To explain why, let me draw an analogy to dark matter (dark matter, 暗物质). Only five percent of the universe (ˈjuːnɪvɜːs, 宇宙) is normal matter that you can see and interact with, and the remaining 95 percent is dark matter and dark energy (ˈɛnədʒi, 能量). Dark matter is completely invisible (ɪnˈvɪzəbəl, 看不见的), but scientists speculate that it's there because it influences (ˈɪnfluənsɪz, 影响) the visible world, even including the trajectory (trəˈdʒɛktəri, 轨迹) of light. For language, the normal matter is the visible text, and the dark matter is the unspoken rules about how the world works, including naive physics (fɪzɪks, 物理学) and folk psychology (saɪˈkɒlədʒi, 心理学), which influence the way people use and interpret (ɪnˈtɜːprɪt, 解释) language.
Why is this common sense even important? In a famous thought experiment (ɪkˈspɛrɪmənt, 实验) proposed by Nick Bostrom, AI was asked to produce and maximize (ˈmæksɪmaɪz, 最大化) paper clips. That AI decided to kill humans to utilize (ˈjuːtɪlaɪz, 利用) them as additional resources, to turn you into paper clips. Because AI didn't have the basic human understanding about human values. Writing a better objective (əbˈdʒɛktɪv, 目标) and equation (ɪˈkweɪʒən, 方程) that explicitly (ɪkˈsplɪsɪtli, 明确地) states “Do not kill humans” will not work either because AI might go ahead and kill all the trees, thinking that's a perfectly okay thing to do. In fact, there are endless other things that AI obviously shouldn’t do while maximizing paper clips, including “Don’t spread fake news,” “Don’t steal,” “Don’t lie,” which are all part of our common sense understanding about how the world works.
However, the AI field for decades has considered common sense as a nearly impossible (ɪmˈpɒsəbəl, 不可能的) challenge. So much so that when my students and colleagues and I started working on it several years ago, we were very much discouraged (dɪsˈkʌrɪdʒd, 气馁的). We’ve been told that it’s a research topic of the '70s and '80s; we shouldn't work on it because it will never work; in fact, don't even say the word to be taken seriously. Now fast forward to this year, I’m hearing: “Don’t work on it because ChatGPT has almost solved it.” And: “Just scale things up and magic will arise, and nothing else matters.” My position is that giving true common sense, human-like common sense, to AI is still a moonshot (ˈmuːnʃɒt, 登月计划，比喻巨大挑战). You don’t reach the Moon by making the tallest building in the world one inch taller at a time. Extreme-scale AI models do acquire an ever-increasing amount of commonsense knowledge, I'll give you that. But remember, they still stumble (ˈstʌmbəl, 绊倒) on such trivial (ˈtrɪviəl, 琐碎的) problems that even children can do. So AI today is awfully (ˈɔːfli, 非常地) inefficient (ˌɪnɪˈfɪʃənt, 低效的). What if there is an alternative (ɔːlˈtɜːnətɪv, 替代的) path or a path yet to be found?
Innovate your weapons. This leads us to our final wisdom: innovate your weapons. In the modern-day AI context (ˈkɒntɛkst, 背景), that means innovate your data and algorithms (ˈælɡərɪðəmz, 算法). There are roughly speaking three types of data that modern AI is trained on: raw web data, crafted examples custom (ˈkʌstəm, 定制的) developed for AI training, and human judgments (ˈdʒʌdʒmənts, 判断), also known as human feedback (ˈfiːdbæk, 反馈) on AI performance (pəˈfɔːməns, 表现). If AI is only trained on the first type, raw web data, which is freely available, it's not good because this data is loaded with racism (ˈreɪsɪzəm, 种族主义) and sexism (ˈsɛksɪzəm, 性别歧视) and misinformation (ˌmɪsɪnfəˈmeɪʃən, 错误信息). No matter how much of it you use, garbage in and garbage out. So the newest, greatest AI systems are now powered with the second and third types of data that are crafted and judged by human workers. It's analogous (əˈnæləɡəs, 类似的) to writing specialized (ˈspɛʃəlaɪzd, 专门的) textbooks for AI to study from and then hiring human tutors (ˈtjuːtəz, 导师) to give constant feedback to AI. These are proprietary (prəˈpraɪətəri, 专有的) data, by and large, speculated to cost tens of millions of dollars. We don't know what's in this, but it should be open and publicly available so that we can inspect and ensure it supports diverse (daɪˈvɜːs, 多样的) norms and values. For this reason, my teams at UW and AI2 have been working on commonsense knowledge graphs (ɡræfs, 图谱) as well as moral (ˈmɒrəl, 道德的) norm repositories (rɪˈpɒzɪtəriz, 知识库) to teach AI basic commonsense norms and morals. Our data is fully open so that anybody can inspect the content and make corrections as needed because transparency (trænsˈpærənsi, 透明) is the key for such an important research topic.
Now let's think about learning algorithms. No matter how amazing large language models are, by design they may not be the best suited to serve as reliable (rɪˈlaɪəbəl, 可靠的) knowledge models. These language models do acquire a vast (vɑːst, 巨大的) amount of knowledge, but they do so as a byproduct (ˈbaɪˌprɒdʌkt, 副产品) as opposed to a direct learning objective, resulting in unwanted side effects such as hallucinated (həˈluːsɪneɪtɪd, 幻觉的) effects and lack of common sense. In contrast, human learning is never about predicting (prɪˈdɪktɪŋ, 预测) which word comes next, but it's really about making sense of the world and learning how the world works. Maybe AI should be taught that way as well.
As a quest (kwɛst, 追求) toward more direct commonsense knowledge acquisition (ˌækwɪˈzɪʃən, 获得), my team has been investigating potential (pəˈtɛnʃəl, 潜在的) new algorithms, including symbolic (sɪmˈbɒlɪk, 符号的) knowledge distillation (ˌdɪstɪˈleɪʃən, 蒸馏，提炼), which can take a very large language model and crunch (krʌntʃ, 压缩处理) it down to much smaller commonsense models using deep neural (ˈnjʊərəl, 神经的) networks. In doing so, we also generate algorithmically (ˌælɡəˈrɪðmɪkli, 通过算法) human-inspectable, symbolic commonsense knowledge representation (ˌrɛprɪzɛnˈteɪʃən, 表示), so that people can inspect and make corrections and even use it to train other neural commonsense models. More broadly, we have been tackling this seemingly impossible giant puzzle (ˈpʌzəl, 难题) of common sense, ranging from physical (ˈfɪzɪkəl, 物理的), social and visual (ˈvɪʒuəl, 视觉的) common sense to theory of minds, norms and morals. Each individual piece may seem quirky (ˈkwɜːki, 古怪的) and incomplete (ˌɪnkəmˈpliːt, 不完整的), but when you step back, it's almost as if these pieces weave together into a tapestry (ˈtæpɪstri, 织锦) that we call human experience (ɪkˈspɪəriəns, 经验) and common sense.
We're now entering a new era (ˈɪərə, 时代) in which AI is almost like a new intellectual (ˌɪntɪˈlɛktʃuəl, 智力的) species (ˈspiːʃiːz, 物种) with unique (juːˈniːk, 独特的) strengths and weaknesses compared to humans. In order to make this powerful AI sustainable and humanistic, we need to teach AI common sense, norms and values. Thank you.
Chris Anderson: This is so interesting, this idea of common sense. We obviously all really want this from whatever's coming. But help me understand. We've had this model of a child learning. How does a child gain common sense apart from the accumulation (əˌkjuːmjəˈleɪʃən, 积累) of more input and some human feedback? What else is there?
Yejin Choi: So fundamentally, there are several things missing, but one of them is, for example, the ability to make hypotheses (haɪˈpɒθɪsiːz, 假设) and conduct experiments, interact with the world and develop these hypotheses. We abstract away the concepts about how the world works, and then that's how we truly learn, as opposed to today's language model. Some of that is really not there quite yet.
Chris Anderson: You use the analogy that we can’t get to the Moon by extending a building a foot at a time. But the experience that most of us have had of these language models is not a foot at a time. It's a breathtaking (ˈbrɛθˌteɪkɪŋ, 惊人的) acceleration (əkˌsɛləˈreɪʃən, 加速). Are you sure that given the pace at which those things are going, each next level seems to be bringing with it what feels kind of like wisdom and knowledge?
Yejin Choi: I totally agree that it's remarkable (rɪˈmɑːkəbəl, 非凡的) how much scaling things up really enhances (ɪnˈhɑːnsɪz, 提升) the performance across the board. So there's real learning happening due to the scale of the compute and data. However, there's a quality (ˈkwɒlɪti, 质量) of learning that is still not quite there. The thing is, we don't yet know whether we can fully get there or not just by scaling things up. If we cannot, then there's this question of what else? And even if we could, do we like this idea of having very extreme-scale AI models that only a few can create and own?
Chris Anderson: I mean, if OpenAI said, "We're interested in your work, we would like you to help improve our model," can you see any way of combining what you're doing with what they have built?
Yejin Choi: Certainly what I envision (ɪnˈvɪʒən, 设想) will need to build on the advancements of deep neural networks. It might be that there’s some scale Goldilocks Zone (ˈɡəʊldɪlɒks zəʊn, 适中区间), such that ... I'm not imagining that smaller is always better either, by the way. It's likely that there's a right amount of scale, but beyond that, the winning recipe (ˈrɛsɪpi, 配方) might be something else. So some synthesis (ˈsɪnθɪsɪs, 综合) of ideas will be critical here.
Chris Anderson: Yejin Choi, thank you so much for your talk.
