OpenAI解决300多道数学难题背后,AI下一步是自己开公司?|AI泰斗Kevin Knight TGC 2026 演讲实录

2026-11-12

一个AI,独立解出数学难题、连续数周自主做实验,下一步——自己开公司、自己赚钱?

这不是科幻片的情节。在钛动科技于近日举行的TGC 2026科技增长大会上,世界顶级NLP专家Dr.Kevin Knight把这幅图景摆到了我们面前。他曾任国际计算语言学协会(ACL)主席、当选 AAAI Fellow,从南加州大学计算机科学教授到如今的 Threeven Labs CEO,他走完了AI从论文到商业化的全流程,并带出Transformer一作Ashish Vaswani、DeepMind研究副总裁Barret Zoph。他的开场演讲,直击本届TGC核心命题“AI is Profitable”。以下为Dr. Kevin Knight的演讲内容整理:

今天,我想围绕人工智能谈五个问题:我们是如何走到今天的?AI 究竟是如何工作的?技术进步是否正在加速?下一次重大突破可能来自哪里?以及,AI 将如何改变我们的生活?

我想先讨论 AI 在科学领域的发展历程与未来。

十年前,我自己编写了一个下棋机器人。它可以在真正的棋盘上下棋。直到今天,我每天早上还是会先给自己煮一杯咖啡,然后和这个机器人下一盘棋。这是我非常喜欢的一件事。

大约也是在那个时期,DeepMind 开发了著名的 AlphaGo 和 AlphaZero 等系统。它们最初通过学习人类棋手的棋谱进行训练,后来出现了更强大的系统,能够完全通过自我对弈来学习。它们通过自己的经验,逐渐掌握哪些策略更有利于赢得比赛。

你可以想象,AlphaGo 回到自己的房间里,不断练习围棋,一个星期之后再走出来,就已经成为世界冠军。

不过,这并没有让我太担心。毕竟,下棋本来就是我的爱好,我每天早上依然很享受和机器人对弈的过程。

图片

Dr. Kevin Knight 开发的国际象棋机器人与 DeepMind AlphaGo(图片来源:Dr. Kevin Knight TGC 大会演讲 PPT)

多年来,我真正从事的研究领域是自然语言处理(NLP)。我们尝试让机器完成语言翻译、回答问题等任务。在早期,这些问题极其困难。读研究生时,我妻子甚至会告诉我们的朋友,她觉得我很辛苦,因为我一直在研究一些几乎不可能解决的问题。

多年来,我们一直在尝试构建语言模型。为什么要让机器做语言建模?其中一个任务看起来甚至有些奇怪:例如,给机器一句话:“Mary saw toys in the window and wanted...”,然后让它预测接下来最可能出现的词。人类当然知道,她想要的是橱窗里的玩具,而不是窗户。但机器必须通过学习才能理解这一点。

通过这样的训练,机器逐渐学习到的不只是语言本身,还包括世界、人类社会,以及事物之间的关系。虽然这看起来只是一个简单的预测任务,但它实际上为机器理解世界提供了一条重要路径,也帮助我们解决了翻译等一系列复杂问题。

图片

从 1940 年到 ChatGPT 时代语言模型的发展脉络(图片来源:Dr. Kevin Knight TGC 大会演讲 PPT)

到了 20 世纪 80 年代,IBM 在语言建模领域取得了重要突破。研究人员尝试利用前面几个词来预测下一个词。机器开始通过阅读文本学习语言。

后来,Google 也基于这些技术构建了规模更大的语言模型。我们当时以为,自然语言处理将沿着这条路线持续发展,逐步取得进步。但 2016 年前后,情况发生了巨大变化。循环神经网络、Transformer,以及后来的 ChatGPT,推动了语言模型能力的快速跃迁。

到了今天,我们在很大程度上已经解决了自然语言处理领域的一些核心问题:机器能够理解语言、生成语言,也能够进行翻译。这是一场非常重要的技术突破。

 

一、AI为什么能够取得突破

AI 取得突破,有两个非常重要的原因。 

第一个是训练规模。现在,我们可以使用互联网级别的数据来训练,机器能够从这些海量样本中学习规律。第二个是后训练(Post-training)。以聊天系统为例,人类会对机器的回答进行评价:这个回答好,那个回答不好。通过这种方式,机器能够利用人类反馈持续改善自己的表现。

所以,AI 的进步不仅来自更大规模的数据和计算,也来自更加有效的学习机制。

图片

1980 年至今 NLP 领域研究者心态的起伏变化(图片来源:Dr. Kevin Knight TGC 大会演讲 PPT)

我想再展示一张图。这张图的时间跨度依然是从 1950 年到 2030 年,但这次我不想展示机器的准确率,而是想展示作为 AI 研究者,我们在不同阶段的心情。

一开始,我们很开心,因为研究本身很有趣。虽然进步不大,但我们并不知道自己正处于停滞阶段。 

后来,统计学习方法出现了。一部分研究者很兴奋,因为机器终于能够从文本中学习;另一部分研究者却不太高兴,因为他们更喜欢语言学,而不是统计学。 

再后来,我们在机器翻译等领域取得了持续进步。比如,中英文翻译的效果一点点变好。当时也有人开始创办机器翻译公司,但整体而言,AI 创业仍然是一件非常困难的事情。

接着,神经网络革命到来。许多研究者不得不放弃过去积累的技术和经验,转而采用全新的方法。虽然这让整个研究领域经历了巨大的震荡,但技术确实取得了更好的结果。 

到了 GPT 时代,自然语言处理领域的研究格局再次发生变化。从某种程度上说,我们解决了过去长期追求的研究问题,但与此同时,也出现了数以百万计的新应用可能性。

对于研究者来说,每一次技术革命都意味着一次巨大的冲击。而我的选择是:既然语言问题已经取得了巨大进展,那就去研究更加困难的问题。

过去,我认为文字非常复杂。一个词可能有很多含义,不同的词组合在一起,就会产生更加复杂的意思。但机器最终学会了处理这些问题。于是,我把研究方向转向了数学。

我原本以为数字会更加规整,比如乘法口诀。但真正深入研究后,我发现数学同样充满了复杂性,甚至世界上最优秀的数学家,也无法解决一些看起来非常简单的数学猜想。我希望利用 AI 来加速数学研究的进步。

几年前,我开始研究如何利用 AI 推动数学研究。而今年夏天,数学领域出现了一场巨大的变化。AI 已经能够解决一些极其困难的数学问题,其中甚至包括千禧年数学难题。这些问题由世界上最优秀的数学家选出,被认为是数学领域最困难、最有价值的挑战之一。

再看看OpenAI,他们近期宣布自己攻克了 300 多个重要的数学难题,提出了全新的证明过程与证明思路。这些成果展示了 AI 在数学证明和提出新证明思路方面的巨大潜力。这对数学界产生了很大的冲击,就像过去神经网络技术改变自然语言处理领域一样。

为什么 AI 能够做到这一点? 首先,数学领域存在大量公开的研究论文。机器能够阅读这些论文,学习不同数学家在不同领域积累的知识。这使得 AI 不再局限于某一个数学分支,而是能够建立跨领域的知识联系。 

但仅仅阅读还不够。另一个重要能力是推理与验证。就像 AlphaGo 能够判断自己是否下出了正确的一步棋一样,在数学研究中,机器也需要判断自己的证明是否成立。如果证明可以被验证,AI 就能获得明确的反馈,并据此继续改进。

于是,我们再次看到了一种熟悉的模式:机器可以自己进入一个房间,持续研究数学问题,经过一段时间后,再带着新的成果回来。这意味着 AI 正在获得一种重要能力:它不再只是等待人类给出答案,而是能够通过自主探索、推理和验证,不断推动自己的进步。

 

二、下一次AI突破可能发生在哪里

从大家的个人体验来看,AI 的进步似乎正在变得越来越快。我也有同样的感受。

无论是在语言领域还是数学领域,几乎每个月都会出现一些让我感到惊讶的新进展。

那么,如何衡量 AI 能力的加速?我认为,一个很有意思的指标是:AI 能够在多长时间内独立解决一个问题,而不需要人类持续与它互动。比如:15 秒回答一个问题,然后等待人类提出下一个问题;10 分钟独立进行网络研究,并生成一份完整的研究报告;1 小时根据需求独立构建一个小型软件系统;10 小时持续工作,独立完成一个大型软件系统。

这些能力在今年开始变得越来越现实,而在去年还很难实现。

如果把这个指标推向极端,我们就会看到更加惊人的案例。今年夏天,有一个非常复杂的数学问题被成功解决。整个过程相当于机器进行了长达 100 年的思考,而不需要人类持续干预。当然,机器并不是真的运行了 100 年,而是由一万个智能体自主协作,连续工作了 88 个小时,最终完成了任务。

所以,如果你想衡量 AI 的进步,不妨关注它能够独立工作多长时间。在未来,机器会越来越擅长处理需要长时间思考、规划和执行的复杂任务。这意味着 AI 的能力边界正在不断扩大。

刚才我介绍了自己亲身参与过、并且 AI 已取得巨大突破的领域:语言和数学。那么,接下来会是什么?

图片

五大 AI 未来方向(图片来源:Dr. Kevin Knight TGC 大会演讲 PPT)

我认为,机器人、生物医药研究、用于商业任务的AI、娱乐,以及AI独立经营一家企业,都可能成为 AI 下一轮突破的重要方向。对于这些领域来说,有一个共同的关键因素:反馈。AI 究竟如何变得更好?它通过什么信号判断自己的表现?

在机器人领域,如果机器成功完成一项任务,就应该继续采用有效的方法;如果失败了,就需要分析原因并尝试新的方式。这与 AlphaGo 通过自我对弈学习的逻辑类似。但机器人不能完全依靠内部模拟,还需要大量真实世界的实验与验证。

在生物医药领域,AI 已经开始参与药物发现、分子研究等工作。它可以像阅读数学论文一样,阅读大量生物化学文献,建立超越单一研究者知识边界的跨领域认知。但要真正证明一个新想法有效,仍然需要真实的实验室实验,甚至需要针对人体药物开展临床试验。所以,这个领域还有很多工作需要完成。

在娱乐领域,AI 已经开始创造内容,但它还不能完全独立创作一部人们真正想看的小说或电影。这里的反馈来自人类:人们喜欢什么,不喜欢什么。只有理解这些反馈,AI 才能持续提升创作能力。

而我认为,还有一个可能让所有人感到意外的方向: AI 未来不仅能够为企业工作,还可能独立经营一家企业。在某些领域,由机器经营的企业可能会与人类经营的企业展开竞争。这将是一件非常值得关注的事情。

因为对于一家企业来说,AI 能够获得的反馈信号其实非常清晰:利润。如果企业赚钱,说明它正在做一些正确的事情;如果企业亏损,它就需要调整策略。当然,这也存在挑战。机器可能在还没有学会如何成功经营之前,就已经耗尽了资源。但如果有足够的资源,AI 就可以尝试更多不同的经营策略,通过持续试错,最终找到在某个商业领域取得成功的方法。

利润,将成为 AI 经营一家企业时最直接的反馈信号。

 

三、AI的未来,不只是思考,而是学习与行动

未来有一天,当你在公交车站等车时,可能会发现旁边站着一个机器人。你和机器人一起上车。第一次看到这种场景,你可能会觉得非常震惊。但几周之后,你可能就会习惯它的存在,觉得这不过是生活的一部分。

又比如,当 AI 真正设计出一种能够治愈重大疾病的疗法时,这同样会给人类带来巨大的惊喜和福祉。

但这些进步也会让很多人担忧。一些 AI 研究者非常乐观,他们相信 AI 能够攻克重大疾病,推动人类社会进步。也有一些人非常悲观,甚至开始担心 AI 最终会对人类造成威胁。

我自己并不完全认同这些极端观点。但有一件事,我确实非常在意。

我认为,把乘法交给计算器没有问题。如果你不想自己种粮食,也完全没有问题;如果你不喜欢做饭,也可以让别人帮你准备食物。这些事情都能够让我们拥有更多时间,去做自己真正想做的事。

但问题在于:当我们不再思考时,会发生什么? 在日常生活中,我经常会遇到一些问题。我可以自己思考、尝试解决,也可以直接问 AI,让它替我思考。如果我们把越来越多的认知活动都交给机器,最终可能会让个人、家庭乃至整个社会陷入一种困难的处境。

图片

AI 下一阶段的突破方向,不只是会思考,而是能行动、自主学习、创造价值(图片来源:Dr. Kevin Knight TGC 大会演讲 PPT)

因此,我想强调的是: AI 可以替我们完成很多工作,但我们不能因此放弃自己的思考能力。

最后,我想回到今天最初提出的那些问题。 AI 不仅仅是在思考,它正在学习,也正在学会采取行动。它能够阅读整个互联网,理解其中的信息,并将这些知识应用于越来越复杂的问题。

如果 AI 已经能够解决这些令人难以置信的难题,那么它显然也将创造巨大的商业价值,解决更多现实世界的问题。未来的突破还会不断到来。

在我的职业生涯中,我已经多次改变自己的研究方向,因为每当 AI 取得新的突破,就意味着我们需要重新思考自己正在研究的问题。我相信,未来很多人也会经历这样的变化。我们需要做好准备,持续适应 AI 带来的变化。

而我认为,最重要的一点是: 保持思考,不要把所有思考都交给 AI。

谢谢大家。

 

附:英文原文

Today, I want to talk about five questions about AI. How did we get here? How does that work? Is progress accelerating? What could drive the next leap, and how will AI change things? Are we going to be okay?

Ten years ago, I programmed a chess robot for my house. It plays on a real chessboard. And every morning still today, I make myself a cup of coffee first, and I play chess against the robot. I enjoy it.

Around this time, DeepMind developed two very famous programs for playing Go and chess: AlphaGo and AlphaChess.

They initially trained these systems using master games, but later developed much stronger systems that could play by themselves. Through their own experience, the machines learned which strategies were effective for winning.

You can imagine AlphaGo going up to its bedroom, playing Go by itself, and coming back downstairs a week later as the world champion Go player.

This doesn't bother me because I enjoy playing chess. It's not my field, and I still enjoy playing every morning.

The Evolution of Natural Language Processing

My field, for many years, was natural language processing. We worked on things like translating between Chinese and English, teaching machines to answer our questions, and many other language-related tasks.

In the early days, these were extremely difficult problems. My wife used to tell our friends in graduate school that she felt sorry for me because I was working on an impossible research problem.

What we were trying to build all those years, going back to 1950 and Shannon, were language models.

What exactly were we trying to build?

It's a rather unusual task. We might ask a machine to complete a sentence: "Mary saw toys in the window and wanted what, it or them?" The answer is probably "them." She probably wants the toys, not the window.

You and I understand this naturally, but the machine has to learn it.

By training machines to perform this task, they learn about the world, human society, and how everything works. Although it seems like a strange task, it allows us to solve problems such as translation and many others.

The First Breakthroughs in Language Models

In the 1980s, IBM achieved a major breakthrough. To predict the next word, they looked only at the previous two words.

We then used similar models to build the first generation of language systems.

The machine learned by reading text. Eventually, we could use these systems to communicate with people. This was a breakthrough at the time. Today, it has become commonplace.

Google later built a very large language model using this older technology.

We thought things would simply continue this way, with steady improvements in natural language processing. But in 2016, something very significant happened: our ability to predict the next word improved dramatically.

On this chart, lower values indicate better performance.

Recurrent neural networks and then transformers and then ChatGPT. And by the time we reach here as a research field, we kind of solved it. So the machine can understand language, can generate language, can translate language. And there's a lot. Yeah, so how does it work? Uh, one thing was just the scale of the training.

Now we're looking at internet-sized collections of examples. Someone types a bunch of words and they show us what the next word is and the machine can learn from those examples. No human in the loop. And then also for applications like chat, we used post-training. So here we have a human say something and the chat system replied, and another human says, that's a good reply. That's a bad reply. So this is a kind of human feedback.

Now I want to show the same chart here, with the same years: 1950, 2000, and 2030.

Instead of plotting the performance, accuracy, or perplexity of our systems, I want to plot our feelings as scientists working to develop natural language processing.

At the beginning, we were happy because we were having fun trying to build these systems. We were actually stuck and not making much progress, but we didn't know it.

Then the first wave of statistical methods came along. Some people were very happy because machines could now learn from text, and we could improve our systems and perhaps solve NLP. But others were unhappy because they loved language, not statistics.

There were some challenges, but things gradually got better. For many years, we have made steady progress.

Translation, for example, particularly Chinese-English translation, was slowly but consistently improving. During this period, we could even start companies focused on machine translation. There were some successes, but starting AI companies was generally difficult at that time.

Then came the revolution in neural networks, and everybody became very stressed out. We had to throw away all the techniques we had developed through years of experience and adopt these new methods. But that was okay, because things got better.

We were happy again. We had mastered the new techniques and could move forward.

By the time we reached GPT, the field of natural language processing research was in a state of disarray. Even today, people don't quite know how to feel about it.

We have more or less solved the research problem. Now, we are creating millions of possible applications.

Machine translation, for example, is what makes it possible for me to participate in this meeting today. All these things are now possible.

But for researchers, the arrival of new technology can be extremely disruptive. Everybody reacts to it in a different way.

My reaction was to change the research field and say, "Okay, we can say we won. Now let's try something more difficult."

I used to think words were very tricky and slippery. They're easy for humans, but every word has so many meanings, and putting them together is very difficult. But then it turned out that machines could do it. Words were easy.

I moved to mathematics and numbers. I thought numbers were very well-behaved, if you think about the multiplication table. But if you look at numbers for a few weeks, you'll find out that they are extremely chaotic. Even the best mathematicians in the world can't solve some of the simplest conjectures in mathematics.

I wanted to try moving into this area and use AI to accelerate progress in mathematics. I did that a few years ago, publishing journal articles and developing a deeper understanding of this new domain.

You may have noticed that this summer brought a huge revolution in AI for mathematics. AI can now solve one of the one-million-dollar Millennium Prize Problems. These are the problems that the world's greatest mathematicians have identified as the hardest and most valuable to solve.

Just two weeks ago, I think OpenAI released more than 300 papers that solved more than 300 important mathematical problems, introducing new proofs and new proof ideas.

This is shaking up the mathematics world, just as previous technology shook up the natural language processing world. Mathematicians are currently wondering what to do next.

How does this work?

Mathematics is interesting because there are so many math papers on the web, written by famous mathematicians, both living and deceased, over the years. These include the four mathematicians who became particularly famous this year for winning the highest award in mathematics, the Fields Medal.

A machine can now read through all this mathematical knowledge. It becomes a kind of superintelligence because it doesn't just know one field of mathematics; it knows something about many different areas.

This is important for large language models, but equally important is their ability to reason and verify their conclusions.

Just like AlphaGo, the machine knows when it has found the right proof. Its work can be verified.

A machine goes up to its bedroom, does mathematics for a week, and comes back as the world champion mathematician.

This is a huge breakthrough, and it happened just two weeks ago. There are many more developments coming in mathematics. The machine can also generate its own feedback.

The Acceleration of AI: How Long Can Machines Work Autonomously?

From your personal experience, you might think that the progress of AI is getting faster. I think so too.

From my experience with language and mathematics, every month brings something new that takes me by surprise. This is probably happening to you as well.

One way to understand this acceleration is to look at how long a machine can work independently on a problem without constantly interacting with you.

Perhaps it can work for 15 seconds, answer your question, and then you need to ask another one.

Or it might spend 10 minutes researching something on the web and return with a well-researched document. You then tell it what to do next.

It might work for an hour and build a small software system according to your specifications. Or it might spend 10 hours writing a large software system.

During those 10 hours, it works continuously without needing to communicate with you.

This kind of capability has only become possible this year. It wasn't possible last year.

At the extreme end, consider the very large mathematical problem that was solved over the summer. It represented approximately 100 years of machine thinking without anyone needing to interact with it.

Of course, the machine didn't actually spend 100 years. It involved 10,000 agents working autonomously for 88 hours. At the end, the system came back and said, "I solved this million-dollar problem."

The problems you encounter in your applications or businesses probably fall somewhere in the middle. Some problems may require more than 10 hours of thinking, while others can be solved much faster.

AI is definitely accelerating. Every month or every six months, machines become increasingly capable.

What's Next for AI?

I mentioned two fields I personally worked in that have been almost completely conquered by AI. What's next?

Some of the next areas include robotics, biomedical research, AI for work, entertainment, and AI running a business.

For all these fields, feedback is essential. How does AI improve, and what signals do it use to get better?

Robotics

For robotics, completing a task means success, and the robot should continue doing what works. If it fails, it needs to understand why and try something different.

AlphaGo can conduct experiments in its own mind. Robots can conduct similar experiments through simulation.

There is already a huge amount of work and many successful experiments involving robots operating in simulated environments.

Biomedical Research

Similar developments are happening in biomedical research.

There is already significant progress in drug discovery, CRISPR molecules, and other areas.

Just as a machine can read all the mathematics available online, it can also read all the biochemistry available online. It develops a kind of superintelligence, knowing more about biochemistry than any single individual.

But to establish whether its ideas are actually good, we need real laboratory experiments. For drugs intended for humans, we also need clinical trials.

There is still more work to be done in this area.

AI for Work

I'm not going to say much about AI for work and business tasks because we have excellent presentations coming up, including the next presentation by Shuhao.

Entertainment

AI is now creating entertainment.

It cannot yet completely write an entire novel or produce an entire movie that people genuinely want to watch.

The feedback here comes from people's preferences. Machines will only get better if we can tell them what people like and what they don't like.

That is also coming.

AI Running an Entire Business

Another development might be surprising, but it will be a real shock when a machine can not only work for a business but actually run an entire business.

In some fields, a machine-run company will compete with a human-run company. We'll see what happens.

The feedback for a machine running a business is clear: profit.

However, the machine might fail or run out of money before it has a chance to learn what it needs to know to build a successful business.

If it is given enough money, it can try many different things and eventually learn what works in that particular field.

The Changes AI Will Bring to Everyday Life

If you ride the bus, one day there will be a robot standing with you at the bus stop.

You'll get on the bus, and the robot will get on too. The first time, it will be a shock. But within a couple of weeks, you'll probably think, "Okay, that's just the way it is."

When AI designs a cure for a major disease, such as some form of cancer, I think it will also bring a major surprise and benefit to humanity. And I think this will happen soon.

Should We Be Optimistic or Worried About AI?

You can look at all these developments and think they're wonderful, or you can find them worrying.

The fields I've worked in have been completely transformed by AI. Other fields may experience the same transformation, and perhaps AI could even turn evil.

A lot of people are worried about this.

Among the people I personally know who work in AI research, there are both optimists and pessimists.

Some optimists have even started smoking cigarettes because they believe AI will cure lung cancer, so they don't need to worry about it.

Some pessimists have stopped saving for retirement because they think AI will kill them anyway. They see no reason to save money, so they spend everything they have.

I don't agree with either position. I'm not going to start smoking, and I'm still going to save money. I proudly recommend doing both.

But it's worth knowing that people who actually work in this field hold these very different views.

The One Thing I Worry About: We Must Keep Thinking

There is one thing I do worry about.

I think it's perfectly fine to offload multiplication to a calculator. You don't have to do multiplication yourself.

If you don't want to grow your own food, that's fine. I don't think anybody here probably grows their own food. Someone else or an automated system can grow it for you.

I like cooking, but if you don't, that's fine too. Someone can simply bring food to your house. You have more time to do other things.

The problem begins when we stop thinking.

In my daily life, I often find myself facing a problem and deciding whether to think about it and try to solve it myself, or simply ask AI.

I worry about this for myself, my family, and everyone else.

If we start offloading all our cognitive work to machines, we may find ourselves, as individuals and as a society, in a difficult situation.

Conclusion: AI Is Learning to Take Action

This is my final slide.

I've tried to address these questions, and I hope you can see that AI is not just about thinking. It is about learning and learning to take action.

AI is reading the entire internet and understanding it. If it can solve incredibly difficult problems, it seems clear to me that it will create business value and solve all kinds of problems.

The breakthroughs are coming.

Throughout my career, I've had to keep changing what I work on because AI keeps coming in and mastering the fields I've been working in. I think this will happen to many people.

We have to prepare ourselves to adapt as things change.

The best way to make sure we can do that is to keep thinking. 

Don't let AI do all your thinking for you.

Thank you!

END