In the summer of 2017, in an ordinary office at Google Brain, eight young people typed the last character of their paper, "Attention Is All You Need." At the time, almost no one realized that this paper would ignite a trillion-dollar AI revolution in the following years. In Lucas Kaiser's own words, " For us, it was just another ordinary day in the office. "
These eight authors subsequently left Google, becoming the most legendary "Transformer Eight" in Silicon Valley history.
Nearly ten years later, Lucas, now a senior scientist at OpenAI, uses Cursor to assist in his research every day. Interestingly, the first thing he asked the AI to do was to spend two days perfectly reproducing his academic paper that he could no longer run 15 years ago because he lost the source code.
The person who personally participated in the design of the world's largest "statistical machine" frankly admitted at his desk: " We haven't actually grasped the true mystery of 'learning' itself. "
In Lucas's view, the industry today has fallen into a strange, blind frenzy. The big model is like an extremely inefficient learner; it must devour trillions of words from the entire internet, exhausting all erroneous patterns of appearance, before it can passively "understand" a fundamental concept. This not only runs counter to human learning methods but is also causing the current Scaling Law to hit an inefficiency iceberg.
Last year, when Lucas spoke with Li Jianzhong, Senior Vice President of CSDN, he poured cold water on the idea, pointing out that "inference models are currently only equivalent to the very early stages of RNNs." A year later, as the rapid progress in pre-training large models has slowed significantly and the industry has fully shifted its focus to agent deployment and engineering exploration, time is precisely proving his "cold thinking" correct.
Here is a quick overview of the key points of this conversation:
Large language models can indeed learn a concept, but only if all other possibilities are exhausted. It generalizes, but it does so using a unique, "alien" way of thinking that we cannot fully comprehend.
The model is currently unable to accurately detect whether it is going further and further down a dead end.
As more and more people integrate these systems into their daily work, we will accumulate massive amounts of real-world human workflow data, spanning weeks or even months. Applying reinforcement learning to these complex workflows might bring unexpected surprises.
The AI industry experiences a technological tsunami every now and then; you must always bet on the trend that represents "tomorrow," rather than blindly coveting the prosperity of "today."
Once you leave the protection of the main lab and see the astronomical figures required to purchase graphics cards and their scarcity, you may face a harsh reality.
While OpenAI's core strength was being overshadowed by the phenomenal success of ChatGPT, Anthropic made an extremely wise strategic choice: to concentrate all its resources on the battlefield of "code," building a solid moat in the blind spot of the giants with absolute focus.
Our intuition tells us it should be smarter.
Host: I think there's no better opening topic than "generalization." This is arguably the focus of the entire industry right now. Last November, I heard you mention a core question: is inference alone sufficient to achieve generalization, or must we find entirely new paths? That was half a year ago, which is equivalent to several years in the rapidly evolving AI industry. How have your views on this issue changed during this time?
Lukasz : If you look at the current Transformers, which combine reasoning and agents and give them access to the system shell and various tools, their capabilities are truly astonishing. Compared to two years ago, not to mention before the birth of the Transformer, the progress we've made is simply incredible. If someone had told me before that simply taking a simple model of "predicting the next token," adding thought chains, reinforcement learning, and tools could unleash such power, I would never have believed it. Personally, I spend several hours every day communicating with the Cursor, and its performance is outstanding. When you discuss work-related problems with it, it not only understands perfectly but also helps you implement them directly. It's truly amazing.
However, on the other hand, we always feel that it's still quite different from humans, seemingly falling short of the limits we expect. Intuitively, we think it should be smarter, capable of generalization with less data, achieving greater leaps in thinking, and acquiring new concepts with minimal information.
I recently made an analogy: someone joked that Americans always do the right thing only after exhausting all the wrong options. Large language models are similar. They can indeed learn a concept, but only after exhausting all other possibilities. You need to feed it trillions of tokens, letting it explore all the apparent patterns. Only when these apparent patterns can no longer explain new things will it passively try to understand the underlying logic. But this is by no means how humans learn. We need very little data to master a concept, and sometimes we can even invent concepts out of thin air—though not perfectly, we do it. Therefore, we always feel that there must be some other mechanism hidden behind this, achieving much higher generalization efficiency and possessing a more fundamental and long-term understanding.
However, this is currently just an intuition. Every time we try to pinpoint this missing mechanism, it seems to vanish—or more accurately, the Transformer quickly catches up. During this time, both paths have grown. The Transformer has become increasingly powerful, but the calls for alternatives have also become more resolute.
Currently, many labs are exploring new architectures for the "post-Transformer era" and have already achieved some intriguing research results. There are indeed some interesting changes happening in the industry. As for who will ultimately prevail, I have no way of knowing at this point. I believe both sides have very solid arguments, and this process of negotiation will be extremely fascinating.
Host: This must be very appealing to our audience. In your recent NeurIPS talk, you also alluded to this "anomaly in the air"—it seems something is quietly pushing some emerging labs (neolabs) and researchers to set up their own labs and explore alternatives to the architectures dominated by mainstream companies. Where does this subtle feeling originate? Is it from some early experimental breakthroughs, or simply from researchers' instinctive intuition? Could you describe it more concretely for our audience?
Lukasz : I think it's largely intuition , and we have to stay alert because this atmosphere often ferments in the various parties and casual conversations of San Francisco. It can be self-reinforcing to some extent. But I believe there's also something very fundamental in it. In fact, Yann LeCun expressed similar views many years ago.
While our models are called "neural networks" and designed to mimic the human brain, they don't truly achieve that. Even with some similarities, there are fundamental differences. If you observe how humans learn and act, you'll find that we can accomplish far more complex things with far less data than existing models. As a kind of "learning machine," humans seem to possess some underlying, core ability that current models lack. Therefore, fundamentally, there must be some undiscovered scientific principle at play, not just a fleeting emotional atmosphere.
Of course, the opposing argument is also clear: these models consume trillions of tokens during training, and humans never encounter that much data in a lifetime. Therefore , we haven't actually optimized these models for "small data training." If you had the same amount of computing power but faced data limitations, you could easily fine-tune the Transformer to achieve far better performance. At this point, some might question: why go through all that trouble? We have plenty of data, and it's become a massive industry. But even if we tried to train with the same amount of data as humans—and then again, humans receive massive amounts of visual input, move and take actions in the real world—the dimension of this data is completely different from plain text, making a simple comparison difficult. This is why it's currently difficult to reach a definitive scientific conclusion.
But this intuition lingers: there are still incredibly valuable unexplored territories in the field of machine learning. The exciting thing is that once we find this missing piece, existing technologies could undergo a transformative leap. Of course, it might not; perhaps the gap will be insignificant in the face of massive amounts of data. Who knows? But as a researcher, it undoubtedly fascinates me, and I believe many of my colleagues share this sentiment.
The allure of Transformers is undeniable; their reasoning capabilities can even solve cutting-edge mathematical research problems. You've likely heard about the recent breakthroughs AI has made in mathematics. Having done mathematical research myself in my early years, this is simply amazing to me. I never imagined that in such a short time, a computer could engage in mathematical discussions with me at such a high level, like a true scholar. But it truly did, and it's incredible.
However, as a machine learning researcher, I reconsidered: we haven't truly grasped the mysteries of "learning" itself. The model does learn, that's undeniable, but the sheer volume of data and computing power it requires makes it seem like we're still one step away from the ultimate truth. Is this merely an intuition, or a fleeting atmosphere? It's more like a reality in some ways, but only time will tell.
Host : The research value behind exploring this mystery is undeniable. However, some might argue the opposite: so what if the model differs from humans? Since we possess massive amounts of data and this method works, that's enough. Of course, some fields do face data scarcity, such as new drug development. In these areas, efficient learning using limited data is crucial. But many core challenges in the real world don't actually involve such severe data bottlenecks.
Sometimes I feel like these two groups are talking past each other. It's understandable that people in mainstream labs might scoff at Yang Likun's views. After all, considering the massive amount of money currently pouring into the AI field, those data-unrestricted problems are indeed being tackled at an astonishing pace.
Lukasz : But soon, all the remaining bottlenecks will evolve into data-constrained problems, or rather, this trend is already evident. Especially to deliver satisfactory results in the physical world, you have to address this issue to some extent. Because the physical world, unlike the virtual world of text or the internet, cannot infinitely expand its data, the efficiency of data expansion is greatly reduced once you've trained on certain robotic hardware. The physical world is a huge challenge. Of course, people are currently trying to utilize simulated data and first-person video data, which are lower-cost alternatives.
I'm a huge Waymo fan. Whenever someone asks me, "Where are the promised self-driving cars?" I always joke that I ride in them every day, they're right here! But recently they canceled highway driving, simply because they couldn't handle certain construction zones. It feels like they've been struggling with these construction zone issues for years. I believe they've run millions of miles in their simulation system and accumulated a considerable amount of mileage in real-world conditions, yet the system still can't smoothly generalize its experience from "urban construction zones" to "highway construction zones." This seems illogical.
I don't know the specifics of the problem, but no teenager who gets a driver's license, or even any ordinary person, would face this kind of confusion. We humans have many shortcomings, but we would never be able to drive on a city construction zone but be completely lost on a highway construction zone. A construction zone is a construction zone; the principle remains the same.
Host: Do you think some of these challenges can be addressed through internal improvements to the Transformer? What do you hope to see in the next few years that will provide a clearer answer to this question?
Lukasz : The most exciting thing about machine learning research is its incredibly broad scope . You can never predict in advance whether you'll need to adjust the architecture, data, loss function, or optimization process. Each approach has its own rationale, and ultimately, you might need to make adjustments across various dimensions. The Transformer is indeed great, but its greatness is inseparable from the loss function of "predicting the next word"; you can also combine it with reinforcement learning (RL), but only if you introduce a chain of thought. These technological pieces only reveal their power when they are seamlessly pieced together.
If disruptive new paths emerge in the future, each step may need to be re-evaluated. However, it's also possible that some core components of the Transformer will continue to exist. For example, the attention mechanism will likely remain, albeit with the assistance of other new mechanisms.
My machine learning journey began with recurrent neural networks (RNNs), so the concept of "recurrence" has always held a special place in my mind. I adore its logical beauty. In a sense, the rise of reasoning capabilities has brought recurring mechanisms back to the forefront, because large models are essentially repeating the same set of weights every time a new token is generated . However, in practical applications, this reinforcement learning approach, accompanied by sparse loss, can handle so much computation and ultimately succeed. It's truly amazing.
Whenever we try to introduce looping mechanisms in other ways, we always seem to fall short at that final step. But this raises an old question: how much effort have we actually put into trying it? You or your audience may not be aware of looping models such as TRM and HRM. Despite their extremely small size, they have demonstrated remarkable performance in highly challenging benchmarks like Sudoku and even ARC-AGI. Although they are currently in toy testing, their performance is truly impressive. I believe that many cutting-edge " post-Transformer architectures " are fundamentally trying to integrate this looping mechanism with large language models. This is objectively very interesting.
The pure Transformer architecture doesn't perform ideally when dealing with these kinds of logical challenges, but by injecting some loop mechanisms, making some architectural tweaks, or slightly changing the loss function, its performance can improve dramatically. Even on a very small scale, you can achieve astonishing breakthroughs. Could this approach eventually generalize to the language level and bring us the capabilities we've always dreamed of? That's something to look forward to. Fortunately, several labs are currently working on this path.
In addition, this year we have witnessed an explosion in intelligent agents. For me, this is probably the biggest change in my daily work style that I have experienced in my twenty years of machine learning research.
This is equivalent to shortening a week's workday to one day.
Host: I wonder if you have tried to quantify how much AI has improved your work efficiency?
Lukasz : I can actually give a relatively accurate quantitative metric for this. I recently tried to reproduce some old papers that I've always been interested in on a personal computer, including several that I wrote myself but whose source code is now lost. I had previously tried to reproduce at least one of them manually, and I know that just getting the code running would take about three weeks. But with the help of Cursor, I completed it in just two days.
This is equivalent to reducing a week's work to a single day—a leap in efficiency of 5 to 10 times. Perhaps I could have gone even faster if I had worked a little harder back then, but this change undoubtedly reshaped my research rhythm, allowing me to try new ideas without any reservations. I can now even run three parallel experiments simultaneously and let them run autonomously, whereas before, when I was writing code by hand, I could only focus on one thing at a time. It not only greatly improves speed but also brings the ability for multi-threaded parallelism.
When working on private projects outside of production environments, I've largely stopped scrutinizing code line by line. A friend once asked me if this might have dulled my mental acuity. I've thought about this carefully, and the answer is quite the opposite. While I no longer need to micromanage every class name or every minute function, I'm well aware that the agent can easily veer off course. For example, once during an evaluation run, it encountered some auxiliary losses and inexplicably added a completely unrelated and utterly wrong auxiliary loss on its own.
Therefore, your brain must maintain complete and absolute control over exactly what the system is doing. What is its loss function? What is its underlying architecture? You must know these things. You just don't need to worry about the class names or the specific spelling details of a function. The sense of trust that allows the agent to precisely implement the ideas in your mind is amazing. Most of the time, when we check it, we find that it has been executed flawlessly.
Because your brain must be highly focused on the machine learning logic itself—how to design the loss function, how to adjust the batch size—I actually feel that I have a greater grasp of the underlying aspects of research projects now than when I was doing everything myself. In the past, when implementing an idea, before the code actually ran, I had to spend my energy on countless trivial debugging details, and then I had to jump back to the macro perspective. In this process, I often missed some ingenious designs.
Now, you're completely immersed in flow. You simply need to start from the essence of machine learning, think about what should happen , transmit instructions to the agent, verify its output, and everything will naturally fall into place. This not only saves time but also transforms the research itself into a tremendous enjoyment. I think this might be a kind of mild obsession that's become popular among researchers lately—we just can't stop.
Host: OpenAI has publicly stated that their goal is to bring AI to the level of a research assistant (intern) by November of this year. As a scholar who heavily uses Cursor in my daily research, how close do you think we are to this goal? What are your thoughts on this milestone?
Lukasz : Its performance is indeed very close to that of an intern, but you still have to keep a close eye on its output. As I mentioned earlier, it might arbitrarily add some loss functions that you didn't even ask for, purely because it sounds reasonable in its logic. I don't know if real interns would do that, maybe when they're very creative.
Sometimes I try to let it run on its own overnight, setting a macro-level goal, such as "improving the model to reduce perplexity." But this never works. It just starts making minor, insignificant tweaks with no research value. So, it's certainly not at the level of an independent researcher yet.
Host: What are the feasible paths to achieve a breakthrough in this direction?
Lukasz : This brings us back to our initial discussion. Actually, long before the Transformer was invented, I had been dedicated to researching "long context" and "memory mechanisms" in machine learning. Later, we brought long context into the Transformer era, achieving token lengths in the millions, which is an extremely impressive scale within the attention mechanism framework.
However, in the current era of intelligent agents, I've found that tools like grep and even ripgrep are the true solution for long contexts. We simply need to write massive amounts of content to files, give the agent the ability to use grep for retrieval, let it build index files, and operate like a small library.
As a researcher, if someone had told me five years ago that this was a way to handle long contexts, I would have scoffed, thinking it was nothing more than a stopgap measure. But in the field of machine learning, many great inventions initially seemed like expedients, such as Dropout. We shouldn't judge technology by its origins; if it's truly effective, we should embrace it. And this method has indeed proven remarkably effective.
You only need to add a small amount of reinforcement learning (RL), such as a compaction mechanism. If there's anything that made me firmly choose Cursor over the web version of Claude, it's Cursor's excellent context compression capabilities.
You can extend a conversation for a long time because it excels at extracting core information. Why does it do this so well? I don't think there's any profound secret; it's simply that the development team designed clever cue words and applied some reinforcement learning techniques to them. If you had told me a few years ago that the ultimate solution for long contexts was simply to use some reinforcement learning to teach it to use tools, search through files, and then extract the content concise enough to maintain the context, I would have said that it was just a band-aid and didn't touch on any scientific depth. But we don't judge a solution by its superficial elegance; we only look at its effectiveness, and this one is indeed extremely effective.
Therefore, regarding whether it can truly evolve into an independent researcher: some are pessimistic, believing it's impossible unless we develop some entirely new "post-Transformer" architecture that can understand larger macro concepts and have long-term goal orientation. This is indeed a reasonable argument. Currently, the existing approach seems capable of solving many problems.
Others argue that after a month of continuous conversations with the Cursor, you can guide it to review these conversations, identify meta-patterns, archive them, and then consider how to utilize them. Perhaps, if we collect similar interaction data from thousands of people and train reinforcement learning on it, the AI will begin to behave like a true scholar. To some extent, human scholars learn in this way: we observe the research of our predecessors, conduct repeated experiments, and summarize a set of most effective methodologies.
Host: Why isn't this method working now? I'm sure someone has already tried it.
Lukasz : I don't think people have put enough effort into this direction yet. Some people have written some prompts, which have had some effect, but that's about it. In my opinion, the real "Cursor era" started around last Christmas. Although Cursor existed before that, and we used it, and Claude also existed, everyone truly felt a kind of transformation during Christmas.
This seems to involve more than just a model upgrade; it also involves the meticulous refinement of the entire Harness and a series of post-training processes. And this has only been going on for about six months. If you step outside the AI industry circle in San Francisco, you'll find that many people haven't fully grasped this change, and they might even think that those of us who heavily rely on it are a bit overly enthusiastic.
This system only recently began to truly unleash its power. We can't even fully explain this leap in theory. It wasn't the result of a massive leap in pre-trained parameters, even though more powerful base models did emerge during that time. Back when we transitioned from RNNs to Transformers, it was easy to attribute that transformation to a complete overhaul of the underlying architecture. Now, while the importance of inference capabilities is undeniable, the transformation around Christmas last year remains somewhat enigmatic. Framework upgrades, post-training optimizations, and the timely emergence of new pre-trained models—multiple factors intertwined to bring about this astonishing leap, making it difficult to simply attribute it to a single driving force.
This is fraught with various cross-influences, as we are constantly optimizing every aspect of the system. However, precisely because its effects are so remarkable and crucial, in the face of fierce market competition, everyone is racing against time to commercialize and promote it across various application scenarios. This has resulted in insufficient time for in-depth theoretical analysis at the "meta-level." While some exploration has begun, conducting research at the meta-level means spending a week capturing a pattern and then attempting to implement it, requiring several weeks of system iteration.
Under existing reinforcement learning mechanisms, each iteration of the solution requires large-scale rollout testing. If a testing cycle lasts for several weeks, the training time for a single session will be extended to several months, which is completely impractical in engineering practice.
This perfectly illustrates the idea that human learning and research methods may offer profound insights for machine learning. Humans can spend years delving into a single research project, with very few attempts during that time. Some mathematicians spend twenty years tackling a single problem, which becomes their most brilliant achievement. They don't have two hundred twenty-year research cycles to repeatedly learn and experiment, yet they still manage to do it. What is the secret behind this? This is undoubtedly an extremely fascinating topic, highly relevant to the current development of AI. We haven't yet solved this mystery. However, as more and more people begin to integrate these systems into their daily work, we will accumulate massive amounts of real-world human workflow data, lasting for weeks or even months. Once someone applies reinforcement learning to these complex workflows, it may bring us unexpected surprises.
Host : This is a very insightful point. In the past, when we expanded the scale of pre-training or developed the first-generation inference model, the optimization path was extremely clear and logical—we knew very well which dimension to focus our computing power on. However, the rapid advancements of Cursor and Claude last Christmas seemed somewhat mysterious. If we cannot accurately pinpoint the true source of this change, it will be difficult to see which direction to focus our efforts on in order to continuously improve the core capabilities of the system.
Lukasz : That's true, it's definitely a bit baffling. Just because I don't know the specific tricks doesn't mean nobody in the industry knows. Perhaps some peers have strong confidence in the real breakthrough, but I think, at least for now, it's by no means an obvious consensus. Technical strength has actually been quietly accumulating for a long time, but after that transformation, many ideas that seemed like pipe dreams became reality overnight. This is clearly a benefit brought about by a clever scaling in the field of reinforcement learning.
Taste is difficult to define and break down using concrete language.
Host: A question that's currently receiving a lot of attention is this: we've already witnessed revolutionary changes in highly "verifiable" fields like code and mathematics. But regarding reinforcement learning, two core questions linger: First, how far can it go in "unverifiable" subjective domains? Second, can we achieve efficient generalization in entirely new domains without relying on massive amounts of proprietary data? In your opinion, how should we overcome the core challenges in "unverifiable domains"? Beyond code and mathematics, what do you think will be the next area to see a breakthrough?
Lukasz : Actually, we've made considerable progress in those "unverifiable" areas. Take Harvey in the legal field or some medical vertical applications as examples. While these tasks lack absolutely rigid verification standards, they still contain a large number of verifiable steps that can be cross-referenced. The results in these areas are quite encouraging. Furthermore, benchmark tests like GPQA also assess these comprehensive capabilities to some extent. There's a strong intrinsic motivation within the industry to expand into these areas.
In fact, simply labeling them as "unverifiable" may not be entirely objective. They certainly don't have the clear-cut rules of code or pure mathematics, but I think people have exaggerated the so-called "verifiability" of mathematics.
In the context of programming competitions, code is indeed very easy to verify. However, once you get into complex system front-end interactions, it becomes equally difficult to define using black-and-white standards. In mathematics, genuine academic proofs are rarely absolutely pure or easily automatically verifiable. You can certainly use formal tools like Lean, but most mathematical derivations produced by large language models haven't undergone rigorous formalization and are therefore not absolutely verifiable. It's a spectrum from easy to difficult, with verifiability gradually decreasing.
I once had a personal project—trying to translate English poetry into Polish, which sounds like an extremely subjective art. But when you let these large models act as reviewers, you find that they can actually capture very subtle details. They meticulously check rhyme, rhythm, and even cultural fit. It turns out that if we refer to past human review mechanisms, subjective art can also be quantified and verified to some extent.
However, this poetry translation project also revealed another truth to me: you can flawlessly verify all objective standards (rhyme, literal meaning, meter), yet the entire poem still feels soulless and lacks "taste." This is because taste is difficult to define and break down using concrete language. If it could be easily articulated, it would have already been formulaically validated. However, the inability to articulate it doesn't mean we can't perceive it. When you read it, a certain intuition in your brain stubbornly reminds you that something is still missing—a certain spark.
To some extent, this is precisely because of the current reinforcement learning paradigm that we've willingly fallen into a trap of our own making. Its operating logic is very simple: as long as there's a judge who can tell you what's good and what's bad, the model can iterate and become stronger in a targeted manner. This is the current growth mechanism of large models. Whenever I complain, "I think this line of text is poorly translated," someone always tells me, "Then you teach it what good taste is," and after a lot of corrections, the model can indeed eventually make up for this specific shortcoming. Just like image generation, it's difficult to define "beautiful" or "ugly," but you can significantly improve the overall aesthetic quality of the system-generated artwork by having thousands of people continuously click on more aesthetically pleasing images during training.
Therefore, verifiable boundaries are very vague and flexible. You can obtain sparse but extremely valuable data signals by collecting human preferences. Why do I find some writing lacking in aesthetics? This clearly stems from my life experience, accumulated knowledge, and my way of perceiving the world. And why can't the model produce this kind of inspiration? There are two possibilities: first, it hasn't experienced enough depth of experience; second, its logical mechanism for processing these experiences is flawed. I believe both reasons are involved. But even with the existing underlying processing logic, as long as you feed it richer real human experiences—for example, by collecting subjective feedback from thousands of people—its taste will improve dramatically.
Any vulnerability can be patched up by constantly patching it. But how wonderful it would be if we didn't have to go through all that trouble to patch vulnerabilities. Once you patch one vulnerability, it ceases to be an obstacle, and the next hidden leak is exposed as a new bottleneck. We seem to be trapped in this endless cycle. How perfect it would be if we had a core learning mechanism like the human brain, so that from the very beginning we wouldn't need to painstakingly patch up every loophole in the rules.
Host: Does this mean that, under the existing underlying architecture, any specific industry problem that people focus on can ultimately be solved? However, as you said, this might require a much larger amount of carefully selected proprietary data, and the whole process is far less natural than the more elegant learning mechanisms of the future. In your opinion, are there really certain types of problems or domains that current reinforcement learning methods simply cannot overcome?
Lukasz : Currently, there don't seem to be any insurmountable obstacles, but we must take commercial and economic costs into account. Under the existing technological path, if you want a model to demonstrate extremely impressive performance in a specific domain, you must first have a top-tier closed-source foundation model that is massive in scale and extremely expensive. Moreover, it is often a closed-source ivory tower, and you simply cannot access its underlying core weights.
While OpenAI offers some reinforcement learning fine-tuning APIs that I really like, and several other major companies are following suit, this incomplete control model still has limitations. Even with API-based fine-tuning, the process remains extremely challenging due to the incredibly high costs of data cleaning and computational resources required. This often necessitates a well-resourced company, long-term contracts, and a wealth of professional resources. If the problem itself has significant commercial value, then this path is certainly worthwhile; but wouldn't we much rather see a scenario where you simply chat with the model, and it single-handedly solves the problem flawlessly?
Host: Does the current foundational model demonstrate a leap in general, underlying capabilities? Let's imagine this scenario: we start by writing code, then conquer mathematics, and finally apply this mechanism to law and medicine, achieving breakthroughs one by one—even without immediately pursuing cross-domain generalization. Ideally, can we expect that after a series of explorations in reinforcement learning across different domains, similar to the pre-training phase, the large model will spontaneously develop cross-domain generalization capabilities in the reinforcement learning dimension?
Lukasz : That's true; signs of this spontaneous generalization are already emerging. For example, in the legal field, which isn't typically part of the standard reinforcement learning pipeline, when you talk to developers of vertical applications like Harvey, they find that a certain understanding either emerges spontaneously or, with only extremely weak guidance at the top level, the system instantly grasps and integrates the concepts. Generalization does exist, but its boundaries seem to be narrower than we expected. Sometimes, it can't even smoothly transfer between two subfields of mathematics.
For example, in the International Mathematical Olympiad (IMO), geometry problems were a formidable obstacle for models for a long time. While models could easily solve extremely difficult problems in other areas, when faced with geometry, people would always sigh: "It really lacks spatial reasoning." However, after encountering more geometry problems, models began to handle them with ease—they didn't encounter any entirely new physical or spatial data, but simply did more practice in geometric derivations.
The model's generalization curve exhibits a peculiar "sawtooth" shape. It may have made a significant leap forward in one dimension, but in another area that seems within reach, simply because its internal thought process doesn't perfectly align with our understanding, it completely stalls. It is generalizing, but it's using a unique, "alien" way of thinking that we can't fully comprehend, which is somewhat at odds with common human generalization principles. Perhaps as training data continues to accumulate, the blind spots it can cover will decrease. But I also fully understand why many decision-makers remain wary of such a system, hesitant to entrust it with significant responsibilities—because you can never predict in advance where its fatal blind spots might be hiding, and you must constantly be on guard against its errors.
As a machine learning scholar, I must remain highly vigilant and cautious when using these systems, as any slight oversight could lead me astray. From an academic research perspective, this rigorous training certainly keeps us sharp, but from a practical technical standpoint, it is undoubtedly a tremendous challenge, as we all hope it can become more rounded and gentle, rather than still being full of sharp edges.
The bottleneck of hardware architecture lagging behind scientific research ideas is rapidly dissolving.
Host: You just mentioned application-oriented companies that benefit from iterative model capabilities. Currently, the industry faces a crucial choice: as an application-oriented company, should we choose to establish deep collaborations with top-tier laboratories, sharing our evaluation systems and industry insights with them, or should we carefully protect our proprietary data and build our own models based on it, avoiding the loss of core assets to large companies? I'm very curious, what are your thoughts on the current application-layer ecosystem that relies on core foundational models?
Lukasz : The larger and more powerful your pre-trained pedestal model, the smoother those so-called "sharp edges" will be. Overall, this will make your application development much smoother. Whether doing reinforcement learning or fine-tuning on large models, a powerful pedestal will make subsequent work much more efficient. The enduring validity of this principle is truly admirable.
A year or two ago, a prevailing sentiment in the industry proclaimed, " Large models are dead; small models (SLMs) are the future ." Today, we've indeed witnessed a number of exceptionally good small models, such as the Gemma series with only a few byte (B) of parameters. Back in the GPT-3 era, everyone firmly believed that reliable zero-shot learning was impossible with fewer than 100 B parameters, but now even a 3 B model can demonstrate astonishing business capabilities. This is certainly exciting, but if you need to tackle extremely complex underlying challenges and want a model that can seamlessly integrate into your specific data and massive context, nothing can replace a truly massive supermodel. Of course, their training and inference costs are extremely high, and their deployment barriers deter many.
Host: For the general public outside of cutting-edge fields, a fact that isn't easily perceived is the extent to which new-generation hardware has actually liberated algorithms. For example, with the launch of NVIDIA's Blackwell chip, model capabilities have also taken a leap forward. It's difficult to distinguish whether this is due to powerful new hardware granting us computational possibilities that were previously unattainable, or simply a coincidence in the timeline. Do you think that the underlying architecture will naturally continue to become stronger with each leap in hardware computing performance?
Lukasz : Hardware performance upgrades are essentially reflected in two dimensions: floating-point operations per second (FLOPs) and memory access bandwidth. You must have sufficiently fast memory transfer efficiency to prevent massive computing power from being idle. This is an extremely direct and hard performance indicator.
I recently installed a 5090 graphics card on my personal computer. Its power is truly astonishing. A single 5090 can provide approximately 200 Teraflops of computing power (it can even reach 400 at certain mixed precision levels, but some are disabled). To put this in perspective, when we were writing the Transformer paper, the GPUs we used had a single-card computing power of only 9 Teraflops, and the entire machine had 8 cards. Considering the overall system overhead, the total computing power of the server was only about 70 to 80 Teraflops.
Now, the single graphics card in my ordinary small tower computer under my desk has the computing power equivalent to five high-performance servers back then. This means you could run all the experiments in the original Transformer paper in your study or kitchen using a single graphics card. And this was done in less than ten years. This is nothing short of a miracle in the history of science and technology. Today we run calculations at BF16 precision, but we can actually use even lower precision, especially after the introduction of Hybrid Expert Models (MoE), which allows you to cram more information into the inference stage.
The hardware requirements for running these models have been significantly lowered, which in turn broadens the scope of academic research. You can rely on massive amounts of high-speed graphics cards to train supermodels. Both NVIDIA's GPUs and Google's TPUs maintain an extremely rapid pace of iteration and are constantly improving in terms of parallelization.
But I think what's even more exciting is how much it unleashes the creativity of researchers. I remember when I first joined Google, the scientific community was still fiercely debating how many FLOPs it would take to simulate the entire human brain. Decades of calculations had ultimately pointed to between 1 and 100 petaflops. Back then, we all thought it would take decades of hardware evolution to reach that. Now, you can easily reach that threshold by simply buying a single GPU or renting a few machines in the cloud. Theoretically, you could now spend just a few hundred or a thousand dollars (instead of the previously prohibitive millions) to process the equivalent of a year's worth of data that the human brain would process in a single day.
If a new idea sparks in your mind about the human brain's learning mechanism, you can run and simulate years of learning in just a few days. In my opinion, this kind of empowerment is more disruptive than simply building a massive model. It can help you clear away all obstacles to implementation. In the past, when I was researching RNNs, I often felt restricted because they are highly serial and run extremely slowly in PyTorch. Although you could solve the speedup problem by writing a CUDA kernel by hand, the barrier to entry for writing a CUDA kernel is incredibly high. Now, you can let an agent write the CUDA kernel for you, allowing it to cross-correct with relatively slow unit tests, instantly turning what was once an insurmountable research barrier into a smooth path. Although their current kernel writing technology is not perfect, it has fully demonstrated feasibility. Before long, more advanced models will be able to do this: you only need to give it a single instruction to "maximize hardware performance," and it will provide you with perfect low-level code.
The bottleneck of "hardware architecture not keeping up with research ideas" is rapidly disappearing. Although the hardware architecture is still parallel, the boundaries you can explore are now vastly different because intelligent agents are continuously writing custom underlying kernels for you.
Host: There's a prevailing view in the industry that without the supercomputing power of top-tier laboratories, individuals or ordinary institutions can hardly conduct in-depth, practically significant academic research. You can certainly do some basic exploration, but truth ultimately needs to be tested in massive computing pools, a platform rarely available to ordinary people. Hearing your continued optimism about academia, amateur researchers, and single-card startups is truly encouraging. Do you think this kind of popularized scientific research can truly continue in the future?
Lukasz : It depends on my mood that day. On more optimistic days, I firmly believe this. Countless times, the history of science has shown us that truly beautiful ideas often arise from pure research, and there's no reason to interrupt future exploration. But at the same time, the mainstream technologies we currently possess have indeed demonstrated tremendous vitality, and it would be a huge mistake to abandon them and fail to further explore their potential. Fortunately, the current laboratory ecosystem in the industry is diverse enough.
Before transitioning to a lab, I was deeply involved in academia. The most captivating aspect of doing research in the ivory tower is that you can unleash your creativity and let your thoughts roam freely. You certainly can't compete head-on with large companies in terms of scale, but even on a relatively smaller scale—and this "small scale" is vastly different today—you can still produce cutting-edge and unconventional research. You should try ideas that completely break free from current mainstream frameworks, ideas that are highly inspired and aesthetically pleasing. That's the truest joy of doing research.
Of course, not all inspirations bear fruit; some methods perform amazingly well on a small scale, but quickly collapse once the parameter scale is increased. But if you have a mainstream 8-GPU server at hand, your research starting point is already far ahead of what it was five years ago. Five years ago, people could only make minor adjustments on toy datasets like MNIST; now, even on a single physical node, you are no longer doing minor tweaks.
Personally, I frequently use Andrej Karpathy's nanoGPT in my spare time. It's a GPT-2 level model that you can get running on a single-card machine in just a few hours. While current hardware is indeed quite expensive, older cards will eventually become more common as newer generations of GPUs are released. The possibilities for independent experimentation are vast, and even if not all methods work with supercomputing power, the intellectual stimulation and exhilaration of this process of exploration are unparalleled.
Host: Before we make a final decision, I'd like to ask you about another cutting-edge area—multimodal models. In a previous podcast, you mentioned that we haven't yet made truly disruptive progress in the field of multimodal models. Do you still hold this view? What are your thoughts on the current development landscape of the multimodal field?
Lukasz : People are clearly making breakthroughs. Perhaps the solution points in a direction similar to the Joint Embedded Prediction Architecture (JEPA). The multimodal learning path we currently take in Transformers and even diffusion models ultimately boils down to the extremely inefficient prediction of every single pixel.
But if you turn your attention to humans, our brains are constantly absorbing and processing massive amounts of information. Our neurons are relatively slow to respond—often requiring hundreds of milliseconds—yet our senses are receiving this deluge of data around the clock, from all angles. We successfully learn from this vast stream of data without needing to autoregressively predict every single pixel like large models. Human interactive learning mechanisms are highly parallel and unfold on an incredibly vast scale. Personally, I believe that existing models haven't truly grasped this core essence . This may require new breakthroughs in our research on underlying architectures.
New ideas are emerging in the industry, such as multi-stream Transformers. The standard Transformer performs attention calculations on the tokens mentioned earlier. However, you can design multiple parallel information streams to operate synchronously. While this is only a relatively straightforward, small architectural improvement, it possesses immense power.
When I'm working with Cursor, I might inadvertently forget something, verbally tell it, and then have to wait for it to run a Bash command in the terminal—a process that takes three minutes. It has to wait for my input to calibrate its orientation; this semi-automated state hardly qualifies as highly real-time, two-way interaction. We've introduced various temporary patchwork techniques to improve the experience, but in an ideal world, everything should be working collaboratively at all times. Human vision, hearing, and expression are bidirectional and integrated, and our models should possess the same real-time throughput characteristic. As major labs begin to focus on this dimension, this vision will likely become a reality.
However, the current impression is that while we are working on "multimodal" technologies, we lack a truly revolutionary architectural upgrade that can support "parallel absorption" at the underlying level. Under the existing operating mechanism, the Transformer simply cannot easily process a high-resolution image per millisecond because at the input end, it has to first tear the image into discrete small patches and then inefficiently stitch them together serially. This always feels awkward; we shouldn't cut images into these tiny patches. Sensory information should flow in unimpeded like a waterfall and be digested instantly as a whole. I believe we haven't yet achieved a truly fundamental breakthrough in this area, but I'm glad to see many colleagues diligently working on this direction.
You must always bet on trends that represent tomorrow.
Host: I'd love to talk to you about your time at OpenAI and your personal journey, because those years were truly eventful, and the company was constantly in the spotlight. During your tenure, what difficult yet crucial decisions ultimately reshaped and established the company's DNA?
Lukasz : I wasn't involved in the company's initial decisions, but during my tenure, the entire company faced a major choice: whether to completely transform and go all in on "Reasoning." At the time, the company's management and all of our R&D staff showed great courage, deciding to go down this path at all costs, elevating "Reasoning" to a strategic level of equal importance to "Pre-training," and dedicating ourselves to creating a new generation of models with a true inference core and bringing them to market.
In the early stages of development, these models, focused on logical reasoning, appeared somewhat impersonal in everyday conversations, struggling to imbue them with vivid human characteristics; their reaction speed was also quite slow, a bottleneck that persists even today. Many began to waver: Did we really have to go down this path? Wouldn't users actually prefer ordinary, fluent dialogue models? But OpenAI excels at making top-level decisions at crucial junctures, resolutely forging ahead on this thorny road, and in the process, developing an efficient management mechanism.
At that time, we had two completely different model product lines running concurrently, and merging them into a unified whole was an extremely challenging project. It took us a long time to complete this integration because all the underlying technologies were evolving rapidly. It was a very difficult decision, but if we hadn't gone all in and pushed forward, many of the powerful capabilities we enjoy today might never have existed. Even some of today's leading companies still face immense pressure to catch up in terms of the fine alignment quality of reinforcement learning, which perfectly illustrates the absolute leading advantage that "unwavering focus" can bring.
Since then, OpenAI and several other top labs in the industry have experienced exponential growth in size, as has Anthropic. Having worked at Google for a long time, I know how incredibly difficult it is for a large giant to make such a bold bet, given their many inescapable assets and cumbersome reporting processes. I sincerely hope that OpenAI and other emerging labs can continue this unconventional spirit. While existing technologies are undoubtedly excellent and can still help us expand our reach, if one day a glimmer of scientific light truly emerges in the "post-Transformer era," will these super labs have the courage to make drastic sacrifices to survive and embrace the future, or will they become increasingly conservative and hesitant?
When we were tackling the "inference" stage, we only saw a faint early glimmer of hope. We lacked massive amounts of comprehensive data to validate our findings; everyone was simply giving it their all with unwavering conviction. Now, although we haven't yet seen a next-generation general-purpose architecture with overwhelming advantages, once it begins to emerge, do we need a completely new laboratory to reignite this flame, or can we expect the existing leading players to continue taking on this unknown risk? At least, I believe OpenAI still possesses the gene for making risky decisions.
Host: That’s exactly why so many new forces, known as “Neolabs,” are emerging in the industry right now. As Jerry Tworek said when he chose to go it alone, breaking free from the shackles of traditional big companies allows them to be more focused and determined on the right direction they believe in.
Lukasz : That certainly has its merits. However, once you leave the protection of the main research labs and see the astronomical figures required to purchase graphics cards and their scarcity, you might face a harsh reality. Of course, graphics card computing power doesn't represent the entirety of scientific research. A diverse ecosystem built by small, dedicated teams and large, comprehensive super labs is very healthy for the entire industry.
In San Francisco, the epicenter of AI, you'll be enveloped in an atmosphere of intense competition and change, because the current technological advantages are far from exhausted. Countless sophisticated algorithms await implementation, data engineering needs upgrading, models with larger parameter sets need training, and a constant stream of new ideas are emerging. While they are not yet mature, various forces are relentlessly pushing them forward with massive amounts of funding and expertise.
However, once you leave San Francisco, you'll find that many outsiders view AI as if it hit a ceiling last year and will never progress further. This is undoubtedly a huge misjudgment. For me personally, the current batch of intelligent code agents is nothing short of a technological apocalypse. I'd call them the prototypes of Artificial General Intelligence (AGI)—of course, everyone can have their own unique definition of AGI. We might even gradually forget the term AGI in the not-too-distant future, just as we said goodbye to the Turing Test. Few scholars seriously debate whether AI has passed the Turing Test anymore, because technology has already easily crossed that finish line. The systems we rely on for our daily coding have undoubtedly demonstrated remarkable intelligence, which in itself is a groundbreaking leap forward.
Host: The competition in the current AI code generation field is extremely fierce. In your opinion, what are the key factors determining the success or failure of these code products, and how can they differentiate themselves? What are your expectations for the future frontiers of tools like Cursor and Claude? I believe this code market is large enough to accommodate multiple top players.
Lukasz : The more fundamental issue lies in how these tools can seamlessly transition to other work domains. Code is indeed crucial for us tech people, but the same underlying logic can empower countless industries. Whenever I strongly recommend Cursor to friends outside the tech field, it immediately requires you to link a GitHub repository, instantly deterring many ordinary users with no prior experience. Although its learning curve is being continuously lowered, because it's firmly anchored in the vertical category of "code development," people are completely unaware that it can also be a powerful tool for financial auditing. Compared to ChatGPT, which only requires typing a few words, Cursor definitely requires an adjustment period, especially when using Claude's in-depth development interface.
The most crucial point of contention lies in how we can bring this extreme efficiency to other ordinary industries. Anthropic is clearly putting Claude on this mission, striving to give these extremely robust underlying core capabilities a more approachable and accessible makeover.
Host : This cross-disciplinary capability absolutely exists. As a machine learning practitioner, I often see models effortlessly handling extremely complex Excel data and various office tasks. But frankly, truly mastering and guiding them still requires a certain level of expertise. While this is a skill that can be learned, most working professionals are extremely busy every day and may find it difficult to find ample free time to delve into it. Therefore, we must minimize the interaction barriers of our products. Furthermore, for fundamental considerations of data security and system stability, you can't allow them to run rampant without oversight. Building trust requires a long period of accumulation.
The question is, how do we convince the public to invest the initial time and effort to build this valuable trust? Looking back, why do you think Anthropic was able to get ahead and achieve significant success in the core area of code?
Lukasz : Anthropic made an extremely wise strategic choice at the time—to concentrate all its resources on the battlefield of "code." Meanwhile, OpenAI's core strength was being overshadowed by the phenomenal success of ChatGPT.
While ordinary conversational chat certainly holds boundless potential, Anthropic's core consideration in making this decision lies in its attempt to find a specific dimension where they can establish an absolute advantage and build a solid moat. This is purely a matter of strategic choice and top-level wisdom. The AI industry experiences a technological tsunami every now and then; you must always bet on the trend that represents "tomorrow," rather than blindly clinging to the prosperity of "today." ChatGPT was astonishing in 2025, but by 2026, the public's threshold will have been significantly raised, and by 2027, we will inevitably witness another disruptive reshuffle.
The shift in momentum was faster than anyone imagined. Once you launch a relentless, saturation attack on a small, overlooked niche, you can reap incredibly rich technical rewards. It's not that OpenAI neglected code—we've always been deeply involved in this area, which is why we were able to catch up so quickly later—but it wasn't the company's absolute core focus at the time. When a relatively lean startup team experiences a volcanic eruption of user volume in a very short period, knowing how to streamline and maintain absolute focus is the only way to avoid the collapse of the entire engineering architecture.
Host: You just mentioned a classic game: on the one hand, you need to squeeze every last drop of profit from the current opportunities, and on the other hand, you need to maintain an open perspective so that when new, unknown areas emerge, you can quickly free up your resources to invest heavily. OpenAI recently went through a well-known period of resource consolidation. We can see that they poured a lot of effort into delivering code and productivity tools, while proactively adjusting and cooling down their strategy for relatively exploratory frontiers like Sora. In your opinion, how can one skillfully balance the inherent tension between "thoroughly refining the current business" and "nurturing those small sparks that may ignite a future prairie fire"?
Lukasz : It all depends on a team's underlying DNA, organizational size, financial resources, and strategic vision. For example, Google strives to maintain a massive, multi-faceted research system. This is often used as a point of criticism, with the outside world arguing that while they've invented countless disruptive technologies (like the Transformer itself), they haven't been able to commercialize them with the most astute approach. However, this extremely strong academic foundation also gives them a significant asymmetric advantage: once an outsider breaks through a bottleneck in a certain area, Google, with its readily available and incredibly powerful research team, can close the gap and catch up in a very short time.
Host : Do you think they have really caught up completely? There is still a lot of discussion in the outside world that they are a step behind.
Lukasz : In terms of regular chat conversations, I think they've undoubtedly caught up or even surpassed each other. Their only weakness might be... I don't know if you've personally tested the latest generation of Gemini. I tried it out after Google I/O and was shocked to find that I could hardly distinguish whether I was using Cursor or facing Gemini itself. There were also plenty of related discussions on Twitter at the time, full of jokes and praises. It was indeed a big deal.
Meanwhile, I recently tried running some of my daily coding projects on the newly released Gemini 3.5 Flash, but its performance was slightly inferior, seemingly not yet having truly overcome the capability gap we experienced last Christmas. For my personal high-level work, it is indeed still somewhat inadequate, but I believe it is about to undergo a qualitative change.
When you choose a broad, multi-pronged research approach, you inadvertently build an excellent safety net, ensuring you won't fall behind in future competition. However, the cost is that you're unlikely to achieve the kind of absolute "first-mover advantage" in a particular emerging field, like Anthropic's success in code generation. It's incredibly inspiring to see such small, dedicated, and daring labs pioneering these breakthroughs, transforming isolated, unknown capabilities into solid technological pathways. This is how scientific progress should be.
OpenAI once boasted an extremely fervent and pure culture of "taking risks," but it has now grown into a behemoth. When your models support the daily productivity of hundreds of millions of users, or, like Google Search, safeguard queries for billions of people, any disruption or loss of balance in the system has significant implications. We certainly want speed to be as good as possible, but if the entire road is destroyed in the process of rapid advancement, the cost will be incalculable. Therefore, allowing these leading companies to maintain a sense of awe and stability towards the underlying infrastructure while sprinting forward may be the best arrangement for the entire ecosystem.
Be brave and test the wildest ideas in your mind.
Host: Many people are pondering the gap between open-source and closed-source large models. We seem to feel two completely opposite underlying forces pulling at each other: on the one hand, the barrier to model distillation seems to be getting lower and lower, with a large number of developers using the powerful output of closed-source models to feed back into and refine open-source models; on the other hand, the most advanced and massive super models have such high operating costs that even giants cannot directly provide services to the public, forcing them to internally perform multiple distillations on large models.
In your intuition, will the technological gap between open source and closed source be bridged in the next few years, or will it become an increasingly insurmountable barrier?
Lukasz : It's hard to generalize. At the current stage, model size still largely determines the upper limit of its capabilities. You can indeed slim down a model through distillation, but distilled versions rarely demonstrate the same capabilities as the full-size base version in truly extreme tasks. For example, I can always feel a visible performance gap between those touted as lightweight and cost-effective Flash versions and the traditional Pro or Sonnet versions—they are often distilled products made after compromising on limitations, while advanced players often have to patiently wait for the unreserved, top-of-the-line version. Even within the same model family, I can't remember the last time I had to rely on the "Mini" series. While they are cheap and easy to use, in my daily development, they will sooner or later make an extremely basic mistake, forcing me to spend a lot of time troubleshooting and debugging, and ultimately I will always switch back to the largest and most expensive flagship version.
You can indeed learn from and extract the wisdom of closed-source models, although major closed-source labs instinctively resist this direct distillation, they usually won't choose to eradicate them completely. If open-source models collapse and fall far behind in the future, it will be a disaster for the entire industry. However, I believe this worst-case scenario is unlikely to happen because there are currently enough powerful commercial groups and geek organizations wholeheartedly supporting the open-source ecosystem. I also understand the concerns of countries around the world: if you allow a country's municipal services and core hospital administrative systems to be highly dependent on the closed-source interfaces of a commercial company across the ocean, the consequences would be unimaginable if that company suddenly crashed or restricted access. This has created a rigid demand for "sovereign models."
Even if these sovereign models are slightly behind in absolute power, they don't need to utilize their highest computing power for routine administrative tasks. Therefore, the underlying driving force for a thriving open-source ecosystem is long-term. Meanwhile, closed-source labs, to ensure users are willing to pay for the most exclusive "golden privileges," will relentlessly pave the way at the forefront. This fascinating dynamic of mutual checks and balances, and spiraling upwards, will be the dominant theme of the industry for a long time to come. Of course, predicting the future of AI is always a high-risk endeavor.
Host: That's certainly a very insightful observation. In the past year, what aspect of the AI field has completely changed your perspective?
Lukasz : I honestly never imagined that we could develop such powerful code generation capabilities, comparable to human assistants, in such a short time. This completely overturned my previous prejudices. In the past, I rarely relied heavily on AI assistants in my daily work. People often asked me how to use ChatGPT, and I would reply that it was simply a matter of asking it a simple question every few days.
I never imagined I'd be spending every day in front of a computer, but now it's become my inseparable work partner. I also never thought I'd completely abandon traditional code editors; now I almost never type code myself anymore. Instead, I'm more like a commander sitting behind a screen, ordering assistants to make various code tweaks. This is a radical shift in my personal research paradigm.
Host: After being deeply involved in the development of these large models over the past few years, have your concerns about safety and the so-called "existential risk" increased or decreased year by year?
Lukasz : My core stance has always been quite stable. I firmly believe that we neither need to live in constant, unfounded anxiety, nor do we have any right to be blindly optimistic. Given the current level of programming development, what we should really be focusing on are the very real and immediate security threats that are already upon us —for example, the possibility that systems could be used by malicious actors to launch hacking attacks or cause paralyzing damage to critical infrastructure.
I still believe that this is the true main battleground that we urgently need to tackle together. This does not mean we can ignore the "existential risks"; on the contrary, a large number of specialized scholars are working to investigate this issue and to build a robust framework for the development of science and technology.
