15 Observations About Artificial Intelligence

Thoughts about DeepSeek, ChatGPT, and the future of humanity.

One of the things I most love about Substack is the ability to share my thoughts with you in a direct, unfiltered way. And one of the new formats I’ve discovered for myself in the process is this series of Observations, in which I try to make sense of a big topic—whether it be a place like France or China, or a subject like AI—by sharing my most interesting thoughts about it without needing to pretend that they add up to one clean, coherent whole.

Thanks for being along for the ride. If you want to support this project, and get access to all of my weekly essays—as well as every episode of my podcast, The Good Fight—please become a paying subscriber today.

- Yascha

(Photo by Lionel Bonaventure/AFP via Getty Images.)

1. Many of my friends still seem to think that AI is “all hype.” On social media, viral posts claim that the progress of AI has stalled; that the technology can at best create shoddier copies of work done by humans; and that it’s only useful to people interested in plagiarizing their school assignments. But all of that is cope. If you haven’t yet figured out how to use apps like ChatGPT, Claude and DeepSeek in your own life, you’re missing out on a miraculous—and extremely useful—technology.

2. Here are some of the things for which I have used AI over the course of the last few weeks: Find a good dim sum spot. Translate my articles and podcast transcripts into French and German. Teach me poker by presenting different scenarios and analyzing the pros and cons of different moves. Practice spoken Chinese with me. Come up with a recipe based on leftover ingredients in my fridge. Proofread an essay draft. Compare and contrast national conservatism and neo-integralism in preparation for a class I’m teaching. Set me Chinese grammar exercises. Find the name of a writer about whom I only remember vague details. Help draft an email in formal French. Analyze the respective advantages of stock and real estate investments. Locate the footnote in which John Rawls discusses abortion. Tell me about a striking building based on a photograph. Design custom-tailored cartoon images to send to a friend as a joke. Explain the importance of a psychologist whose work I have never read. Troubleshoot a computer problem. Plan out activities for a day trip.

3. Are most of these things that, given a lot of time and money, human experts could have done just as well, and perhaps even a little better? Sure! Does the range of things these apps can do for you instantly and at (virtually) no cost make them the most miraculous personal assistants ever invented? Without a doubt. If you’re not using them, you’re losing out.

4. One of the deflationary arguments I often hear is that AI is not artificial intelligence at all; rather, skeptics claim, it is a kind of clever conjuring machine that simply recognizes patterns in the data on which it was trained, allowing it to predict what word (or number or pixel or line of code) is most likely to follow upon the heels of the previous one. In a literal sense, this is true: that is roughly speaking how “generative pre-trained transformers” used by companies like OpenAI work. To say that this does not constitute true intelligence is nevertheless wrong, just as it would be strange to say that you’re not really intelligent because the smart argument you just made turns out to be a product of neurons firing inside your brain. If a device is capable of doing the kinds of tasks, from writing poems to solving hard math problems, which we have traditionally taken to be constitutive of intelligence, then pointing out how it does so is a strange way to deny that it does.

   ---

   Would you (or someone you know) like to read my articles in German or French? Please subscribe to my sister Substacks!

   ---

5. More profoundly, humans too seem to be extremely proficient pattern-recognizing machines. Neuroscience remains in its infancy. We don’t truly understand how human intelligence, let alone consciousness, works. And there are likely to be some big differences between humans and machines, ones that make it naive to think that our brains are simply wet GPTs; one neuroscientist, for example, told me that we should think of AI as a third kind of intelligence on Earth, alongside vertebrate and cephalopod intelligence. But once we understand the nature of our own brains better, the scientific explanation for the physical bases of our intelligence may turn out to be no more flattering than the scientific explanation for the physical bases of artificial intelligence—and there is a chance that these two sets of explanations may even turn out to be more similar than many now believe.

6. One of the deflationary arguments about AI is that its progress has supposedly slowed. I do not in any way pretend to be a technical expert on AI, and certainly don’t have a crystal ball that would allow me to predict its future development. And it does appear to be the case that optimistic assumptions about the speed with which further increases in the scale of data and compute fed into training AI models would lead to further improvements have been frustrated over the course of the last year. But I’m skeptical of these arguments for two reasons. First, in the domains I can assess, progress has been extremely fast. When OpenAI first made its GPT-3.5 model publicly available, a student could have earned a B or perhaps B+ by submitting an AI-written essay to a college class in the humanities or social sciences; on the latest model of Claude or DeepSeek, it could (in part but not only because of grade inflation) get an A- or even an A in the great majority of humanities and social science classes at top American universities. Second, it seems plausible that there may be something to arguments that the current approach of major AI labs in the United States, which puts a lot of emphasis on scaling the data on which these models are trained and throwing ever more computing power at the task, may eventually bring diminishing returns. But that is the way progress tends to work, especially when technologies are protean: when the current approach ceases to work, these AI labs will innovate in other ways.

7. The release of DeepSeek—a Chinese rival to ChatGPT that’s taken the market by storm since last week—is one illustration of that point. It took me a few days to recognize its importance; after all, its performance, while impressive, does not seem to exceed the most sophisticated models that were already publicly accessible. So what’s the big deal? The release of DeepSeek is significant for at least three reasons. First, it demonstrates that China and the United States are genuine competitors for global leadership in AI. Second, it is a genuine gift to humanity. Like Meta’s earlier Llama, the model has been released in open source form, meaning that it will forever be available for any human being to use at little to no cost. And third, DeepSeek is notable not just for its performance but also for its innovative development approach. For example, DeepSeek introduced key technical advancements, such as optimizing how the model processes information. While most language models analyze text in chunks (like words or parts of words), DeepSeek enhances this by focusing on larger, more meaningful units—like phrases or sentences—allowing it to better understand context and relationships within the text. Additionally, DeepSeek employs techniques that activate only the most relevant parts of the model for a given task, effectively “delegating” work to the parts of the system best suited to handle it.¹ Taken together, these well-documented changes allowed DeepSeek to reduce the cost, speed and energy use of the model, both in development and everyday use. This demonstrates how much space for creativity remains in the development of these models and (in my admittedly layman opinion) should increase our estimation of how likely future progress remains.²

8. The immediate economic impact of DeepSeek’s release has been a sharp fall in tech stocks. There are some rational reasons for this. Clearly, the most heavily impacted American tech companies are not as far ahead of their Chinese competition as their extremely steep valuations seemed to assume. And if the computing power needed to train AI models really turns out to be much lower than expected, it is imaginable (if far from certain) that manufacturers of advanced chips like Nvidia will enjoy lower profits. But this does not at all mean that the release of DeepSeek is bad for the economy as a whole. For it is a boon to humanity if the development of AI turns out to be cheaper than expected. And all kinds of developers now have access to state-of-the-art AI technology on the cheap, likely spurring further innovations over the coming years. Every major development has cons as well as pros; but the fact that a groundbreaking technology has just been democratized should be cause for celebration.

9. Here, for example, is one area in which applying AI, even at its current levels of performance, can still lead to huge improvements in the real world. The ways in which AI can help you learn a language are incredible. You can have a conversation in the language you’re practicing with the world’s cheapest and most patient tutor. You can practice reading a text, or having a written exchange, based on your most niche interests. You can get grammar lessons custom-tailored to your personal strengths and weaknesses. And you can get instant feedback on your pronunciation, including detailed tips on how to improve. But for now, all of this requires that you design your own course, instructing your AI language teacher what to do at each step. There is clearly vast room for improvement here, with somebody building an app that systematically asks about your personal preferences and then integrates all of this functionality into a streamlined experience. The obstacles to achieving such a learning interface are much more rooted in challenges of design, pedagogy and user experience than they are in the need for further progress in the power of the underlying model. Somebody will almost certainly fix these problems, creating the most powerful language learning tool ever invented—but for now, such an app does not exist.

   ---

   Make sure that you don’t miss any episodes of my podcast! Sign up for ad-free access to all of its episodes by becoming a paying subscriber and adding the private feed to The Good Fight now.

   ---

10. In some corners of the internet, DeepSeek has also been celebrated for another reason. Unlike Western AI models like ChatGPT and Claude, these commentators claim, DeepSeek does not engage in censorship on myriad topics. I am at some level sympathetic to the underlying concerns. It’s clear that companies like OpenAI and especially Google seem to have put a heavy political hand on the scales during the “fine-tuning” phase of development, in which human moderators reward the model for answers they like and punish it for ones they don’t. When social media first became big, the political mainstream was far too slow to recognize how socially corrosive it is for Silicon Valley—or government bureaucrats pressuring Silicon Valley—to be arbiters of what citizens can and can’t say. Now, the mainstream is even slower to recognize that political control over AI should be even more offensive to anyone who has genuine liberal values. If you believe in human freedom, you should be passionately opposed to a world in which governments collude with powerful billionaires to decide what facts or even opinions the world’s most powerful research tools and teaching assistants can express.

11. It is nevertheless naive to ask both ChatGPT and DeepSeek a set of questions that are controversial in Western political discourse and then draw the inference that the former censors more than the latter. Each country’s political discourse contains its own taboos. Some of these are obvious to outsiders; others wouldn’t even occur to anybody who doesn’t know the country well. Having asked both models similar, politically sensitive questions over the last couple of days, my preliminary impression is twofold. First, both models are dishearteningly conventional in their views, in part because conventional views—even ones I believe to be wrong—probably represent the majority of the texts used for training them; when I asked the models to list some highly contentious or taboo topics in American politics, and assess whether they are true, they both ended up sounding a lot like the least imaginative writers employed by The Guardian or The New Republic. And second, just as it is easy to find the areas on which ChatGPT refuses to offer an answer, it is not at all hard to find topics on which DeepSeek refuses to engage. After asking about America, I instructed both models to run a similar analysis of taboos for China. ChatGPT was happy to volunteer a number of areas in which it believes Chinese taboos to be false or dangerous. (Unsurprisingly, these judgments were largely in line with conventional American views on China.) DeepSeek, by contrast, told me that this topic “is beyond my current scope. Let’s talk about something else!”

12. Europe is just nowhere on any of this. The few AI companies located on the continent, like France’s Mistral, are now far behind both America and China. Because European entrepreneurs find it hard to raise the necessary funds and the best European talent is now concentrated in the Bay Area, it is hard to see how the continent can catch up anytime soon. (It is, for example, telling that SAP, the largest German tech company, was founded over fifty years ago, and remains economically viable mostly because large corporations that adopted its software decades ago find it hard to transition to the much better business enterprise solutions offered by American competitors like SalesForce.) And so just about every European politician and intellectual I have spoken to about this topic resorts to the copiest of copes in a realm replete with cope: The idea that Europe could remain relevant by becoming the world’s leader in AI regulation. This ambition is sad, reminiscent of a schoolchild’s dreams of growing up to be a hall monitor. Humiliatingly, it is not even realistic: How can Europe hope to influence the global development of AI if the continent remains a complete sideshow in its development?

13. When ChatGPT was first released, I wondered whether AI may one day be remembered as the “third humiliation of humanity.” According to Sigmund Freud, the first great humiliation came when Copernicus argued that the universe does not revolve around the Earth. The second great humiliation came when Charles Darwin discovered that we evolved from apes.³ This makes me think that AI may be the author of our third great humiliation. The most advanced models are already better than any human at chess, certain video games, and some scientific tasks like predicting how proteins will fold. They are better than most humans, including professionals, at a huge range of tasks from medical diagnostics to translation. And they have made genuine inroads into creative fields, like writing poems, composing songs, or creating images. If next-generation models should one day outperform humans at these tasks, this wouldn’t just pose an existential crisis for writers such as me; it would also pose a critical question for humanity as a whole: What remains of humanity’s self-perception if machines start to outperform us at some of the creative tasks that we used to consider most quintessentially human?⁴

14. Humans throughout history have proven incapable of resisting the temptation of chronocentrism, the false belief that their particular era had outsized historical importance. There is every possibility that predictions according to which AI may prove to be the end of humanity, or at least the end of humanity’s reign as the most powerful species on earth, will turn out to be similarly chronocentric. But while I do not think that breathless predictions according to which the end of humanity is imminent are likely to prove correct, it would be a mistake to rule such a possibility out altogether. Three things are now clearly true: The best AI models are starting to rival human intelligence. The field of mechanical robotics, from humanoid robots to self-driving cars, is rapidly advancing. And the problem of alignment—of making sure that artificial intelligences behave as instructed, ensuring that they could never harm humanity—remains far from solved. p(doom) may not be very high; but nor is it extremely low.

15. Our thinking about ethics, both in the public and in the “professional” realm, remains ill-equipped for the questions AI raises. Here’s one of the many questions I’ve been trying to wrap my head around, without much success: It is clear that we humans have reason to resist AI taking over. Some forms of partiality in moral thinking are inevitable, and perhaps even desirable. As humans we must defend the future of human existence. But what kind of tragedy would be entailed by our species being supplanted by hyper-intelligent machines? Would the new world be one bereft of meaning, populated by creatures that lack consciousness and therefore joy and perception and moral worth? Or would such hyper-intelligent AI systems constitute a form of life that, in its own, vastly different way, has a value of its own, one in whose creation we humans can take some small modicum of pride even as we rue our own demise?

1. Since I am not a technical expert, and to illustrate the usefulness of AI, I asked DeepSeek with help in checking the accuracy and suggesting a more precise formulation of the last four sentences. My original draft of them was: “And third, DeepSeek is notable for the form of its development as much as for the level of its performance. DeepSeek changed key technical aspects such as instructing the model to operate phrase-by-phrase rather than word-by-word and finding a way for the most relevant parts of the model to operate while leaving others dormant when answering questions—essentially, delegating tasks to those parts of the system best equipped to deal with them.”

2. This does not, however, mean that we should take the claim that the model only cost $6 million to develop on trust. For one, this is likely to reflect the cost of electricity use and other key inputs, not the salaries of employees and other costs of doing business. For another, it is at this point impossible to verify whether the true costs of these technical inputs may itself have been understated for effect.

3. Sigmund Freud also identified a third great humiliation: Freud’s own discovery of the unconscious, revealing that we are not in control of our own minds. But this always struck me as a rather grandiose piece of self-promotion, one that exaggerates the supposedly world-historical importance of his own work.

4. I plan to write about this observation more fully in the coming weeks.