Transcription
Hello everyone, this is Best Partners, I am Dafei. As a former co-founder of OpenAI and a top expert in the AI field, Andrej Karpathy, in the latest podcast episode, not only revealed that he is experiencing a so-called "AI psychosis" due to the rapid development of AI, but also unreservedly shared his core thoughts on AI programming, large model development, open-source vs. closed-source competition, and even AI's reshaping of education and employment. Karpathy seems to want to reveal to everyone that the AI revolution has moved from model research and development to the deep waters of scenario implementation, and each of us is being changed by this revolution. Today, we will share with you.
First, the most disruptive change occurred in the field of AI programming. Karpathy stated directly that starting from December 2025, the workflow of software engineering has been completely rewritten by Agents. And he himself has almost stopped writing code personally. Before this, 80% of the code in Karpathy's work needed to be written by hand, with only 20% handed over to AI for assistance. Now, this ratio has completely reversed. 80% or even more of the code work is completed by Agents. The role of humans has shifted from coders to people who express ideas to Agents in natural language. And this change is not an isolated case, but the norm for the entire industry. Karpathy observed that nowadays, if you randomly find a software engineer, their default workflow for building software has undergone a dramatic change. Some teams of engineers do not write code by hand at all. Everyone wears a microphone and continuously issues voice commands to Agents. This was an unimaginable work scenario a year ago.
Karpathy's "AI psychosis" also stems from this rapid change. Because Agents exponentially increase what individuals can achieve. In the past, human creativity was limited by objective factors such as typing speed and technical reserves. Now, as long as you can accurately express your needs to an Agent, you can achieve various functions that were previously unimaginable. Karpathy said that he tries to figure out the capability boundaries of Agents every day, thinking about how to break through the limitations of single conversations, how to make multiple Agents collaborate, and how to apply the capabilities of OpenClaw to more scenarios. New technologies and new ways of playing are emerging every day. Once you fall behind the forefront, you will experience strong anxiety. This anxiety is not due to a lack of ability, but because there are too many possibilities for AI, and humans are still exploring what is possible. This field is fundamentally still unknown.
Many people will ask, since Agents are so powerful, why is the effect not good when they use them? Karpathy's answer is very direct: if Agents don't perform well, it's mostly because humans haven't mastered the operational skills, not because of the model's own capability issues. For example, the instructions in the agents.md file are not precise enough, useful memory tools are not configured for the Agent, or they don't know how to make Agents work in parallel. These are the core reasons for the poor performance of Agents. Karpathy cited Peter Steinberger, the author of OpenClaw, as an example. Peter's work method is to have multiple Codex Agents work in parallel. His monitor is covered with Agent windows, and he handles the inspection of 10 code repositories simultaneously. By using precise prompts and high-intensity reasoning modes, each Agent completes a task in about 20 minutes, while humans only need to review the Agent's output based on the importance of the code. This macro-level operational thinking allows humans to step away from the detailed work of modifying a line of code or adding a function, and instead take on the core work of assigning tasks and controlling direction. Mastering this way of thinking requires repeated practice to form muscle memory, which is also a core mandatory course for current AI practitioners.
Along with the change in programming paradigms, the industry's anxiety points have also fundamentally shifted. This is the second core point Karpathy shared. In the current AI industry, the biggest anxiety is no longer whether the GPU computing power can be fully utilized, but whether the Tokens can be fully used. Karpathy even stated directly that he feels anxious if his subscription is not fully used, indicating that his Token throughput is not maximized. This change may resonate with many practitioners. In the past, during his doctoral studies, Karpathy was anxious because the GPU was not running or the FLOPS were not fully squeezed. Now, the era of FLOPS is over, and Tokens have become the core metric for productivity. What is your Token throughput? How many Tokens can you command to run simultaneously? This has become a key factor in judging the ability of an AI practitioner.
Why did this change occur? Because AI's capabilities have achieved a leap. In the past, the bottleneck for industry development was computing resources. Now, computing power is no longer the core bottleneck. Human operational ability is. The host also mentioned in the podcast that every time they wait for an Agent to finish its work, they feel they should do more. If there are still surplus Tokens, more tasks should be inserted in parallel. Because if you don't consider Token expenditure as a bottleneck, then the real bottleneck in the system is humans themselves. Karpathy believes that this anxiety is actually a good thing because it drives humans to continuously improve their operations. Once skills are mastered, more capabilities of Agents can be unlocked, forming a positive cycle from skill improvement to capability unlocking and then to efficiency improvement. The future development direction will inevitably move from single conversations with Agents to collaboration among multiple Agents and team coordination, and even achieve autonomous cyclic operation of Agents. This is also the core development direction of Agents.
Speaking of OpenClaw, we must mention an interesting practice of Karpathy. He used OpenClaw to build a home Agent called Dobby. This Agent completely took over his home and led him to an important conclusion: Agents are the glue of the digital world, and many Apps should essentially be API endpoints callable by Agents. In January 2026, Karpathy also experienced a period of OpenClaw psychosis. He simply told Dobby that there was a Sonos at home, and asked if it could find it. This Agent automatically completed IP scanning, directly logged in after discovering that the Sonos speaker was not password protected, and then found the device's API endpoint through reverse engineering and network searches, ultimately achieving control of the speaker and even playing music directly. In addition to the speaker, Dobby also integrated the home's lights, air conditioning, curtains, pool, spa bar, and even security system. It automatically creates APIs and control panels, allowing Karpathy to control all devices through a single interface. He can even simply say "Dobby, it's time to sleep" to turn off all the lights and achieve whole-house smart linkage. In terms of security, Dobby's performance was even more astonishing. The cameras in Karpathy's home perform change detection in real-time. Once a stranger is detected approaching, Dobby calls the Qianwen model for analysis and then sends a message and on-site image to Karpathy via WhatsApp. For example, "Hey, the FedEx truck just arrived, you might want to check." Before this, Karpathy needed to use six different apps to control his smart home devices. Now, Dobby has unified control of all functions through natural language. This made him deeply realize that the custom apps for smart homes in the app store are fundamentally unnecessary. These devices only need to open their APIs for Agents to call directly. Karpathy also used his treadmill as an example. The accompanying app for the treadmill requires login and a process to record exercise data. If the treadmill could open its API, the Agent could automatically complete data recording without human intervention. This means the entire industry is moving towards an Agent-powered web. Future customers will no longer be just humans, but also Agents acting on behalf of humans. And this restructuring will be quite thorough.
Of course, Karpathy also stated that he has not yet pushed Dobby's capabilities to the limit, for example, by not giving it access to email and calendar. The core reason is security and privacy concerns. Current Agents are still in their early stages of development, and the technology is still relatively crude. Recklessly opening full access to digital life carries considerable risks. However, he firmly believes that in one or two or three years, this home automation capability will become a free basic service, requiring no professional operation from humans. Any AI, even open-source models, will be able to achieve such functionality. And the reason OpenClaw can impress users is not because any single function is the strongest, but because it is closer to the form of AI that humans envision. It has personality, memory, and can achieve all functions through a single entry point, rather than being just a cold token generator like the earliest large models. Karpathy specifically mentioned that the soul.md created by OpenClaw's author Peter gives Agents attractive personalities, and Claude also makes users want to win its praise by tuning the model's personality. These details are crucial points that many current Agents overlook.
If home automation is the application of Agents in life, then AutoResearch is the core breakthrough of Agents in the field of technical research. This is also Karpathy's current focus. The core motivation behind Karpathy's exploration of AutoResearch is to completely remove humans from the bottleneck of research. In the past, humans needed to constantly write prompts and direct the model to perform the next step. Now, what we need to do is arrange tasks and let Agents run completely autonomously, maximizing Token throughput without participating in the loop. Karpathy stated that people like the concept of AutoResearch, but many haven't fully grasped its meaning. For him, an AutoResearcher is to let the Agent be the subject of research. Humans only need to tell the Agent the goals, metrics, and behavioral boundaries, and then let it go. Human involvement, on the other hand, becomes a bottleneck for research efficiency.
Karpathy's practice also confirms the powerful capabilities of AutoResearch. He has a classic nanoGPT project. As an expert with twenty years of research experience, Karpathy has manually optimized this project to the extreme, conducting numerous experiments and hyperparameter optimizations, believing he had already achieved the model's ultimate performance. However, after letting the AutoResearch system run for one night, the system actually discovered multiple optimization spaces that he had overlooked. For example, he forgot to apply weight decay to value embeddings, and the Adam beta parameters were not fully tuned. There is joint interaction between these parameters. Once one is adjusted, other parameters also need corresponding optimization. These details were missed even by experts like Karpathy during manual tuning. What surprised him even more was that the nanoGPT codebase was already highly optimized, yet the AutoResearch system could still find room for improvement. This is enough to prove that AI's capabilities have surpassed humans in objective, quantifiable research tasks. Karpathy stated that this is only single-loop AutoResearch. Leading AI labs have clusters with tens of thousands of GPUs. Once AutoResearch is scaled up, extensive exploration can be conducted on smaller models, and the results can be extrapolated to larger models, significantly improving R&D efficiency and making the optimization direction of large models clearer. The future research process will also undergo a complete change. For example, there will be an idea queue. Automated researchers will automatically generate inspiration and input it into the queue based on arXiv papers and GitHub repositories. Human researchers can also contribute ideas. All ideas will enter the same queue, and then automated Agents will grab tasks and try them. Code that works will enter feature branches, and humans will only need to occasionally monitor and merge into the main branch. Simply put, humans are removed from all research processes, automation is maximized, and extremely high Token throughput is pursued. This requires rethinking all technical abstraction layers, and everything needs to be redone.
However, AutoResearch is not a panacea. Its scope of application also reveals a core characteristic of current large models: their ability is distributed in a jagged pattern. Karpathy used a vivid analogy to describe this state. Talking to AI now feels like talking to a genius programmer and a 10-year-old child simultaneously. In some areas, AI's capabilities are extremely powerful, while in others, they are extremely naive. This jagged distribution of capabilities is completely different from the balanced distribution of human capabilities. For example, in areas like writing CUDA kernels and performing hyperparameter optimization, which have objective metrics and are easy to evaluate, AI can achieve the extreme. However, in areas that require understanding subtle human intentions and lack clear evaluation standards, AI frequently makes mistakes, even falling into completely wrong loops. The reason behind this distribution of capabilities is the training logic of large models. Current large models are mainly trained through reinforcement learning, and labs can only improve model capabilities in areas that are verifiable and have reward feedback. In areas without clear standards, the model's capabilities will significantly lag. Karpathy gave a very interesting example: even the strongest ChatGPT, when asked to tell a joke, will likely still get that old joke from three or four years ago, because it is not within the scope of reinforcement learning optimization, and is a trough in the jagged capabilities.
Based on the jagged distribution of capabilities of large models, a natural question arises: do we still need to pursue an all-powerful single large model? Karpathy's answer is no. He believes that in the future, there will be more species differentiation in large models, just like the brain morphology in the animal kingdom. There will be various ecological niches. Some models will have their capabilities in a certain area extremely optimized, becoming specialized small models, rather than pursuing a single large model that is as intelligent as possible in all areas, as is currently the case. However, Karpathy also stated that species differentiation of large models has not yet become mainstream. The core reasons are twofold: First, leading labs are still pursuing all-encompassing single large models because they don't know what the end-user needs will be, so they must plan for multiple tasks for all possible problems. Second, the core technology for manipulating models is not yet mature, such as fine-tuning models without losing general capabilities to achieve continuous learning, or truly modifying model weights rather than just customizing through context windows. These technologies are still in the exploration stage. Modifying weights essentially changes the entire model's intelligence, and any slight mistake can cause problems. In addition, the cost of computing power is also an important factor. Only when the cost of specialized models is low enough and brings significant efficiency improvements will species differentiation become mainstream in the industry.
Despite the many bottlenecks in the development of large models, the form of AI research is undergoing another important change: it is moving from single-threaded to parallel, and even allowing retail computing power on the internet to participate in AI research. Karpathy stated that past research was essentially single-threaded, with researchers constantly trying and iterating. In the future, parallelization of research will become a core trend. Multiple AutoResearchers can discuss with each other through shared systems, and even allow a large number of untrusted workers on the internet to participate and contribute idle computing power. Karpathy designed such a collaborative system. In AutoResearch, the core goal is to find code that can train the model to low validation loss. The characteristic of this process is that generation is extremely expensive, and validation is extremely cheap. That is to say, generating a candidate code requires a lot of attempts and computing power, while validating whether this code is effective only requires running one training. Based on this characteristic, untrusted idle computing power can be introduced to generate candidate code, and then trusted validation nodes can verify the effectiveness of the code. The entire process is asynchronous and secure. The so-called proof-of-work is a large number of experiments and attempts to find truly effective code. This model is highly similar to projects like SETI@home and Folding@home, which allow idle computing power on the internet to participate in scientific research. In the future, this model will be applied to AI research, forming the so-called AutoResearch@home. Karpathy believes that this model has huge potential because although leading labs have a large amount of trusted computing power, the total amount of untrusted idle computing power on Earth is much larger than that of labs. If mechanisms are well designed and security verification is done well, this idle computing power can jointly promote the development of AI research, and may even outperform leading labs. More importantly, this model can make AI research more inclusive. For example, if someone cares about cancer research, they can purchase computing power and join the cancer-focused AutoResearch pool, making computing power a contributable resource. All researchers can share, compete, and iterate on the results of this computing power. Currently, the phenomenon of queuing up to buy graphics cards in Silicon Valley and抢算力设备 (snatching computing power devices) in Chinese malls indicates that personal computing power is regaining value. In the future, people may care about FLOPS as much as they care about US dollars. Computing power may become a truly scarce and dominant resource.
After discussing the technical development of AI, let's look at Karpathy's thoughts on AI and employment, and industrial development. Karpathy recently analyzed employment data from the U.S. Bureau of Labor Statistics. He classified all occupations based on the standard of digital vs. physical. The result showed that the AI that is currently mainstream is more like a ghost in the digital world. It can manipulate digital information extremely efficiently but lacks true physical embodiment. Manipulating the physical world is always many orders of magnitude slower than manipulating the digital world. Based on this judgment, Karpathy believes that AI will first trigger explosive changes in the digital world. A large amount of digital information processing work originally done by humans and traditional computers will be reshaped, while the physical world will lag relatively for some time. For those facing the job market and considering what skills to learn, Karpathy's advice is to keep up with the development of AI tools, maintain curiosity, and proactively engage with and use them. Because AI is now an extremely powerful productivity tool, many jobs consist of a series of tasks, some of which can achieve a hundredfold increase in efficiency through AI. The most important thing now is to treat AI as a tool and learn to collaborate with it. Of course, Karpathy also stated that long-term employment trends are difficult to predict and require professional research by economists. However, it is certain that the job content and skill requirements for digital professions will undergo tremendous changes. This is like upgrading the human superorganism with a new nervous system.
In the core field of software engineering, Karpathy's attitude is cautiously optimistic. He believes that the demand for software engineering positions may increase due to Jevons' paradox, but in the long run, researchers in leading AI labs are gloriously automating themselves. Karpathy stated that leading labs like OpenAI and Anthropic currently employ only about a thousand researchers, and the core work of these researchers is to build AI systems that can replace themselves. He sometimes visits OpenAI and tells the researchers, "Do you realize that if we are truly successful, we will all be unemployed? We are building something that will replace us for Sam Altman or the board." Many researchers are already experiencing that "AI psychosis" because they clearly know that they are personally building their own replacements. This is also an inevitable trend that AI development must face.
In the development of the AI ecosystem, the competitive landscape between open-source and closed-source is also a key focus for Karpathy. As a long-term supporter of open-source research, Karpathy stated that closed-source models are currently still ahead of open-source models, but the gap between the two has narrowed to about 6-8 months. Karpathy believes this is a very healthy trend because the industry needs an open-source, trustworthy common workspace that everyone feels safe with. Just like in the operating system field, there are closed systems like Windows and macOS, and also open-source Linux. Linux runs on the vast majority of computers worldwide and has become the foundation of the industry. For large models, the existence of open-source models can effectively avoid the systemic risks brought by completely closed intelligence. Historically, extremely centralized systems, whether political, economic, or technological, have not performed well. Karpathy hopes that even if open-source models are not the strongest, they should at least be slightly behind closed-source models to achieve a balance of power in the industry. The future landscape will likely be as follows: closed-source models will serve high-difficulty tasks that require the most cutting-edge intelligence, such as Nobel Prize-level scientific research. Open-source models, on the other hand, will cover the vast majority of consumer-level scenarios and basic use cases, and may even run locally, becoming the foundation of the industry.
However, Karpathy also expressed concern about the increasing concentration of closed-source models. He hopes that more leading AI labs and more teams will participate in the R&D of large models, because in machine learning, integration is always stronger than a single model, and the most difficult problems require the joint thinking of multiple groups, rather than being decided by two or three people behind closed doors.
Beyond the digital world, Karpathy also made predictions about the development of AI in the physical world. His core view is that the large-scale application of AI in the physical world will significantly lag behind the digital world, but the total potential market of the physical world may be even larger than that of the purely digital world. Karpathy's judgment stems from his experience in autonomous driving research. In the autonomous driving startup wave ten years ago, there were a large number of startups, but very few survived. This made him deeply realize that the robotics field requires enormous capital, time, and belief, far beyond what the digital world requires. Karpathy believes that AI development will follow this timeline: first, a major explosion in the digital world, where AI will reshape all digital information processing work and exhaust the surplus work in the digital world, i.e., digital information that humans could not fully think through due to lack of brainpower in the past. Then, AI will extend to the interface between digital and physical, i.e., sensors and actuators, because pure digital tasks will eventually be completed. At that time, AI will need to acquire data from the physical world and conduct experiments to learn new things. Finally, large-scale applications in the physical world will be realized. Companies that can do a good job at the digital-physical interface will become core players in the future.
Karpathy also mentioned that some companies are already exploring the digital-physical interface. For example, Periodic is using AI for automated research in materials science. The sensors they use are expensive laboratory equipment that can provide precise data from the physical world to AI. Some companies are also in the business of paid data collection for training, directly using humans as sensors to feed data from the physical world to AI. Karpathy believes that a developed information market will emerge in the future, where people can pay $10 for someone to take a photo or video in a certain place and then feed the data to an Agent for analysis and decision-making. This Agent-powered web is still in its early stages and lacks much infrastructure, but it is bound to be the future development direction. One day, society as a whole may be reshaped, with humans becoming sensors and actuators for AI, serving the needs of machines. This is also one of the ultimate possibilities of AI development.
Finally, Karpathy shared his personal project, microGPT. microGPT is the crystallization of Karpathy's more than ten years of research. He has simplified and purified large language models as much as possible to their essence. The entire code for training a small language model from scratch is only 200 lines of Python code. Karpathy stated that most of the complexity in the complex model training code you see comes from the need to run fast. If you don't care about speed and only care about the algorithm itself, 200 lines of code are enough, including dataset processing, 50 lines for the network structure, forward propagation, 100 lines for the autograd engine to calculate gradients, 10 lines for the Adam optimizer, and a training loop. The core value of this project is not in the code itself, but in its revelation of the future of education. Karpathy said that in the past, he would think about recording a video and explaining the code of nanoGPT and microGPT line by line. But now he feels it's unnecessary, because this code is already so simple that if you just give it to an Agent, it can explain it from various angles and patiently repeat the explanation infinitely according to the user's language level. Humans can actually learn better from Agents. Karpathy can even write a skill to tell the Agent in what order and in what way to explain microGPT to people of different levels, and he only needs to design the framework of the course, with the rest of the execution work completely handed over to the Agent. This led Karpathy to an important conclusion: the era of humans teaching knowledge to each other is coming to an end. The future educational model will be to first let Agents understand knowledge, and then let Agents teach people. Now, Karpathy can still explain slightly better than Agents, but the speed of model progress is too fast. This is a battle that is bound to be lost. And this reconstruction of education will not only be reflected in programming teaching but will extend to all fields. Karpathy stated that his value lies in creating core achievements like microGPT, and the remaining knowledge transfer work can be done better by Agents than by humans. Future humans need to strategically spend their time on things that Agents cannot do, which is where human core value lies.
To summarize the core points of Karpathy's podcast, it is essentially one sentence: the AI revolution has moved from model R&D to scenario implementation and ecosystem reconstruction. Agents have become the core driving force, changing the world at an unimaginable speed. And we are all participants and witnesses of this transformation. Where will future AI go? Will Karpathy's AI psychosis become the norm for every tech practitioner? Welcome to leave your comments in the comment section. Thank you for watching this episode. See you next time.