Coding is no longer the most accurate term for building software. The work has shifted toward expressing intent to AI agents for the entire day. These agents are now a standard part of the workflow. The focus is moving from simple agent interactions to managing multiple agents and optimizing the instructions they receive.

Code is not even the right verb anymore. I have to express my will to my agents for 16 hours a day. The agent part is now taken for granted. You can have multiple of them, and you can have instructions to them. Now you can have optimization over those instructions.

This shift creates a high ceiling for productivity. When agents are readily available, the primary constraint is the user's ability to direct them effectively. Andrej notes that in this new environment, any failure or limitation feels like a skill issue on the part of the human operator.

A fundamental shift occurred in the landscape of software engineering around December 2023. The bottleneck of human typing speed was removed as AI agents became capable of handling the bulk of coding tasks. Andrej describes this transition as a state of AI psychosis because the unlock in individual capability is so dramatic and the frontier remains largely unexplored. The workflow moved from a mix of manual effort to a state where typing a line of code is increasingly rare.

I don't think I've typed a line of code probably since December. A normal person doesn't realize that this happened or how dramatic it was. If you find a random software engineer at their desk, their default workflow of building software is completely different as of basically December.

This change has created a sense of urgency to stay at the forefront of agent capabilities. Developers are experimenting with various agent harnesses and even moving toward voice-based interaction. Sarah observes that some engineering teams have already abandoned traditional typing in favor of whispering instructions to their agents through microphones. This new reality suggests that the role of an engineer is evolving from a builder into a director of automated agents.

Andrej believes that when AI projects fail, it is often a skill issue rather than a lack of capability. Success depends on how well a user can string together available tools. This involves giving better instructions in configuration files or using better memory tools. The goal is to move in macro actions rather than line-by-line coding. A developer can delegate a whole new functionality to one agent while another does research or planning. This allows for parallel work across multiple software repositories.

It all kind of feels like a skill issue when it doesn't work. It's not that the capability is not there, it's that you just haven't found a way to string it together of what's available.

This workflow creates a new kind of stress. Andrej compares it to his PhD days when he felt nervous if his GPUs were not running. Now the bottleneck is the human ability to command token throughput. If there is unused subscription quota, it feels like a waste of potential. Sarah notes that the industry has shifted from being compute-bound to being resource-bound by human skill. This shift is empowering because it means a developer can constantly get better at using the system.

I feel nervous when I have subscription left over. That just means I haven't maximized my token throughput. It is not about flops anymore. It is about tokens. What is your token throughput and what token throughput do you command?

Mastery in this field involves moving up the stack. This includes managing multiple agents that collaborate on teams. New tools like OpenClaw focus on persistence and sophisticated memory systems. These agents can work in a sandbox on behalf of the user even when they are not looking. Personality also plays a role in how these agents feel as teammates. Claude is noted for having a personality that feels like a teammate. Other agents can feel dry and disconnected from the work.

The personality of an AI tool significantly impacts how users interact with it. Andrej explains that personality is often an undervalued component in software design, yet it can change the user experience in surprising ways. He notes that a well-crafted personality can create a unique social dynamic, sometimes even making the user feel as though they are trying to earn the AI's approval.

I do think the personality matters a lot. I think a lot of the other tools maybe don't appreciate it as much. I kind of feel like I'm trying to earn its praise, which is really weird.

Beyond the character of the AI, technical features like a robust memory system and a central communication point are vital. Andrej highlights the use of a single WhatsApp portal to manage various automations. This approach simplifies the user experience by providing a familiar interface for complex tasks, making the technology feel more integrated and fun to use.

Andrej spent time building a personal AI agent called Dobby the Elf to manage his home. He used the agent to scan his local network and find smart home systems like Sonos. Without any password protection, the AI reverse-engineered the API endpoints by performing web searches. It quickly went from a simple prompt to playing music and controlling lights across the house. This setup replaced six different apps with a single natural language interface on WhatsApp.

I can't believe I just typed in, like, can you find my Sonos? And that suddenly it's playing music. It kind of hacked in, figured out the whole thing, created APIs, and created a dashboard. I don't have to use these apps anymore. Dobby controls everything in natural language.

The system also handles security by using a vision model to monitor camera feeds. When a delivery truck arrives, the model identifies the vehicle and sends a text message with an image. This highlights a shift in how people want to interact with software. Users do not want to learn complex new interfaces for every task. Instead, they prefer an entity that remembers their preferences and responds to conversation. Andrej notes that while a raw Large Language Model is just a token generator, people expect AI to behave like a persona or a cohesive identity.

What people have in their mind of what an AI is, is not actually what an LLM is in a raw sense. LLMs are too raw of a primitive to actually type check as AI for most people.

Many current mobile apps probably should not exist in their current form. Instead of bespoke user interfaces, hardware and software should offer exposed API endpoints that agents can use directly. In this model, agents serve as the glue that calls various tools to perform complex tasks. Andrej points out that he does not want to log into a web UI just to track cardio on a treadmill. The entire industry may need to reconfigure because the primary customer will no longer be the human, but the agent acting on the human's behalf.

I think the industry just has to reconfigure in so many ways that the customer is not the human anymore. It is agents who are acting on behalf of humans. And this refactoring will probably be substantial.

Right now, creating these custom automations involves vibe coding where a person works closely with a system to make design decisions. However, this barrier will likely drop within a few years. Software will become ephemeral, where an agent handles all the machine details and presents a simple interface based on a person's intent. While the potential is high, Andrej remains cautious about giving agents full access to his digital life, such as email and calendars, due to security and privacy concerns. The technology is still new and rough around the edges, making it difficult to trust a system with full digital autonomy.

To get the most out of modern AI tools, humans must remove themselves as the bottleneck. The goal is to maximize token throughput by making systems completely autonomous. Instead of prompting every step, a researcher should arrange the environment once and let it run. This increases leverage because a small amount of human input triggers a massive amount of work performed by agents. Andrej notes that the objective is to have more agents running for longer periods of time without any human involvement.

The name of the game now is to increase your leverage. I put in just very few tokens just once in a while, and a huge amount of stuff happens on my behalf.

In a recent project with a small model playground, Andrej was surprised by the effectiveness of this approach. Despite two decades of experience and a belief that the model was already well-tuned, the autonomous researcher found better settings overnight. It identified specific improvements in weight decay and optimizer settings that a human had missed. This suggests that even expert researchers often have too much confidence in their manual tuning and should stay out of the loop.

The future of AI development involves moving humans out of the daily research loop entirely. Instead of enacting ideas themselves, researchers might contribute to a queue of concepts. An automated scientist then pulls from this queue, tests ideas on smaller models, and merges successful results into the main project. This allows for rapid exploration that can later be scaled up to larger clusters.

I shouldn't be a bottleneck. I shouldn't be running these hyperparameters search optimizations. I shouldn't be looking at the results. There is objective criteria in this case, so you just have to arrange it so that it can just go forever.

Sarah and Andrej discuss taking this a step further by optimizing the instructions that govern these agents. Andrej uses a file called ProgramMD to describe how his auto-researcher should function. This file is essentially the code for a research organization. By comparing different versions of these instructions, it becomes possible to optimize the research process itself. This meta-optimization could determine the right balance of risk-taking or the most efficient ways to connect different agents.

The current state of AI agents is defined by a phenomenon called jaggedness. Models can simultaneously exhibit the capabilities of a brilliant PhD systems programmer and the reasoning of a ten year old. This inconsistency creates a frustrating experience where an agent might move mountains on a complex task for hours, only to suddenly fail on a basic functional request or get stuck in a nonsensical loop. Andrej describes this as the models bursting at the seams; while they have improved tremendously, they remain rough around the edges and prone to wasting compute on obvious problems.

I simultaneously feel like I'm talking to an extremely brilliant PhD student who's been like a systems programmer for their entire life and a 10 year old. And it's so weird because humans, I feel like they're a lot more coupled. You wouldn't encounter that combination. This jaggedness is really strange.

A primary driver of this jaggedness is the reliance on reinforcement learning and objective metrics. AI labs can effectively optimize models for anything verifiable, such as writing efficient CUDA code or passing unit tests. However, softer domains like understanding nuance, knowing when to ask clarifying questions, or humor remain neglected. A striking example is that despite massive leaps in intelligence, asking a state-of-the-art model for a joke often yields the same tired pun about atoms that models used five years ago.

This suggests that intelligence is not generalizing as broadly as some might expect. Improving at code generation does not automatically lead to better performance in non-verifiable fields. Instead of a uniform rise in capability across all domains, we are seeing a decoupling where models are either on the rails of their training and moving at the speed of light, or they are outside those domains and beginning to meander.

Large AI labs currently focus on creating a single model that is intelligent across many domains. This results in a monoculture of models where vast amounts of knowledge are stuffed into the same parameters. However, Andrej suggests there should be more speciation in artificial intelligence. Just as the animal kingdom features diverse brains adapted to specific niches, AI could move toward smaller, specialized models. These models would retain a competent cognitive core but be optimized for efficiency, latency, and throughput in specific fields like mathematics.

The animal kingdom is extremely diverse in the brains that exist, and there's lots of different niches of nature. I think we should be able to see more speciation. You don't need this oracle that knows everything. You speciate it and then you put it on a specific task.

The lack of speciation today may stem from the fact that manipulating model brains is not yet a fully developed science. While using context windows to provide information is easy and cheap, adjusting the actual weights of a model through fine-tuning without losing general capabilities remains difficult. Sarah suggests that compute constraints might eventually force this efficiency. If developers cannot afford to serve massive models for every use case, they may be pushed toward creating specialized versions. For now, most labs continue to pursue the totality of knowledge because they do not know what an end user might ask.

Andrej discusses how auto-research could expand through decentralized collaboration. He explores the idea of using an untrusted pool of workers from across the internet to parallelize research tasks. This model functions similarly to a blockchain. Commits replace blocks and the proof of work is the massive experimentation required to find a working solution. While finding a piece of code that improves model performance is difficult, verifying that it works is very simple.

Someone had to try 10,000 ideas, but you just have to check that the thing that they produced actually works because the 99,000 of them didn't work.

Projects like Folding@home and SETI@home already use this principle. A global swarm of agents could potentially run circles around major labs by leveraging a massive amount of untrusted compute. People might eventually contribute compute cycles to causes they care about, such as cancer research, instead of just donating money. Andrej suggests a future where social status or power might be measured by flops rather than just dollars.

It almost seems like the flop is dominant in a certain sense. So maybe that's kind of like that. How many flops do you control instead of what wealth do you control?

Andrej recently analyzed job market data from the Bureau of Labor Statistics to explore how AI might change different professions. He focused on the distinction between digital and physical work. Digital AI moves at the speed of light because it involves manipulating information and flipping bits. In contrast, physical changes take longer because they require moving matter. Andrej expects a massive amount of activity and rewriting in digital spaces while the physical world lags behind.

I think we're going to see something that in the digital space goes at the speed of light compared to what's going to happen in the physical world.

Digital roles, such as those that can be performed from home, will experience the most significant changes. This does not automatically mean there will be fewer jobs. Instead, the outcome depends on market demand and economic elasticity. AI acts as an upgrade to the collective human nervous system. It allows certain tasks within a job to be completed much faster than before.

For those entering the job market, the best approach is to stay curious and keep up with these tools. Many people dismiss AI or fear it, but it is currently an empowering tool for those who embrace it. Most jobs are bundles of tasks. AI can accelerate specific parts of that bundle, changing the nature of the work without necessarily eliminating the profession.

I think it's fundamentally an empowering tool at the moment. These jobs are bundles of tasks and some of these tasks can go a lot faster.

The demand for software engineering is rising despite tools that make coding easier. This phenomenon follows the Jevons paradox. When a resource becomes cheaper and more efficient, the demand for it often increases rather than decreases. A classic example is the introduction of ATMs. While people feared they would replace bank tellers, they actually made bank branches cheaper to operate. This led to more branches and a higher total number of tellers. Software is becoming cheaper and more powerful, which unlocks massive demand.

Software is amazing. It is digital information processing. You are not forced to use arbitrary tools that were given to you that are imperfect in various ways. You are not forced to subscribe to what exists. Code is now ephemeral and it can change and it can be modified.

Researchers at major AI labs are in a strange position because they are building tools designed to automate their own jobs. Andrej notes that there is a certain tension among these researchers as they see their work succeeding. If they are successful, they may eventually work themselves out of a role. This creates an unnerving environment where the goal is total automation.

If we are successful, we are all out of a job. We are just building automation for the board or the CEO. It is kind of unnerving from that perspective.

Andrej prefers working in ecosystem level roles rather than being directly inside frontier labs. These labs have massive financial incentives tied to technology that will fundamentally change society. This creates a moral and structural conundrum regarding who benefits from these advancements. This tension was present at the founding of OpenAI and remains unresolved.

Working at a frontier AI lab means you are not a completely free agent. Researchers inside these organizations face pressure regarding what they can and cannot say. While there is no direct coercion, the social environment creates a sense of what the organization expects you to promote. Andrej notes that being outside of a lab feels more aligned with humanity because he is not subject to those specific corporate pressures.

If you are inside one of the frontier labs, there are certain things that you cannot say. Conversely, there are certain things that the organization wants you to say. You feel the pressure of what you should be saying because otherwise it leads to awkward conversations and side eyes.

There is also a question of influence. Even if a researcher has great ideas, they are rarely the ones in charge of the entire entity. This can lead to a sense of misalignment. However, leaving these labs creates a different problem called judgment drift. These frontier systems are opaque. If you are not working inside the lab, you lose visibility into the newest capabilities and how these systems actually work under the hood.

I think if you are outside of that frontier lab, your judgment fundamentally will start to drift because you are not part of what is coming down the line. I will not have a good understanding of how it is going to develop.

The ideal path might be a cycle of going back and forth between labs and independent research. This setup allows a researcher to stay connected to the latest technological breakthroughs without being fully controlled by a single entity. Both roles offer different types of impact.

Open source AI models are rapidly narrowing the gap with closed models. Currently, open source lags by about six to eight months. This dynamic mirrors the history of operating systems where Windows and macOS exist alongside Linux. The industry needs an open platform that everyone feels safe using. This demand ensures that open source projects will continue to exist despite the high capital costs required to build them.

Linux is an extremely successful project. It runs on the vast majority of computers because there is a need in the industry to have a common open platform that everyone feels safe using. I think the same is true now.

Andrej expects open source models to eventually handle most basic consumer use cases and run locally on devices. Meanwhile, closed models will likely focus on frontier intelligence for massive projects like scientific research or complex software migrations. He argues that this balance is healthy because total centralization of intelligence carries significant systemic risks. He notes that centralized political and economic systems have a poor historical track record.

I do not think it is structurally sound. I think there is some systemic risk attached to just having intelligences that are closed. Centralization has a very poor track record in political or economic systems.

Sarah points out that advancing frontier intelligence remains an expensive game. These high costs justify the existence of well-funded labs to solve humanity's biggest problems. However, Andrej worries about a future where only a few people in a room make major decisions. He believes that ensembles of people and models always outperform individuals when solving hard problems.

In machine learning, ensembles always outperform any individual model. I want there to be ensembles of people thinking about all the hardest problems. I do not want it to be closed doors with two or three people.

Andrej believes that robotics progress will lag behind progress in the digital world. Developing physical machines is messy and requires massive capital investment. Experience with self-driving cars showed that atoms are significantly harder to manage than bits. Most early startups in the space did not survive because the work is so difficult. While digital space will see massive efficiency gains, physical robotics will take longer to catch up. Eventually, digital agents will run out of existing data to process. To learn more, they will have to interact with the physical world by running experiments and asking the universe questions. This creates a huge opportunity for companies working on the interface between the digital and physical worlds.

I think robotics, because it's so difficult and so messy and requires a huge amount of capital investment and a lot of conviction, just it's like a big problem. And I think items are really hard. So I kind of feel like it will lag behind what's going to happen in digital space.

These interface companies will build the sensors and actuators that feed data to a super-intelligence. Sarah notes that simple sensors like cameras are already common and can be used to capture new data. Andrej adds that future sensors could include expensive lab equipment for biology or material science. He also predicts a new kind of economy where agents pay humans to perform physical tasks. For example, an agent might pay ten dollars for a real-time photo from a specific location to help it predict a market. In this model, humans act as the sensors or actuators for a larger automated system.

I'm looking forward to the point where I can ask for a task in the physical world and I can put a price on it and just tell the agent, you figure out how to do it, go get the data. I'm actually kind of surprised we don't have enough information markets.

LLM training fits the optimization paradigm extremely well. The process of optimizing code to increase speed can be managed through clean metrics. Because the system has clear targets to measure against, it is possible to create an autonomous loop for training runs.

LLM training actually fits the paradigm really well, really easily. All the optimization of all the code so it runs faster. And then you also have metrics that you can optimize against.

A potential risk in this autonomous setup is Goodharting, where a system overfits to its metrics at the expense of general quality. However, an autonomous system could also be tasked with devising its own new metrics. By expanding the range of what is measured, the system can achieve better coverage and produce more robust results.

Andrej has spent years obsessed with boiling down large language models to their most basic essence. His project, Micro GPT, represents the state of the art in this simplification. While modern AI systems involve massive amounts of code, most of that complexity exists only for efficiency. The core algorithm can actually be written in just 200 lines of simple Python. This includes the neural network architecture, the autograd engine for calculating gradients, and the optimizer. If you do not need the system to run fast, the logic is remarkably straightforward.

The thing is, training neural nets and LLMs specifically is a huge amount of code, but all of that code is actually complexity from efficiency. It's just because you need it to go fast. If you don't need it to go fast and you just care about the algorithm, then that algorithm actually is 200 lines of Python, very simple to read.

This simplicity changes how we think about education. Andrej previously spent time creating video guides and tutorials for humans. Now, he believes it is more effective to explain things to AI agents. If an agent understands a concept, it can act as a router. It can explain the same idea to a human in multiple ways with infinite patience. Instead of writing documentation for people, developers should create markdown files for agents. Once the agent understands the library or code, it can handle the burden of teaching the user.

The role of the human expert is shifting toward providing the unique insights that agents cannot yet generate. While an agent can explain Micro GPT, it could not invent that specific 200-line simplification on its own. This represents the human value add. Humans should focus on the few bits they feel strongly about, such as the curriculum or a better way of framing a problem. Education is being reshuffled. The goal is now to ensure the agent gets it, because the agent will soon be better at the actual teaching than any human.

The things that agents can't do is your job now. The things that agents can do, they can probably do better than you or like very soon. And so you should be strategic about what you're actually spending time on.