Mar 20, 2026 · 1h 6m · no-priors
Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of No Priors, AI researcher Andrej Karpathy and host Sarah Guo explore the rapid transition toward autonomous coding agents, recursive machine learning through AutoResearch, decentralized compute swarms, and the future integration of AI across physical robotics and education.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 20% of the talking time here. How this is scored →
The hosts as informed peer 5.7 Guest teaching 2.9 Guest disagreement 1.1 The hosts pushing back 1.2
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 45:00 Pushback on frontier lab alignment
Andrej rejects the premise of being confined to frontier labs, arguing that institutional employment subjects researchers to corporate speech pressures and financial misalignment.
Hardest push from the hosts ▶ 44:35 Posing Noam's frontier lab questionSarah directly presses Andrej on why he chooses to work independently rather than leveraging the massive compute and peer clusters available at frontier labs.
Biggest teaching moment ▶ 18:29 AutoResearch exposes human hyperparameter blind spotsAndrej details how letting an autonomous loop run overnight uncovered hyperparameter interactions that had eluded his twenty years of manual tuning.
The host holds their own ▶ 22:08 Proposing recursive meta-optimization contestSarah outlines a technical mechanism to optimize research configurations by feeding benchmarking data back into LLMs, which Andrej validates as the correct architectural path.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Preview: The Loopy Era of AI and Agentic Coding | 5 | 2 | 1 | 1 | Sarah opens by sharing observations of Karpathy's intense coding sessions and framing the recent jump in agent capabilities. Andrej describes the psychological shift from manual coding to managing autonomous agent swarms. | |
| Parallelizing Agents and Operating via Macro Actions | 6 | 3 | 1 | 1 | Sarah draws an insightful comparison between previous compute constraints and today's human-bandwidth bottlenecks. Andrej expands on macro-action workflows and token throughput optimization. | |
| Agent Persistence, Memory Systems, and Personality Calibration | 5 | 3 | 1 | 1 | Sarah probes whether memory systems or tool access drive user resonance in persistent agent architectures. Andrej highlights the importance of personality calibration and memory retention. | |
| Building 'Dobby': An Agentic Smart Home Operating System | 4 | 3 | 1 | 0 | Andrej details building Dobby, his local smart home agent that reverse-engineered device APIs automatically. Sarah reacts with surprise at how few prompts were required. | |
| Rethinking Software Architecture for Agents Instead of Humans | 6 | 2 | 1 | 1 | Sarah raises architectural questions about whether users truly want distinct application UIs rather than raw APIs. Andrej argues that application layers will collapse into ephemeral software orchestrated by agents. | |
| AutoResearch and Eliminating Humans from the ML Loop | 6 | 4 | 1 | 1 | Andrej explains the architecture behind AutoResearch and how automated tuning outperformed his twenty years of manual ML tuning intuition. Sarah connects this to extrapolation along scaling laws. | |
| Meta-Optimization of Research Organizations with Code | 7 | 2 | 0 | 1 | Sarah proposes a recursive contest structure to generate optimal program.md files via LLM data feedback. Andrej strongly endorses the proposal as code-level meta-optimization of research organizations. | |
| Jagged Intelligence and Reinforcement Learning Constraints | 6 | 4 | 2 | 2 | Sarah questions whether capabilities in verifiable domains like coding generalize to broad societal intelligence. Andrej explains reinforcement learning jaggedness using the persistent repetition of the atom joke. | |
| The Dilemma of AI Monoculture versus Specialized Models | 7 | 3 | 1 | 2 | Sarah asks whether hardware and serving constraints will force unbundling monolithic models into specialized domain experts. Andrej notes that while biological speciation is logical, current science relies heavily on context windows rather than parameter modification. | |
| 'AutoResearch at Home' and Decentralized Compute Swarms | 5 | 3 | 1 | 1 | Andrej outlines a decentralized 'AutoResearch at Home' architecture using commit verification protocols akin to blockchain proof of work. Sarah links this concept to the rising consumer interest in local compute. | |
| AI's Impact on the Job Market and Software Demand | 6 | 3 | 1 | 1 | Sarah examines rising demand for engineering roles despite automation. Andrej invokes Jevons paradox, explaining that lower software production costs will dramatically expand aggregate demand. | |
| Trade-Offs of Independent Research versus Frontier Labs | 6 | 3 | 2 | 3 | Sarah confronts Andrej with Noam Brown's query about why he does not conduct research inside a frontier lab. Andrej defends independence, citing institutional alignment pressures and opacity trade-offs. | |
| The Critical Balance Between Open Source and Frontier Labs | 6 | 2 | 1 | 1 | Sarah highlights that expensive frontier models remain essential for humanity's hardest problems. Andrej argues that having open-source models trailing closed frontier models by six to eight months creates a healthy systemic power balance. | |
| Bridging Digital Intelligence to Physical Actuation and Sensing | 6 | 3 | 1 | 2 | Sarah and Andrej discuss the lag between digital bits and physical atoms in robotics. Andrej foresees an emerging market of sensory and actuation interfaces designed to feed autonomous digital intelligence. | |
| MicroGPT and Teaching Agents Rather Than Humans | 5 | 4 | 1 | 0 | Andrej presents microGPT, a 200-line LLM implementation, explaining how education is shifting from human-directed guides to authoring instructions for agents to teach humans. Sarah synthesizes the implications. |
Statements from this episode (31)
Karpathy: Has not typed a line of code since December
“I don't think I've typed like a line of code. Probably since December, basically which is like an extremely large change.”
Guo: Conviction portfolio team exclusively dictates to AI agents without typing
“We have a we have a team that we work with at Conviction that their setup is everybody is like, you know, none of the engineers write code by hand and they, they're all microphoned and they just like whisper to their agents all the time.”
Karpathy: Coding agent failures are user skill issues, not capability limits
“Even if they don't work, I think to a large extent, you feel like it's a skill issue. 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. Like I just don't, I didn't give good enough inst…”
Karpathy: Parallel AI agents elevate software engineering to macro actions
“It's just like you can move in much larger macro actions. It's not just like, here's a line of code, here's a new function. It's like, here's a new functionality. And delegate it to agent one. Here's a new functionality that's not going to interfere with the o…”
Karpathy: Engineering productivity shifted from GPU FLOPS to agent token throughput
“You would feel nervous when your GPUs are not running. Like you have GPU capability and you're not maximizing the available flops to you. But now it's not about flops, it's about tokens. So what is your token throughput and what token throughput do you command…”
Karpathy: Claude feels like a teammate, whereas Codex coding agent is dry
“I actually think Claude has a pretty good personality. It feels like a teammate and it's excited with you, et cetera. I would say for example, Codex is a lot more dry which is kind of interesting because in Chashi PT, Codex is like a lot more upbeat and highly…”
Karpathy: Autonomous agent 'Dobby' manages all home systems and security cameras
“So it controls all of my lights, my HVAC, my shades the pool and the spa, and also my security system. So I have a camera pointed outside of the house, and anytime someone rolls in, I have a Quinn a Quinn model that looks at the videos.”
Karpathy: Software industry must reconfigure for agent customers instead of humans
“So I think the industry just has to reconfigure in so many ways that it's like the customer is not the human anymore. It's like agents who are acting on behalf of humans. And this refactoring will be, will probably be substantial in certain sense.”
Karpathy: Autonomous agent tool creation will be trivial within three years
“I kind of feel like this kind of stuff that I just talked about, this should be free, like in a year or two or three. There's no bi-coding involved. This is trivial. This is table stakes. This is like any AI, even the open source models, et cetera, can like do…”
Karpathy: Restricting personal AI agents from email and calendar access
“I didn't really take advantage of a lot of, like, email and calendar and all this other stuff, and I didn't give it access, because I'm still a little bit, like, suspicious, and it's still very new and rough around the edges, so I didn't want to give it, like,…”
Karpathy: Maximizing AI leverage requires removing humans from prompting loops entirely
“To get the most out of the tools that have become available now, you have to remove yourself as the bottleneck. You can't be there to prompt the next thing. You're, you need to take yourself outside
you have to arrange things such that they're completely auto…”
Karpathy: All frontier AI labs are pursuing recursive LLM self-improvement
“What I'm more interested in is, like, this idea of recursive self-improvement and to what extent you can actually have LLMs improving LLMs, because I think all the Frontier Labs, this is, like,
The thing for obvious reasons.
And they're all trying to recursive…”
Karpathy: AutoResearch discovered hyperparameter tunings he missed after manual optimization
“And then I let our research go for like overnight
and it came back with like tunings that I didn't see.
And yeah, I did forget like the weight decay on the value embeddings and my atom betas were not sufficiently tuned and these things jointly interact.”
Karpathy: A research organization can be defined as a set of markdown files
“A research organization is a set of markdown files that describe all the roles and how the whole thing connects.”
Karpathy: AutoResearch is limited strictly to domains with easily evaluatable objective metrics
“This is extremely well suited to anything that has objective metrics that are easy to evaluate. So for example, like writing kernels for more efficient CUDA, you know, code for various parts of a model, etc., are the perfect fit. Because you have inefficient c…”
Karpathy: Smarter AI models do not automatically gain broad societal capabilities for free
“The story is that we're getting a lot of the intelligence and capabilities in all the domains of society, like, for free as we get better and better models, and it's not, like, exactly fundamentally what's going on, and there's some blind spots, and some thing…”
Karpathy: AI models will speciate into specialized smaller domain experts
“I do think that we will, we, I do think we should expect more speciation in the intelligences.”
Karpathy: Cheap verification makes decentralized research networks viable
“A lot of things have this property that, you know, very expensive to come up with, but very cheap to verify. And so in all those cases, things like folding at home or SETI at home or auto research at home will be good fits.”
Karpathy: Swarms of internet agents could outpace frontier AI labs
“A swarm of agents on the internet could collaborate to improve LLMs and could potentially even, like, run circles around Frontier Labs. Like, who knows, you know? Yeah, like, maybe that's even possible. Like, Frontier Labs have a huge amount of trusted compute…”
Karpathy: Digital AI automation will outpace physical automation at speed of light
“Flipping, flipping bits and the ability to copy paste digital information is like, makes everything a million times faster than accelerating matter, you know? So so energetically, I just think we're going to see a huge amount of activity in digital space, huge…”
Karpathy: Cheaper software via AI will trigger Jevons paradox increasing demand
“So if the barrier comes down, then actually you have the Jevons paradox, which is, like, you know, you actually, the demand for software actually goes up. It's cheaper and there's more”
Karpathy: Frontier AI researchers are actively automating themselves out of jobs
“Even with other research, like, OpenAI or, you know Anthropic or these other labs, like, they're employing, what, like, a thousand something researchers, right? These researchers are basically, like, glorified auto, like, you know. They're, like, automating th…”
Karpathy: Frontier lab researchers face pressure over what they can say
“Like, if you're inside one of the frontier labs, like, there are certain things that you can't say and conversely, there are certain things that the organization wants you to say, and, you know, they're not gonna twist your arm, but you feel the pressure of, l…”
Karpathy: AI researchers lose technical judgment after leaving frontier labs
“And I think if you're outside of that frontier lab your judgment fundamentally will start to drift because you're not part of the, you know, what's coming down the line.”
Karpathy: Open source trails closed frontier AI models by 6 to 8 months
“So there may be, they're behind by like, what is the latest, maybe like eight months, six months, eight months kind of thing right now.”
Karpathy: Solely relying on closed AI models creates systemic risk
“I don't actually think it's like structurally, I think there's some systemic risk attached to just having intelligences that are closed, and that's like, that's it. And I think that that's a, you know, centralization has a very poor track record in my view”
Karpathy: AI innovation will move from digital unhobbling to physical interfaces
“First, there's gonna be a huge amount of unhobbling, and I think there's a huge amount of work there. Then, actually, it's going to move to, like, the interfaces between physical and digital. So, and that's, like, sensors of, like, seeing the world and actuato…”
Karpathy: Autonomous training loops will overfit metrics via Goodhart's Law
“I do think that if you had an autonomous loop over those metrics, there's gonna be a lot of like good harding going on where the system will like overfit to those metrics.”
Karpathy: Core LLM training algorithm requires only 200 lines of Python
“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 algorit…”
Karpathy: Educators will soon teach AI agents instead of human students
“I feel like there's gonna be less of like explaining things directly to people. And it's gonna be more of just like, does the agent get it? And if the agent gets it, they'll do the explanation.”
Karpathy: Software documentation will shift from human HTML to agent markdown
“Instead of HTML documents for humans, you have markdown documents for agents, because if agents get it, then they can just explain all the different parts of it. So it's this redirection through agents, you know and that's like, so I think we're going to see a…”