No PriorsSkill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI
CHAPTERS
- 0:00 – 3:24
From typing code to “manifesting” with agents: the new default workflow
Sarah and Andrej set the stage for how dramatically day-to-day software creation has changed since late 2024. Andrej describes the shift from hand-writing code to delegating almost everything to code agents—and the resulting sense of “AI psychosis” from suddenly-expanded leverage.
- •Coding becomes giving intent/instructions rather than writing lines
- •December as the inflection point: delegation flips from 80/20 to 20/80 (and beyond)
- •The new obsession: exploring what’s possible with agentic workflows
- •Why this change isn’t yet widely internalized outside the AI bubble
- 3:24 – 6:36
Token throughput and parallelism: why the human becomes the bottleneck
They discuss how engineers now optimize for token budgets the way researchers used to optimize for GPU utilization. The mental model shifts to running many parallel agent threads, with the person acting as scheduler, reviewer, and bottleneck remover.
- •Stress comes from realizing you can always spin up “one more agent”
- •Analogy to PhD life: nervousness when GPUs (now tokens) aren’t fully utilized
- •Peter Steinberg’s “tiled agents” workflow as a model for parallel delegation
- •Work moves to macro-actions: delegating features, planning, research, and review in parallel
- 6:36 – 7:39
What mastery of coding agents looks like: climbing the stack beyond single chats
Andrej sketches what “agent mastery” could mean over the next few years: multiple collaborating agents, better persistence, and higher-level abstractions. He introduces the idea of a ‘claw’—a persistent, looping entity that keeps working even when you aren’t watching.
- •Mastery = moving from one agent session to coordinated multi-agent systems
- •‘Claw’ concept: persistent loops, sandboxes, and ongoing background execution
- •Memory systems become a key differentiator (beyond simple context compaction)
- •The goal: higher leverage via fewer human interventions
- 7:39 – 9:17
Why OpenClaw resonates: personality, memory, and a single portal to automation
They break down why OpenClaw feels different from typical agent tooling. Andrej argues that agent “personality” and calibrated feedback matter more than people expect, and that convenient surfaces (e.g., WhatsApp) make the system feel like a real entity.
- •OpenClaw innovates across multiple dimensions, not just one feature
- •Personality as product: agents that feel like teammates vs. dry executors
- •Calibration of praise/sycophancy affects trust and engagement
- •A unified “single portal” (WhatsApp) makes automation intuitive and accessible
- 9:17 – 11:45
Claws in the real world: Dobby the home-automation agent (and its implications)
Andrej shares a concrete ‘claw’ deployment: a WhatsApp-driven home automation system that discovers devices, reverse-engineers APIs, and orchestrates routines. This becomes a broader argument that many bespoke apps may collapse into APIs + agent glue.
- •Agent scans local network, discovers devices (Sonos/lights), and uses APIs
- •Natural-language routines (“sleepy time”) replace multiple vendor apps
- •Vision pipeline: change detection + model (Qwen) to classify events and message updates
- •Prediction: many consumer apps shouldn’t exist—agents should orchestrate APIs directly
- 11:45 – 16:21
Second-order effects: agent-first software, ephemeral code, and trust boundaries
They explore how software UX changes when the ‘customer’ is an agent acting for a human. Andrej notes the near-term reality still involves ‘vibe coding,’ but expects barriers to collapse—and flags privacy/security as a limiting factor for deeper integration (email/calendar).
- •Shift from human-first UI to agent-first APIs and workflows
- •Ephemeral, on-demand software assembled by agents from intent
- •Today requires design choices and technical effort; soon becomes ‘table stakes’
- •Security/privacy concerns slow adoption of deeper personal automation
- 16:21 – 20:54
Why AutoResearch: removing the human from the research loop
Andrej explains AutoResearch as a practical consequence of maximizing leverage: you must structure systems so they run autonomously against objective metrics. He describes using Nanochat as a playground for recursive self-improvement—LLMs improving LLM training runs.
- •Core idea: you can’t be the next-prompt bottleneck—set objective + boundaries and ‘hit go’
- •Nanochat as a harness to study recursive self-improvement in a controlled setting
- •AutoResearch found tuning improvements Andrej missed (hyperparameters and interactions)
- •Frontier-lab analogy: explore on small models, then extrapolate via scaling laws
- 20:54 – 23:16
Meta-optimization: Program.md as the ‘code’ of a research organization
They take AutoResearch one level up: if research processes are encoded in Markdown, those instructions can also be tuned. Sarah proposes a contest where people submit competing Program.md recipes; Andrej agrees this enables systematic improvement of the research org itself.
- •Program.md describes roles, loops, and the ‘operating system’ of research
- •Different org designs can be compared empirically by output improvement per compute
- •Human contributions become modular: ideas funnel into a queue executed by workers/agents
- •Recursive layers (LLM → agent → claw → instruction optimization) create “infinite” complexity
- 23:16 – 28:55
Capability limits that remain: jagged intelligence and the boundary of verifiability
They diagnose why the current moment still feels brittle despite huge gains. Andrej argues progress is strongest where evaluation is objective/verifiable (tests, speedups) and weakest in soft/nuanced domains—producing ‘jaggedness’ like brilliant coding paired with terrible jokes.
- •AutoResearch works best with crisp metrics; weak where evaluation is fuzzy
- •Models feel like a ‘brilliant PhD student’ plus ‘a 10-year-old’—inconsistent competence
- •Reinforcement learning focuses improvement on verifiable domains
- •Evidence of limited generalization across domains (e.g., joke quality stagnation)
- 28:55 – 32:57
Model speciation: why we may need specialized intelligences (and why it’s hard)
Sarah asks whether jaggedness implies unbundling models into domain experts. Andrej expects ‘speciation’—smaller competent cores with specialized skills—but notes the field lacks mature primitives for safely touching weights (continual learning, fine-tuning without regressions).
- •Prediction: more specialized models for niches (math in Lean, domain tools, etc.)
- •Compute constraints and serving costs could pressure specialization
- •Today’s reality: monoculture models because labs must handle unknown user demands
- •Speciation is limited by undeveloped science of weight-editing vs. prompt/context control
- 32:57 – 37:53
Collaboration surfaces for humans + AI: untrusted swarms, verification, and “AutoResearch@home”
Andrej explores how to parallelize research beyond a single trusted cluster by tapping an untrusted internet workforce. He outlines a system reminiscent of blockchains: expensive search to propose candidate commits, cheap verification by a trusted validator—enabling large-scale distributed experimentation.
- •Goal: integrate untrusted contributors who submit candidate code improvements
- •Cheap verification property: many tries are expensive, but checking a winning commit is easy
- •Security challenge: safely running arbitrary code from outsiders
- •Analogy to SETI@home/Folding@home; future: donate compute to causes via AutoResearch swarms
- 37:53 – 48:54
Jobs, skills, and Jevons paradox: how AI reshapes digital work first
Using BLS data, Andrej discusses which professions are likely to change fastest—especially jobs that manipulate digital information. He emphasizes tool adoption, argues demand may rise as software becomes cheaper (Jevons paradox), and notes the uneasy reality that even AI labs automate their own roles.
- •Digital work changes faster than physical work because bits move faster than atoms
- •Jobs are bundles of tasks; near-term: AI is an empowering tool for many roles
- •Jevons paradox in software: cheaper creation can increase total demand
- •Tension inside labs: success implies researchers automate themselves away
- 48:54 – 54:18
Open vs. closed models: converging timelines, ecosystem power balance, and systemic risk
They assess how close open source is to the frontier and why the gap has narrowed. Andrej argues the Linux analogy motivates an open platform for industry trust, while closed frontier models still matter for hardest problems—yet an all-closed world creates systemic risk and centralization concerns.
- •Open source catching up from ‘nothing’ to months behind frontier models
- •Linux analogy: industry wants a common, safe open platform
- •CapEx makes open competition harder, but many consumer use cases will be covered locally
- •Concern: excessive centralization; healthier ecosystem needs multiple strong labs + open alternatives
- 54:18 – 1:01:30
Autonomous robotics and the physical world: digital-first, then sensors/actuators, then atoms
Drawing from self-driving, Andrej predicts robotics will lag digital agents due to capital intensity and real-world messiness. He proposes a trajectory: massive near-term transformation in digital space, then growth at the physical/digital interface (sensors and actuators), and eventually broad physical automation.
- •Self-driving as the first major robotics application; many startups struggled over time/capex
- •Near-term: digital unhobbling and refactoring happens far faster than physical automation
- •Next frontier: interfaces—data collection (sensors) and world manipulation (actuators)
- •Emerging need for information markets to acquire real-world data on demand
- 1:01:30 – 1:06:09
MicroGPT and agentic education: explain to agents, not to humans
Andrej introduces MicroGPT as a distilled, ~200-line expression of LLM training fundamentals, emphasizing that most complexity in real systems is for efficiency. He argues education will reroute through agents: write docs/curricula for AI tutors, while humans focus on the few insights agents can’t yet originate.
- •MicroGPT as the minimal core: architecture + autograd + optimizer + loop in ~200 lines
- •Efficiency drives most real-world training code complexity; algorithms are simpler
- •Education shifts: agents provide infinite explanations; authors provide ‘skills’ and curricula hints
- •Human value concentrates in original distillation and taste—what agents can’t yet invent
- 1:06:09 – 1:06:31
Wrap-up and where to follow the show
Sarah thanks Andrej and closes with ways to subscribe and find transcripts. The episode ends with standard show outro and links.
- •Closing thanks and acknowledgements
- •Where to follow No Priors on social and podcast platforms
- •Transcript and email signup at no-priors.com
