NNaval
← All episodes
28 April 2026

A Return to Code

3Frameworks
14Insights

Listen

Frameworks in this episode

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster

Ten agents talking is not ten people talking

The multi-agent analogy to human collaboration breaks down, in Naval's view, because humans are trained on ten different data sets while ten instances of a model share one. Running them in parallel is functionally identical to spending ten times the tokens, or telling one model to take ten times as long. Different labs' models diverge slightly and may add some insight, but not much. He also notes AI is jagged: brilliant at some things, hopeless at others.

  • Humans collaborate well because each brings a different training set
  • Ten instances of one model share a brain and a data distribution
  • Multi-agent setups mostly buy you more tokens on the same problem
  • Codex, Gemini and Grok Code differ slightly, but not that much
  • AI is jagged intelligence, not uniformly smart

It's like 10 people with the same brain, the same data set talking to each other.

Naval

So all you're doing when your 10 agents are talking to each other is you're just throwing 10 times as many tokens at the problem.

Naval
#multi-agent#ai-limits#groupthink

Hot Take· 3

Hot Take

Vibe coding is a video game with real rewards and no ceiling

Naval maps the addictiveness of vibe coding onto game design. A video game keeps you hooked by holding difficulty at the edge of your capability and paying out constant feedback, but the rewards are fake, the world is bounded, and once you learn the rules it gets boring. Vibe coding runs the same loop with a Turing machine underneath, so the space never closes, the objective is set by you, and the output has real-world relevance.

  • Games tune difficulty to the edge of capability and pay constant feedback
  • Game worlds are bounded and the rewards are fake, so they eventually bore you
  • Vibe coding is unbounded because a Turing machine sits underneath
  • You set the objective, so it keeps expanding rather than filling up
  • Naval's reading, doom-scrolling and gaming time has been reallocated to coding

Except with vibe coding, it's unbounded because now you've got a touring machine running underneath.

Naval

But those rewards are fake. And the video game is bounded, it's created by other humans.

Naval
#motivation#vibe-coding#psychology
Hot Take

Pure software is uninvestable, full stop

Naval sharpens his own tweet, saying the watered-down version was that pure software is rapidly becoming uninvestable and the real claim is that it simply is uninvestable. Two reasons: anyone can hack the thing together today, and agents are improving so fast that within a year or less they will produce scalable software with good architecture. He redirects venture money toward hardware, network effects and AI models, and argues training models is the new building software until auto-research and auto-training arrive.

  • If your only edge is software others cannot build, you have no edge
  • Coding agents will likely produce well-architected scalable code within a year
  • Venture attention should shift to hardware, network effects and AI models
  • Training models is the new building software, for a limited window
  • The counter-question is whether incumbents gain from distribution instead

which is that pure software is uninvestable. I would just full stop right there.

Naval

So if you're a venture investor now, you're looking for hardware, you're looking for network effects, you're looking for AI models.

Naval
#venture-capital#investing#software#moats
Hot Take

Apple giving up on AI is the decade's biggest strategic mistake

Naval argues Apple's moat erodes once interaction moves from tapping apps to talking to an agent. If you say call me an Uber or track my workout instead of opening an app, the phone reduces to a screen, a battery and connectivity, which Android supplies fine. Apple now leans on Google's Gemini, so the software differentiator is gone, leaving hardware margins closer to Samsung or Lenovo. He draws the parallel to Microsoft missing mobile: still valuable, but capped.

  • Conversational and agentic interfaces dissolve the app-tapping habit
  • Agents do not even need APIs; they can create their own on the fly
  • Apple relying on Gemini removes its software differentiation
  • Without the OS advantage, margins compress toward Samsung or Lenovo levels
  • Microsoft is the precedent: still valuable, but it lost the mobile wave

I think Apple giving up an AI will go down as the biggest strategic mistake in the tech industry of this decade.

Naval

It should all be conversational. It should all be agentic.

Naval
#apple#big-tech#strategy#agents

Explainer· 5

Explainer

The moment coding agents stopped being assistants

Naval dates the shift to around December 2025 and the release of Claude Opus 4.5, when agents began staying on track across whole builds rather than returning snippets. The old model was a coding assist: ask for a fix, receive a pile of code, paste it into your IDE. The new model is a long-lived agent that builds an app end to end. He had tried earlier agents with mixed results, but says this time the hype felt real.

  • December 2025 and Claude Opus 4.5 marked the inflection point
  • Agents stay on track and build apps soup to nuts rather than emitting snippets
  • The experience resembles having a fast, free junior programmer
  • Naval has a computer science degree but had not seriously coded in decades
  • Activation energy, not ability, was the old barrier to writing code

So around December of 2025, the coding agents in AI hit an inflection point with the release of Claude Opus 4.5.

Naval

And the activation energy to writing code is really high.

Naval
#ai-agents#vibe-coding#developer-tools
Explainer

Why coding agents are really Unix programs wearing English

Naval explains the architecture behind the terminal-first workflow. Agents are trained on text tokens, and most of the code they trained on was Unix, so they are natively fluent in the shell. They run in a command line interface, execute commands, reach the file system, chain grep, awk, sed and pipes, run cron jobs to stay long-lived, and spawn additional shells and tasks. Modern operating systems are Unix underneath anyway, with macOS famously built on BSD.

  • Agents are text in, text out, which is exactly what Unix is
  • Most training code on GitHub and Stack Overflow was Unix
  • They connect to the shell, the file system and standard commands
  • Cron jobs make them long-lived; spawned shells make them parallel
  • The AI acts as a translator from English into the tool jargon

So these agents are just long-lived coding AIs that are connected to Unix at a core level.

Naval

So underneath these are all Unix, which is all text in, text out.

Naval
#unix#cli#ai-agents#architecture
Explainer

Every recent AI leap is an error-correction trick

Naval reframes the last few years of AI progress as successive applications of error correction. Agents error correct and learn, some of them watching YouTube videos at night to improve at tasks assigned during the day. Reasoning models turn next-token prediction into a pseudo-thinking process that error corrects at each step. Reducing hallucinations was error correction too. He wonders aloud what the next application will be, and floats agents correcting each other, while doubting the analogy holds.

  • Agents go out, error correct, and improve their own skills
  • Thinking models are next-token prediction with step-wise error correction
  • Hallucination reduction was itself an error-correction process
  • The open question is what error correction gets applied to next
  • Agent-to-agent correction is the obvious candidate and the weakest analogy

The most interesting thing to me about agents is their ability to error correct and learn.

Naval

Getting rid of hallucinations was also an error correction process.

Naval
#ai-research#reasoning#agents
Explainer

Why a growing codebase breaks a coding agent

Naval explains the mechanical reason agents degrade on large projects. State of the art is roughly a million tokens of context, which he says will look laughable in future, and the transformer attention mechanism scales as the square of the token count. Once the codebase exceeds what fits, the model starts guessing, approximating and compacting. The symptoms are recognisable: losing the plot, fixing the wrong thing, and fixing the same bug five times.

  • About a million tokens is the current ceiling, roughly a million words
  • Attention cost scales as the square of context length
  • Past the ceiling the model compacts, guesses and approximates
  • Repeat fixes on one bug are the diagnostic signal
  • Guidance shifts onto the operator as complexity grows

And right now, the state of the art is about a million tokens, which will be considered laughable in the future.

Naval

They start losing the plot, they get lost. They start fixing the wrong thing. They fix the same bug five times.

Naval
#context-window#transformers#ai-limits#codebase
Explainer

Great engineers' taste is now inside the models

Naval offers two explanations for the recent jump in coding model quality. One is near-recursive training, where one model helps improve the next. The bigger one, he suspects, is that many of the best software engineers started using these tools in the last few months, so the models absorbed not just their code but their judgement about what is good. He generalises: improving models requires high-taste feedback loops, which are far harder to build than they look.

  • Recursive training, where one model improves the next, is part of it
  • The larger factor may be top engineers adopting the tools en masse
  • The models gained their code and their taste at the same time
  • High-taste feedback loops are the real scarce input
  • Some domains make such loops tractable; others do not

But I think the bigger reason might just be that a lot of the best software engineers started using these models in the last few…

Naval
#model-training#taste#feedback-loops

Story· 1

Story

The compromise tax of building with a team

Naval contrasts building AirChat with eight or nine engineers against rebuilding it alone with an agent. Every team build carries compromises, even when you are nominally in charge, because you cannot keep asking an engineer to move an icon left, then right, then back, and you cannot demand changes you can only justify by gut feel. An agent removes the social cost entirely, which he likens to not feeling self-conscious in a self-driving car.

  • AirChat took eight or nine engineers and nine to twelve months
  • Team building always imposes compromises, even on the person in charge
  • You cannot justify pure-intuition requests to another human indefinitely
  • An autonomous agent has no ego to manage, so idiosyncrasy is free
  • Expect more Minecraft-shaped products: one person's undiluted vision

It's like a self driving car. You don't feel self conscious in a self driving car because there isn't a driver sitting there.

Naval

So you may end up with more things like Minecraft, which Notch famously coded by himself, where there was one person's vision.

Naval
#product#solo-builder#airchat#creativity

Tool· 2

Tool

Which frontier model Naval reaches for, and why

Naval routes questions by each model's underlying data and character. Claude has strong visual presentation through Artifacts and is unusually good at meeting you at your level of understanding. ChatGPT is the OG and strong all round. Gemini is a frustrating product that times out and loses the plot, but it has the Google crawl and YouTube behind it, so search-shaped questions go there. Grok is the least neutered, has X for news, and holds up on hard technical and scientific questions.

  • Claude: Artifacts for visual output, and it calibrates to your level
  • ChatGPT: the original, strong generalist
  • Gemini: weak product, best underlying data via Google crawl and YouTube
  • Grok: least filtered, X access for news, strong on deep technical problems
  • Each model's strengths reflect the biases of the company that trained it

And then Grok is the one I can count at to tell me the truth. It's like the least neutered, least nerfed.

Naval

Gemini is very good at search because it has the Google crawl underneath.

Naval
#model-selection#claude#gemini#grok#chatgpt
Tool

An overnight bug queue that fixes itself into side branches

Naval describes the maintenance loop inside the app he is building. Users tap a button to report a bug, logs and the report upload to a server, and every 24 hours Claude works through the whole queue unattended, landing each fix in a side branch. His only job is review: reject the ones that were not real bugs or not good fixes, ship the rest. He extends this to a future where users request and vote on features and a maintainer in the cloud arbitrates.

  • In-app bug reporting uploads logs automatically to a server
  • An agent clears the queue every 24 hours without intervention
  • Fixes land in side branches rather than shipping directly
  • The human is reduced to a final approval gate
  • Perfect customer service is an indefatigable coder with no ego

And then I have Claude go every 24 hours through all the bug reports, and it just fixes them all by itself without my having…

Naval

I'm just the final gate that decides on what goes out there.

Naval
#automation#bug-triage#agents#product-development

Takeaway· 2

Takeaway

Vibe coding teaches kids what Scratch never could

Naval notes how hard it is to get kids to program, despite tools like Swift Playgrounds and Scratch Junior. Vibe coding gives instant feedback and instant rewards, and because operating the agents still requires skill, the fundamentals arrive as a side effect. Kids end up learning the command line, basic computer architecture, caching, network back-off, streams, writing to disk and latency versus bandwidth trade-offs simply because the work demands it.

  • Swift Playgrounds and Scratch Junior rarely convert kids into programmers
  • Vibe coding supplies instant feedback and rewards instead
  • Operating the agents still requires real skill
  • Fundamentals arrive as a by-product: CLI, caching, latency, streams
  • Learning happens by doing rather than by curriculum

It's hard to get kids to program. You can throw Swift Playgrounds and Scratch Junior and all of that at them and hope that they…

Naval

But if you throw vibe coding at them, they're gonna get instant feedback and instant rewards.

Naval
#education#kids#learning#vibe-coding
Takeaway

One- and two-person companies that scale to millions

Naval expects the small-team outlier to become common. It has happened before with Notch, Satoshi Nakamoto, and the original Instagram and WhatsApp teams making enormous dents with very few people, but agents make it repeatable. He also estimates the share of the population capable of building apps rises from roughly 0.1 percent to one to three percent, while stressing that most people will still treat computers as a magic black box.

  • One- and two-person companies can now scale to millions of users
  • Precedents: Notch, Satoshi Nakamoto, early Instagram, early WhatsApp
  • The app-building share of the population goes from ~0.1% to 1-3%
  • Most people still experience computers as a black box, so 10x easier changes nothing for them
  • The winners are the creative, self-motivated and articulate with a clear vision

You truly can have one person, two-person software companies now that can scale to millions upon millions of users and make billions upon billions of…

Naval

But for the people who are creative, who are self-motivated and who are articulate and have a good vision, you can code now.

Naval
#solo-founder#startups#software-economics