Amjad invokes the Jevons Paradox — when the cost of something falls, total consumption rises rather than falls, as with electricity. He argues software will follow the same pattern: as building gets cheaper, people won't build less, they'll build far more software to improve their lives and work and to start more companies. Lower activation energy also means people build ideas they otherwise never would have.

Jevons Paradox: falling cost raises total consumption, not lowers it
As software gets cheaper to build, people will build much more of it
Lower activation energy unlocks ideas that would never otherwise be built
Cheaper software should mean more startups, not fewer

“when the cost of things go down uh the total consumption of it goes up”
Amjad Masad · 24:00

“I think that's going to be the case of software like as the costs go down people will just like make a lot more soft…”
Amjad Masad · 24:30

#economics#jevons-paradox#ai#startups

Hot Take38:30

Why Software Is How AI Becomes General

Amjad describes a coming autonomous mode: instead of testing an app himself, another agent tests it, and you interface entirely through Slack — e.g. 'get me Taylor Swift tickets the moment they land,' and it builds an app that monitors the web and buys via a wallet. His larger claim: because software runs our lives, the internet, and our businesses, an AI that can write software is how AI becomes more general.

A coming feature uses a second agent to test the app autonomously
You could interface entirely through Slack with natural-language goals
Example: an agent that monitors the web and buys Taylor Swift tickets the moment they drop
Because software runs everything, software-writing AI is the path to generality

“give me Taylor Swift tickets the the moment they land and so it'll build a an app uh that continuously monitors the web”
Amjad Masad · 38:30

“software like agents being able to do software is how AI gets more General because software runs Our Lives runs the internet runs our businesses”
Amjad Masad · 39:00

#autonomous-agents#agi#future#automation

Hot Take47:00

Amjad's Law: The ROI on Learning to Code Doubles Every 6 Months

Amjad shares a coinage others dubbed 'Amjad's Law': the return on investment for learning to code is doubling roughly every six months. Rather than learning coding the traditional way (starting with git), he argues non-engineers should pick up a little coding by talking to an AI and debugging. Counterintuitively, specific engineering skills become more valuable as agents write more of the code — because you use them to unblock the agent.

'Amjad's Law': the ROI on learning to code is doubling roughly every six months
Learn by prompting AI and debugging, not by starting with git in a bootcamp
Specific engineering skills grow more valuable even as agents write the code
The value is in unblocking the agent and understanding the mental model, not knowing everything

“dubbed as I'm Jed's law which is um the return on investment for learn a code is doubling every six months”
Amjad Masad · 47:00

#skills#learning-to-code#careers#engineering

Hot Take54:00

A Billion-Dollar Company With Zero Employees

Amjad imagines that within roughly five years someone could run a billion-dollar company with zero employees — support, development, and maintenance all handled by AI. But he raises the economic catch: if anyone can generate a Salesforce, software prices may fall, so the durable edge becomes being generative — iterating and improving faster than everyone else building the same tools.

Within ~5 years someone could run a billion-dollar company with zero employees
AI could handle support, development, and maintenance
If anyone can generate a Salesforce, software prices may collapse
The durable advantage becomes iterating and generating new ideas faster than competitors

“now uh someone running you know a billion dollar company with zero employees where it's like the support is handled by AI the development is…”
Amjad Masad · 54:00

“can you actually build the next Salesforce if anyone can generate Salesforce”
Amjad Masad · 54:30

#future-of-work#startups#economics#ai

Explainer· 4

Explainer02:30

Why Making Software Is So Hard (And Replit's Fix)

Amjad explains that building software is difficult mostly because the toolchain is fragmented: you need an IDE, a runtime, a package manager, and a way to deploy and share. Replit's vision is to collapse all of that into one place so making software becomes easy, learnable, and fun. He argues the friction of setup is where most would-be coders give up.

Software is hard largely because the toolchain is fragmented across IDE, runtime, package manager, and deployment
The setup friction is where people get stuck and never learn to code
Replit's vision: make software creation easy, in one place, and learnable
New AI products let people jump in with no prior coding experience

“the idea behind repet is that uh making software today is is is very difficult”
Amjad Masad · 02:30

#replit#developer-tools#vision#no-code

Explainer09:00

How Replit Differs From Cursor and Other AI Coding Tools

Amjad positions Replit as an end-to-end platform covering the whole software lifecycle from writing code to deploying and monetizing it, versus point tools like Cursor, which is a VS Code fork that still needs a separate runtime and deployment environment. Being opinionated about everything makes Replit hard to adopt inside big engineering pipelines, but it lets non-engineers — PMs, designers, ops, lawyers — build software.

Cursor is a VS Code fork with great AI tools but still needs a runtime and deployment environment
Some users run Cursor alongside Replit for Replit's runtime and deploy layer
Replit's differentiator is covering the entire software lifecycle end-to-end
Being opinionated makes it hard for big-company pipelines but empowers non-engineers to build

“cursor is a fork of vs code that's made um that has really awesome AI tools but that's an editor you still need a runtime…”
Amjad Masad · 09:00

“we want to Empower everyone to to be able to build software and that means product managers designers”
Amjad Masad · 10:00

#replit#cursor#competitive-landscape#product-managers

Explainer15:30

The AI Agent Is Built as Another 'Player' Coding Alongside You

Replit reused its real-time multiplayer coding system to build the agent, structuring the agent as just another user of the platform. So the human and the AI are coding together in the same live session, and you can open the files, read them, or ask the AI to explain — unlike pure chat interfaces. It feels like watching an engineer code over their shoulder.

Replit already had real-time 'multiplayer' collaborative coding
The agent is structured as another user of that multiplayer system
Human and AI code together in the same live session
Unlike pure chat tools, it's a full IDE where you can inspect and edit files

“you're basically sitting there behind an engineer on a computer and just watching them code is what this feels like”
Lenny Rachitsky · 15:30

“the agent in the code is structured as a another user of the platform so basically we're both coding together together”
Amjad Masad · 16:00

#replit#ai-agent#architecture#multiplayer

Explainer31:30

Building a Computer for the AI: ACI and a 'Society of Models'

Amjad describes a new discipline — AI computer interfaces (ACI), the machine analog of human-computer interaction (HCI). Because LLMs are 'alien creatures,' they need interfaces different from humans': text representations of the shell, dedicated tools for package installs and editing, rather than expensive image-based computer use. Replit runs a multi-agent 'society of models' — Claude's Sonnet for coding, OpenAI models for critique and management, and internally trained embedding models.

AI computer interfaces (ACI) are the machine analog of HCI
LLMs need different interfaces than humans — e.g. text views of the shell instead of images
Image-based computer use (like Anthropic's) is powerful but expensive
Replit runs a multi-agent 'society of models', not a single model
Sonnet from Anthropic is used for coding; OpenAI models handle critique/management; embeddings are trained in-house

“there's like almost like a new discipline um called um AI computer interfaces so sort of like HCI is now a ACI”
Amjad Masad · 31:30

“it's going to be Society of models like products will be made of a lot of different models”
Amjad Masad · 33:00

#ai-architecture#aci#multi-agent#foundation-models

Story· 3

Story07:30

An 11-Year-Old Built a Full App — And the Coding Was the Easy Part

Amjad recounts a viral tweet about an 11-year-old girl who built an app on Replit from an idea. A commenter objected that you also need to host it, build a database, and deploy it — and the point is that Replit abstracts all of that away. The surprising part isn't the coding; it's that all the surrounding infrastructure disappears.

A viral tweet described an 11-year-old building an app on Replit from just an idea
Hosting, database, and deployment are the parts that normally block beginners
Replit abstracts away all the infrastructure 'nonsense' around coding
The impressive part is the abstraction, not the code generation itself

“the the surprising thing about an 11year old building an app is not so much even the coding it is like all the nonsense around…”
Amjad Masad · 08:00

#replit#accessibility#story#deployment

Story20:00

A Few Days of Engineering vs. 15 Cents of Compute

After the agent builds a full-stack feature-request app live, Amjad estimates that a decent engineer would need a few days to a week to build the same thing. The agent did it in roughly five to ten minutes, at an estimated compute cost of about 15 cents. The contrast dramatizes the collapse in the cost and time of building working software.

A decent engineer would take a few days to a week to build the demo app
The agent produced it in roughly five to ten minutes
Estimated compute cost was about 15 cents
The comparison illustrates the collapsing cost of building software

“it probably would take me a few days I would say I'm a I'm like a decent engineer”
Amjad Masad · 20:00

“I would estimate it at like 15 cents or something like that”
Amjad Masad · 20:30

#replit#cost#productivity#demo

Story25:00

How Real Teams Actually Use Replit Right Now

Amjad shares concrete usage: SMBs like real estate agents build internal back-office tools instead of buying ill-fitting SaaS; product managers build a V1 and test it with real users before handing it to engineers; a head of marketing at SpotHero built a full-stack competitor price-benchmarking app. The through-line is unblocking non-engineers from needing engineers for every build.

SMBs build back-office tools as a SaaS replacement that fits their exact needs
PMs build and test a real V1 with users before it goes to engineering
A SpotHero marketing lead built a full-stack competitive pricing analysis app
Sales engineers spin up prototypes quickly for customers
It unblocks PMs from needing engineers for every idea

“a lot of people are seeing it as sort of a SAS replacement for in-house tools”
Amjad Masad · 25:30

“we've seen uh product managers build like I said like a V1 of an app and actually go out and test it with users”
Amjad Masad · 26:00

#use-cases#product-managers#saas-replacement#prototyping

Tool· 1

Tool1:01:00

Agent vs. Assistant: Two Modes for Building on Replit

Amjad previews Assistant, the 'cousin' of Agent. Agent is high-agency — it sets up the project, creates the database, and debugs for you, like handing a developer a PRD to go build the whole thing. Assistant is less powerful but far more controllable and much faster (milliseconds to seconds), for focused, small increments — like sitting next to the developer and asking them to move a button three pixels.

Agent is high-agency: it sets up the project, creates the database, and debugs autonomously
Assistant is less powerful but more controllable and much faster
Assistant responds in milliseconds to seconds for focused, small changes
Mental model: Agent = hand a developer a PRD; Assistant = sit next to them for tweaks

“we're working on assistant so assistant is uh like let's say the the cousin of agent it is a little less powerful but a lot…”
Amjad Masad · 1:01:00

“agent uh is like um you know having a developer uh work you give you give them the PRD right and and and they're going…”
Amjad Masad · 1:02:00

#replit#product#agent#assistant

Takeaway· 3

Takeaway17:00

Where AI Coding Still Breaks: Database Migrations and Iteration

Amjad is candid about the limits: the agent is great for building an MVP and getting initial users, but iterating on a working product is still a struggle. Database migrations are a specific weak spot — changing an app's structure can corrupt the database in an unrecoverable way, which can strand a non-coder. Persistent users often unblock themselves by asking ChatGPT or Claude.

The agent is strong at MVPs and getting initial users
Large iterations on a working product are where it runs into trouble
Database migrations are a specific weakness that can create unrecoverable errors
Non-coders sometimes get stuck; persistent users unblock via ChatGPT or Claude

“it's not very good at data database migrations and so we're trying to fix that”
Amjad Masad · 17:00

“you'll get an MVP past the MVP where it's like a product that's working and you need to change and iterate on it it's it's…”
Amjad Masad · 18:00

#replit#limitations#mvp#database

Takeaway55:30

Humans as the 'God of the Gaps' in AI Tools

Lenny offers a mental model borrowed from 'God of the gaps': humans are currently the gaps in these tools — the human coders you hire on Replit to fix what the agent can't. Over time those gaps shrink as AI closes them itself. Amjad agrees, with the caveat that it depends on whether AI hits some fundamental limit; if not, the trajectory removes the gaps quickly.

Analogy to 'God of the gaps': humans fill what the AI can't yet do
Human coders hired on Replit are today's gap-fillers
Those gaps shrink over time as AI closes them itself
The open question is whether AI hits a fundamental limit (data, Transformers scaling)

“it feels like right now humans are like the gaps in these tools”
Lenny Rachitsky · 55:30

“if there isn't then we are on a massive trajectory of removing these gaps uh quickly”
Amjad Masad · 56:00

#mental-models#future#ai-limits#automation

Takeaway57:00

Kill the Rigid Roadmap and Break the Silos

Amjad's advice to founders and leaders: in an AI-affected world, rigid roadmaps are a liability — when a capability drops (like Anthropic's computer use), you want to switch priorities immediately. Replit itself has no explicit roadmap. He also argues for dissolving the silos between designers, PMs, and engineers, since when designers can code and engineers can design you need a fluid, flexible culture.

Rigid roadmaps are a liability when AI capabilities can drop any week
Replit jumped on Anthropic's computer use immediately because it has no explicit roadmap
Be agile enough to switch priorities right away
Dissolve designer/PM/engineer silos; build a fluid, flexible culture

“one thing that I think is really difficult now is having road maps especially if you're doing anything in AI”
Amjad Masad · 57:00

“being agile not being sort of stuck with with road maps being able to kind of just just say oh we're just going to just…”
Amjad Masad · 57:30

#leadership#roadmaps#org-design#agility

Frameworks in this episode

AI Computer Interfaces (ACI)

Build for the Models Six Months Out

Generative-First Skill Stack for the AI Era

PM Prototype-to-Production Handoff

Roadmap-Less Agility for AI-Era Building

Insights & moments

◆Hot Take· 4

Jevons Paradox for Software: Cheaper Code Means More Code