How AI will change programming with Amjad Masad, CEO of Replit
In this episode of the Founders in Arms podcast, we sit down with Amjad Masad—co-founder and CEO of Replit—to discuss AI coding assistants, the future of software creation, developer productivity, large language models, AGI, and why the barrier between an idea and a working product is collapsing faster than most people realize.
This conversation dives deep into:
AI coding assistants and software creation
Replit’s long-term vision
How LLMs and transformers work
Developer productivity in an AI-first world
AGI, alignment, and reliability
The future of programming careers
If you're building in AI, SaaS, developer tools, or enterprise software, this episode is packed with technical insight and long-term thinking.
In this episode, we cover:
(00:00) Introduction and why this moment in AI matters
Amjad Masad joins the podcast to talk about one of the biggest shifts in computing in years: the rise of AI-assisted software creation.
The conversation opens with Replit’s mission:
Make software the fastest and most accessible it has ever been
Reduce the distance between an idea and a working product
Turn programming into a more collaborative, browser-based, AI-assisted experience
Amjad frames the current AI wave as a historic moment for builders—one that feels comparable to the rise of the modern web, but even faster.
(01:24) What Replit is building and the original product vision
Amjad explains that Replit started with a simple but ambitious idea: a collaborative online programming environment that makes software creation dramatically easier.
Over time, that vision expanded into:
Collaborative development in the browser
Community-driven software creation
Team workflows for developers
AI-powered coding assistance
A faster path from idea to product
His view is that much of Replit’s roadmap was visible early on. The hard part was not imagining it—it was actually building it.
(05:21) Why Replit serves such a broad range of users
Rather than optimizing for a narrow user persona, Amjad thinks about Replit through the lens of jobs to be done.
His core belief:
People do not come to Replit because they fit one demographic profile. They come because they want to make something.
That includes:
Students learning to code
Hobbyists building side projects
Professional developers shipping products
Teams collaborating on software
Founders prototyping startup ideas
The common thread is not background—it is intent.
(09:14) Why developers resist change more than people think
One of the most interesting parts of the conversation is Amjad’s argument that developers—despite being agents of technological change—are often extremely conservative about their own tools.
He points out:
Developers tend to stick with familiar workflows
Tooling habits change slowly
Programming often evolves “one generation at a time”
Innovation in developer tools is harder than outsiders assume
That helps explain why browser-based programming and collaborative coding were slow to emerge, even if the need seemed obvious.
(12:00) The core mission: reducing the distance between an idea and a product
Amjad describes Replit’s deeper mission as shrinking the time between having an idea and getting a real product into people’s hands.
His vision is that software creation keeps compressing:
From complex setup and manual coding
To collaborative development environments
To AI-assisted building
To natural-language prototyping
Eventually to near-instant MVP generation
He shares an example of a user posting an idea and getting a prototype in about 30 minutes—with a human builder accelerated by AI.
The long-term direction is clear: software gets faster to create, easier to test, and more accessible to more people.
(15:15) Why AI will generate prototypes before it replaces full software teams
Amjad’s take is nuanced.
He believes AI will soon be able to generate:
Initial apps
Rough MVPs
Basic software prototypes
Starting points for real products
But he does not believe AI will immediately replace the human work required to:
Iterate on edge cases
Make systems reliable
Maintain and scale software
Understand customers deeply
Turn prototypes into durable businesses
In other words, AI can get you started quickly—but human judgment still matters once software meets reality.
(16:49) How LLMs and transformers actually work
The episode goes deep into the technical side of large language models.
Amjad explains the transformer model in practical terms:
Transformers introduced attention mechanisms
Attention helps models focus on relevant parts of the input
Instead of hand-coding language rules, models learn patterns from data
With enough scale, these systems begin to show emergent reasoning abilities
He describes this shift as part of software 2.0:
Instead of programmers explicitly writing every algorithm, machine learning systems discover algorithms by optimizing over large datasets.
(23:27) Why scale alone was not enough for ChatGPT
Amjad argues that ChatGPT’s leap was not only about more parameters or more data.
He points to two major breakthroughs:
Supervised fine-tuning
Reinforcement learning from human feedback (RLHF)
These helped models become more useful, more conversational, and more aligned with what humans actually want.
His implication is important for founders:
The biggest gains in AI may not come only from scale. They may come from better training methods, better interfaces, and better system design.
(29:01) Why AI is still unreliable for full production software
One of Amjad’s core concerns is reliability.
Traditional software can be tested with deterministic engineering methods:
Unit tests
Program verification
Repeatable behavior
LLMs are different because they are probabilistic and stochastic. That makes them powerful—but also harder to trust in high-stakes environments.
This is one reason Amjad believes AI-generated full-stack products still need human oversight:
Models can hallucinate
Outputs are not always reproducible
Reliability remains hard to guarantee
Traditional engineering workflows do not map cleanly onto LLM behavior
(31:55) Constitutional AI, online learning, and what today’s models still lack
The discussion expands into newer ideas like:
Constitutional AI
Model interrogation by other models
Reinforcement from human feedback
Online learning during deployment
Amjad sees online learning as especially important for the future.
Today’s models usually need retraining after deployment. Truly general intelligence, in his view, would require systems that can:
Learn continuously
Adapt in production
Improve across domains
Update behavior without full retraining cycles
That is one of the key gaps between current LLMs and anything resembling AGI.
(37:54) The economics of AI and whether AI products are too expensive
The episode also covers the economics of inference.
Amjad notes that some AI products may look expensive today—but there is enormous room for optimization across:
Smaller domain-specific models
Better inference routing
Hardware improvements
More efficient chips
Software-level optimization
His view is that current AI economics are not fixed. They are early.
That means founders should be careful not to assume today’s cost structure is permanent.
(42:32) Why software creation may get dramatically cheaper
One of the biggest long-term predictions in the episode is that the cost of creating software will keep falling.
Amjad suggests that:
Basic app creation trends toward zero cost
MVP generation becomes more automated
Cloning or recreating simple software interfaces becomes easier
The real value shifts away from basic implementation and toward judgment, systems, customer understanding, and distribution
This has major implications for startups: building may become cheaper, but winning will still require insight.
(44:21) What happens to software engineers in an AI-first world
Amjad predicts a bimodal future for software talent.
The biggest winners may be:
Platform engineers working on low-level systems, infrastructure, and core tooling
Product-oriented builders who understand customers, markets, and product judgment
The group most at risk:
General-purpose “middle layer” developers doing repetitive application glue work
His argument is that AI will be especially effective at standard implementation tasks, while deeper systems work and product reasoning remain more defensible.
(47:13) How much more productive AI makes developers today
Amjad shares that early estimates suggest meaningful productivity gains from AI coding tools already.
He references a range of outcomes:
Conservative estimates around 20% improvement
Anecdotal reports of much larger gains
Some workflows feeling 2x faster
A belief that 10x productivity improvements may arrive over the next few years
The broader point is that AI copilots are not theoretical anymore. They are already changing how engineers work.
(50:36) Will software be rebuilt for AI—or layered onto existing systems?
A fascinating section of the conversation asks whether AI-native development will require new programming languages and new software architectures.
Amjad’s answer is pragmatic:
Most technological change layers on top of old systems instead of replacing them cleanly.
Just as the internet still carries the baggage of older abstractions, AI will likely be added to existing workflows before fully replacing them.
That means founders should expect messy transitions rather than clean resets.
(53:37) What makes an LLM different from AGI
The conversation then shifts into AGI.
Amjad draws a clear distinction between today’s LLMs and true general intelligence.
His view:
Current models are impressive and generalizable in narrow ways, but they still do not autonomously learn across entirely new domains the way humans can.
For example, current systems still struggle to:
Learn new environments independently
Maintain persistent state in a robust way
operate across domains without retraining
Form durable long-term goals without heavy scaffolding
That leaves a major gap between “useful AI” and true AGI.
(58:08) Consciousness, materialism, and whether intelligence is fully computable
One of the most philosophical sections of the episode centers on the limits of computational models of intelligence.
Amjad raises questions around:
Consciousness
Pain and pleasure as core features of experience
Materialist explanations of the mind
Whether intelligence is fully Turing-computable
The relevance of thinkers like Roger Penrose
The possibility that human reasoning is not fully captured by today’s computational models
He does not dismiss progress in AI—but he cautions against overconfidence in simplistic assumptions about consciousness and machine intelligence.
(1:03:21) Why AI alignment matters even without near-term AGI
Even though Amjad is skeptical of the most extreme AGI doom scenarios, he still takes alignment seriously.
His reasoning is practical:
Humans have historically struggled to align powerful systems with human well-being.
He compares AI alignment to capitalism:
Powerful optimization systems create huge value
But they also create side effects
Harmful outcomes often emerge unintentionally
Alignment is hard even when incentives are visible
So even if near-term AI remains narrow, it still matters how these systems are trained, deployed, and constrained.
(1:07:14) Weaponization, misuse, and the more realistic AI risks
Amjad suggests that more realistic near-term risks may include:
Weaponization
Automated trolling and manipulation
Harmful misuse by bad actors
AI combined with drones or autonomous systems
Dangerous amplification of political or social control
Rather than assuming instant sci-fi catastrophe, he points to a more grounded concern: powerful tools in the hands of humans with bad incentives.
(1:10:34) Why this feels like the biggest moment in tech in years
The episode closes with a strong reflection from Amjad on the pace of change.
He compares today’s AI moment to the early years of the web—but says this feels even bigger.
His message to founders and builders is clear:
This is a rare platform shift
The pace of progress is exhausting but real
There is huge opportunity for people willing to engage deeply
The builders who understand these tools early will have an edge
Key Takeaways for Founders
AI is collapsing the time from idea to prototype.
The biggest shift may be how quickly teams can go from concept to working software.
Developer productivity is already changing.
AI copilots are not hype alone—they are creating real output gains.
The future of software talent will polarize.
Platform engineers and product-minded builders may benefit the most.
LLMs are powerful, but still unreliable.
There is a big difference between useful generation and production-grade reliability.
AGI is uncertain, but alignment is urgent anyway.
Even narrow AI systems can create serious misalignment and misuse problems.
This is a major platform shift.
Founders who treat AI as foundational rather than optional may be better positioned over the next decade.
Listen to Founders in Arms
Founders in Arms is a podcast for ambitious builders—covering startup strategy, AI, developer tools, fintech, and the realities of scaling companies.
If you’re building in AI, SaaS, or developer infrastructure, this is required listening.
🎙 Subscribe to Founders in Arms on your favorite platform.
💬 Join the conversation at TribeChat.com.
🚀 Discover more insights from top founders and operators shaping the future of tech.
