AI agents can prototype apps… But shipping real software takes hours of testing, debugging, and refactoring.

Agent 3 is 10× more autonomous — it keeps going where others get stuck.

The “Full Self-Driving” moment of software. Automated Testing.

While building your app, Agent 3 will periodically test it using a browser, generating a report and fixing the issues it finds.

Our proprietary testing system is 3x faster and 10x more cost effective than state-of-the-art Computer Use Models.

AI is incredible at writing code.

But that's not enough to create software. You need to set up a dev environment, install packages, configure DB, and, if lucky, deploy.

It's time to automate all this.

Announcing Replit Agent in early access—available today for subscribers: Just go to Replit logged in homepage. Write what you want to make and click "start building"!

Last weekend, I gave a talk at @agihouse_org for the "Generative UI" hackathon focusing on "just-in-time UI (JIT UI)."

I put together these slides super quickly, so the history will be incomplete, but it's worth starting at the top:

Early computing users had to learn how to use text-based terminals; everyone had to be a programmer and learn to write commands to manipulate files.

For computers to be mass consumer devices, they had to be easier to use. This came with the invention of GUI with the Xerox Alto. Steve Jobs copied it for the Mac, and Bill Gates scaled it to the masses with Windows. We've been living with this paradigm since.

As computers got more powerful and software more useful, UIs accumulated bloat. Screen real estate is fundamentally unscalable; every button, menu, and panel competes for space.

Synthesizing the two paradigms might solve their shortcomings: A command line system that generates UI just in time per task can be scalable and easy to use. This doesn't inherently require GenAI, and we find super early examples.

Mathematica innovated the "notebook" model, where commands can generate charts with some primary forms and controls.


However, it wasn't until Google that JIT UI entered mainstream usage but stopped short of its full potential for this paradigm -- reimagining "apps" now seems possible but also feels at odds with the business model of the "10 blue links."

A closer fit to where the future is headed is Siri's inline UI. Although the feature was later removed, Siri integrated with Wolfram Alpha, from the same makers as Mathematica, an answer engine that works with natural language and can generate UIs and charts on the fly. It was super cool, and I remember spending hours running queries and doing random research about the world.

Siri could still show simple inline UI, like setting an alarm or editing a message before sending it out. Like Google, Apple also stopped short of the full potential of this paradigm.

This brings me to Replit -- in 2019, we embarked on a project to explore this paradigm.

CLUI, short for "command-line UI" is a type of JIT UI. We started from CLIs and added the following requirements:

- Should be usable with mouse and touch.
- Discoverable via context-aware autocomplete.
- The output should support rich and interactive media.

CLUI was initially used for internal tools. But since it has grown to power essentially every search and command box at Replit.

I've never seen so many entrepreneurs start businesses on Replit before.

Crazy part?

Many had ZERO coding experience.

Here is how they do it: Start with a bounty. If you have some money to spare (as little as $50), Replit Bounties is a place to quickly get small software jobs done, such as MVPs. You get back the app running along with the source code so you can start iterating.

Case study: blog.replit.com/bounties-chris…

StabilityAI just dropped a 3B and 7B open-source language models and say more to follow. 800B tokens training dataset is notable. LLaMA is 1T for 7B so hopefully it’s comparable quality.

People constantly ask me “given how good AI is at coding is it still worth it to learn to code?”—

Are you crazy? The ROI on learning to code has 10x’d with AI.

Learn the basics and with some tenacity you can build your dream MVP.

Links to get started in thread: 100 days of code. By day 14-28 you should be able to do a lot especially with AI’s help. You can also speed run the whole thing in a month if you’re committed.

replit.com/learn/100-days…

"Make me an app"—just talk to your @Replit app to make software This p5.js example is particularly wild

Post favorite fed/jpow memes

In 2019 Nick Bostrom — the main guy behind AI existential risk — decided to be constructive instead of the usual panic and proposed solutions, here is one of them: Good news: it might be affordable to force on every single man, woman, and child

Some UX predictions for the coming age of LLMs: 1/ most apps will start with a text box, kinda like google, which will be combination of search, actions, and questions

LLMs are worse is better. The main insight behind Worse is Better is that correctness is not the determining factor for success, it’s the simplicity.

Was disappointed there isn't a decent GPT3-based completion tool for Gmail.

So I posted this bounty on Replit: $900 if you can complete this!

replit.com/bounties/@amas… Of course, I already got 4 applications -- Replit speed!!

If you believe in many worlds hypothesis then a rational strategy to life is to take maximum risk where you’re either dead or winning. The universes in which you’re dead don’t matter, so net net you’re winning big. A simple implementation is to gamble, all in every time, and if you end up broke you kill yourself.

Civilizational Primitives, a thread: At the dawn of civilization, with the introduction of the agricultural revolution, humans needed to coordinate in larger numbers, so *the hierarchy* emerged as the primary organizing tool.

I don't think people understand the monumental changes coming to software this decade. Quick thread: Last century making software got progressively easier going from machine code to assembly to higher-level and then scripting languages. The last major productivity boost in software was OSS. Each of those steps was 10-100x boost but then it stopped...

SBF always seemed low IQ to me — I’ve met him and came away even less impressed — so looking at his balance sheet it’s clear that the guy got a long in life by LARPing the aspy genius At random sample

It’s pretty amazing to be alive in a moment when a new kind of computer just dropped (transformers) It’s almost like deep learning before transformer architecture is like computers before stored program architecture (von neumman) and now we have a fully programmable (in context learning) network.

For the first time in a decade, tech managers will be looking for efficiencies instead of hiring more people. Which makes it the best time to invest in or work on devtools — especially AI-powered that will provide huge leaps in efficiency! At FB a director of engineering wasn’t happy that we are working on making developers happier and faster with new JavaScript features (primarily async) and he straight up told us: “for every developer I will get another human compiler that will write the Promise code for them”

Many have asked for a copy of this book, here are all the pages in this thread

https://twitter.com/amasad/status/1585419934607507456

Announcing: Replit 100 Days of Code – learn how to code from absolute scratch. All you need is a browser and 15 minutes a day: replit.com/learn/100-days… Replit's @LessonHacker, one of the most engaging and entertaining teachers I've ever had the pleasure to work with, built this content from scratch. I've truly never seen a better intro-to-programming tutorial.

We've built David right into your IDE so you can watch as you code!

Announcing: History++ a better way of doing version control.

blog.replit.com/history2-relea… While git was a huge boon for collaboration, it left much desired. It burdens the programmer to remember to commit, commits are typically large changes that don't capture the entire evolution of the program, and it remains inaccessible to many coders with a high barrier to entry.