Matt's Homepage

JustPaid is an AI-powered billing and payment processes financial platform startup for businesses. What makes them more interesting is that besides the founders, the company was developed completely by AI agents.

What has brought them to the news (The Wall Street Journal) is that they use OpenClaw with Claude Code to create an entire software engineering team with 7 fully autonomous AI agents. The company’s human employees focus on tasks like customer requests.

The company also says it hired an employee who was trained almost entirely by the AI agents. They have just nine employees, three of whom are cofounders. The AI bots nearly equal the humans.

Pinnaka said he was spending $4,000 a week to run Claude Code and OpenClaw when he first started experimenting with them. He now claims to have brought the bill down to $10,000 to $15,000 per a month.

None of this is confirmed, but it’s highly likely there is a lot of truth to the claims because other companies are starting to claim to be doing the same thing.

Having big context windows is critical for LLM’s, but fitting a decent sized context on a GPU is rough work. Even when you do get a decent budget, the reality is that you’ll be running out of context more than a few times in an 8 hour session with a 120,000 or even 250,000 token context window.

It’s worse than that since research seems to indicate contexts start to degrade and lose coherence after you use up about 60% of your context budget. You can try /compress commands, but eventually things degrade enough you have to close and restart – which means you lose all that great context and have to retrain the LLM.

There’s a couple ways to fight this.

On Claude, you can persist sessions to external storage and resume sessions.

On Copilot CLI, the memory already exists. Copilot CLI writes every session to a local SQLite database — ~/.copilot/session-store.db. Desi Villanueva wrote a simple python script that will re-load that context if you use Copilot. Handy if you use it for Visual Studio Code.

Researchers at UC Irvine have a more updated Adversarial Machine Learning attack that appears to work on drones.

By creating an umbrella with a specially painted top, the researchers were able to confuse the AI tracking and navigation of a DJI Mini 4 Pro drone and cause it to crash. The method is called the FlyTrap attack and is a visual pattern that performs a next-gen physical distance pulling (PDP) attack that works across multiple angles, even in motion in real settings.

The printed visual draws victim drones closer as its neural network tracking systems interpret the pattern to be the subject moving further away. As the drone approaches the umbrella, the pattern causes the targeting bounding box to continue shrinking – so the drone moves to get closer. Autonomous drones lured by the pattern can then easily be ensnared using a net gun, or further induced to crash to Earth.

None of this would stop a human-controlled drone of course (like we see in Ukraine). But as drone warfare becomes more and more autonomous – it would take down fully autonomous drones. Perhaps modern camo clothing needs to become anti-AI patterned more than traditional human visual camouflage. Or perhaps screens of this kind of camo can be used to protect bases or encampments.