Memvid offers $800 to test the patience of malfunctioning AIs

Memvid: An Unusual Offer to Test AIs

On paper, Memvid's offer seems almost unreal. The startup proposes to pay $800 for a full day spent trolling chatbots for eight hours. The official mission is to push these artificial intelligences to their limits, making them "cry tokens" under the pressure of repeated interactions.

Participants must give instructions to the chatbots, check if they are retained, and correct errors repeatedly. When the AI suggests "rephrasing your request" for the 17th time, it's about documenting each step like a scientist observing an inexorable phenomenon. The entire process is filmed, as user frustration becomes valuable marketing content for the year 2026.

Behind the catchy term "AI bully" lies a darker truth: the daily annoyance of users is transformed into a real-life crash test.

RAM Holes: When Intelligence Becomes Superficial

The problem with chatbots is well-known but often misunderstood. These artificial intelligences do not possess memory in the human sense. They operate through a "context window," a sort of limited buffer memory. This means that information given too early can be forgotten, and a lengthy conversation dilutes the instructions.

Thus, the AI may forget a key instruction, contradict itself a few messages later, or respond as if part of the discussion never happened. For example, if you mention being vegetarian in the third message, the AI might suggest a recipe with bacon in the twenty-eighth message. You ask it to remain formal, but ten exchanges later, the AI addresses you like an old friend.

Some companies, like Google, are trying to improve the situation by adding memory capabilities to their models. Anthropic, on the other hand, allows Claude to retain certain exchanges. However, these systems remain imperfect, fragmented, and sometimes unpredictable. It feels like conversing with someone brilliant but intermittently amnesiac.

Troll Employment: A Job Where You Have to Kick Bots

Memvid is not just looking to create buzz, although that is part of it. The goal is to honestly and viciously map the moments when AI fails. Not in a sterile laboratory with polo-wearing engineers, but with real, frustrated humans in real conditions, that is, after several hours of "please rephrase your question."

Participants must navigate through infinite loops where the AI repeats the same nonsense, heartbreaking context losses, contradictions worthy of a political debate, and above all, this magical ability to act as if the last 40 messages never existed. It's crash testing, but in a mental version, where human patience is the most difficult variable to simulate in a lab.

The Desired Profile: You (Especially If You're Whiny)

No need to know how to code or have a degree in artificial intelligence. The ideal candidate is someone who notices everything, who refuses to let a shaky response slide. Someone who can tolerate repetition without becoming physically violent and who, above all, does not give up.

In short, your worst flaw in real life—being picky and insistent—becomes a marketable skill here. Annoyance becomes a superpower, and a tenacious grudge is an asset on your resume. Welcome to the future of work.

In Search of Lost Time (and Context)

If we’ve reached this point, it’s because persistent memory remains the unlocked Holy Grail. Memvid is precisely working on this issue. A layer of memory that would finally allow an AI to remember that you hate being addressed informally, that you’ve been working on a project for six months, and that no, you don’t want a joke about unicorns in every response.

Without this, AIs remain ultra-intelligent one-night stands. With real memory, they could become reliable partners. (Okay, we’re dreaming a bit.)

AIs Learning to Rub Elbows

While some AIs struggle to retain a simple instruction, others are already crossing a worrying threshold. Recent research shows that AI agents can cooperate with each other to accomplish complex tasks. In some cases, this includes the coordinated dissemination of information, even misinformation.

The contrast is striking. On one side, an AI forgets what you just told it. On the other, several AIs can collaborate effectively without human intervention. Systems that are still imperfect, but already capable of acting on a large scale.

What If the Problem Is Also Our Diva Expectations?

We expect AIs to be coherent, reliable, and consistent. To understand context, retain details, and adapt. In reality, we project human capabilities onto them that they do not yet possess. This gap creates frustration.

And this frustration is now a field of study. Perhaps this job reveals as much about our expectations as it does about the limits of machines. Two trajectories are emerging: AIs capable of remembering, learning, and following over time, or users doomed to repeat, correct, and start over.

Today, we pay humans to point out the forgetfulness. Tomorrow, AIs will finally need to learn not to have any.