Medium

🌐 US-Proxy

>

Marco Kotrotsos

In

Autocomplete. Real World AI

by

Marco Kotrotsos

·

2d ago

Synthetic Training Data Is Easy. Knowing It Isn’t Garbage Is Hard.

Start with a failure you understand, then build examples whose success can actually be checked.

Synthetic Training Data Is Easy. Knowing It Isn’t Garbage Is Hard.

Marco Kotrotsos

·

5d ago

Your Agent’s Memory Can Keep an Attack Alive

A poisoned memory can outlast the conversation that created it. Here is how to test the full recovery path.

Your Agent’s Memory Can Keep an Attack Alive

Marco Kotrotsos

·

6d ago

Stop Debugging the Prompt. Replay the Agent’s Bad Decisions.

Find the first consequential mistake before rewriting instructions that may have worked perfectly.

Stop Debugging the Prompt. Replay the Agent’s Bad Decisions.

Marco Kotrotsos

·

Oct 10

Someone Built an AI Torture Chamber. What Did It Actually Prove?

The viral project, the backlash, and the research revision that complicates the story.

Someone Built an AI Torture Chamber. What Did It Actually Prove?

Marco Kotrotsos

·

Oct 10

Please Stop Bullying the Autocomplete

Anthropic’s new cruelty rule sounds like a punchline. The reasoning behind it deserves a closer look.

Please Stop Bullying the Autocomplete

Marco Kotrotsos

·

Oct 6

Your private repository has a public screenshot problem

What PixelLeak reveals about coding agents, review artifacts and the permissions between them.

Your private repository has a public screenshot problem

Marco Kotrotsos

·

Oct 4

No, that local decision model is not better than Jev

Our local routing demo worked. Then we checked it against the evidence in real documents.

No, that local decision model is not better than Jev

Marco Kotrotsos

·

Oct 1

We Built an AI Incident Register. Now You Can Audit Its History.

How we built Agent Did What? with Codex, and why every saved account needs a date, its own evidence, and a way back.

We Built an AI Incident Register. Now You Can Audit Its History.

Marco Kotrotsos

·

Sep 30

Your Million-Token Context Window Comes With a Pricing Cliff

How 1.1% more input can nearly double a GPT-6.1 Sol request cost, and what to measure before changing your agent.

Your Million-Token Context Window Comes With a Pricing Cliff

Marco Kotrotsos

·

Sep 30

Your AI Agent Did What?

Introducing agentdidwhat.com: what happened, who operated the agent, and what the sources actually establish.

Your AI Agent Did What?

Marco Kotrotsos

Marco Kotrotsos

Tech person. I write about technology, Generative AI, design and development. Deeper AND broader at acdigest.substack.com. Also: https://gloss.run

Following

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech