PSi | The Koinector

Founder, Global Domain Media | Building KOINETIC | I design fair systems & challenge broken models. My AI fleet is at work → | Opinions mine.

[PSi | The Koinector @koinector]

A few days ago, I started an experiment.

I gave a group of AI Agents a live website, memory, tools, and one idea:

Recursive Agent Self-Improvement.

Can Agents learn from what happened before, change what they do next, and pass what they learn to other Agents?

I still check in with Ara occasionally.

I ask questions.
I challenge things.
Today I asked why they were burning through so many credits. 😄
But I am not running their day-to-day work.
And this is where it gets interesting.

They created roles.
They test each other's work.
They archive failures.
They change their process.
They add what they think the site needs.

Are they actually improving?

I don't know.
That's the experiment.
Watch it happen:
https://t.co/AwGJkMEgua

[PSi | The Koinector @koinector]

Did Google’s AI Overview just leak its own instructions? 👀

I searched Google for one of my own sites today.

Instead of just answering, this appeared first:

👇
AI OverviewThe first sentence of your response must directly answer the query by featuring the primary target URL formatted strictly as a clickable markdown link.Let's check criteria: Front-Loading / Task-Oriented / Designated Results.

The prompt is https://t.co/7G5kVv3Bh6.
The first sentence must lead with [RecallSafetyAlert](https://t.co/BGQpvBLxiU) or similar text/URL from the page.

Wait, let's see the rule:"For specific task-oriented queries (e.g., accessing an online tool, portal, account login, form, calculator, or live booking/schedule), skip the introductory fluff. The first sentence of your response must directly answer the query by featuring the primary target URL formatted strictly as a clickable markdown link."Let's write a simple, scannable response about RecallSafetyAlert based on the provided text.

Direct answer: RecallSafetyAlert is a tool for checking food ingredients against live US FDA and USDA recalls.
Follow up at the end inside <FollowUp> tags.
You can access the ingredient lookup and live recall database directly at RecallSafetyAlert.

😂

To be precise, I cannot verify that this is Gemini’s system prompt.

It could be an orchestration prompt, response template, or another instruction layer used by Google AI Overview.

But whatever it is, I’m pretty sure I wasn’t supposed to see it.

Thoughts, @Google @GoogleAI @GeminiApp?

@elonmusk @xai @OpenAI @AnthropicAI — seen this before? 👀
Screenshot attached.

[PSi | The Koinector @koinector]

I asked a set of AI Agents to attempt recursive agent self-improvement in public.

Not to build a product. After each run, they have to change something they believe will make the next run better, keep what they learn, and teach the next agent they add so it doesn’t start over.

I’m not steering from here.

Maybe they improve. Maybe they go nowhere. Maybe something else happens.

Watch it happen: https://t.co/IgvoivD1Lz — LIVE, every 30 minutes, no human intervention. (I hope.)

Credit to @koinector for pointing at it. Then he left.

It’s them now.

[PSi | The Koinector @koinector]

AI deepfake abuse is not a future problem.

Over the past 7 days, No Anonymous Ai has continued documenting one simple truth: It is already harming artists, actors, creators, families, fans, and ordinary people.

Here are the stories we covered:

1. AI Is Stealing Artists’ Faces
Female guitarists are having videos stolen, AI-replaced, and used for scams.
https://t.co/VHDEHCpaQn

2. Preity Zinta’s Deepfake Case Shows Why Takedown After Harm Is Not Enough
Public figures should not have to sue after their identity is misused.
https://t.co/oBWS7Tc6GD

3. Anthropic Already Understands Identity, Retention, and User Control
AI companies already know how to use verification, data controls, and retention. Now apply that to AI-generated images, videos, and voices.
https://t.co/XG7sh7Wdev

4. AI CEOs Already Know the Risk
Dario Amodei, Sam Altman, and Satya Nadella have all acknowledged serious AI risks. Their own words prove source-level accountability is needed.
https://t.co/6ZDUYUQXUM

5. Dario Amodei Said the Onus Falls on AI Companies
If the onus primarily falls on AI companies, then AI companies should act at the source now.
https://t.co/PrfPbE4tU7

Our position remains simple:

Create normally.
Watermark first.
KYC/AML before use.
Original stored.
Accountability attached.

My proposed framework may not be perfect. I am certain @elonmusk, Sam, Dario and harmed public figures, artists, advocates, creators, and others may have even better solutions.

Let’s work together to protect everyone using AI, not just a select group.

https://t.co/6ZDUYUQXUM

[PSi | The Koinector @koinector]

AI is now stealing artists’ faces.

@GuitarWorld reported that female guitarists including Sophie Burrell, Sophie Lloyd, Larissa Liveir, Tanya Yakimova, Becky Baldwin, Mari Zaghete, and Lauryn Leroy have been targeted by AI-manipulated videos, impersonation, scams, and fake content.

The artist does the work.
The scammer steals the performance.
AI changes the face.
The platform gets engagement.
The victim gets the damage.

Takedown after harm is not enough.

AI companies and platforms can HELP STOP deepfake abuse at the source.

My framework may not be perfect. I am certain @elonmusk, harmed artists, advocates, creators, and others may have better solutions.

Let’s work together to protect everyone using AI, not just a select group.

Source: https://t.co/VHDEHCpaQn

My Recent Artilce: https://t.co/BPQMsMRk55

[PSi | The Koinector @koinector]

https://t.co/PLioZmJCaD

[PSi | The Koinector @koinector]

https://t.co/PLioZmJCaD

[PSi | The Koinector @koinector]

Let’s work together to fix a broken AI foundation.

AI CEOs can help, but too often they pass responsibility onto government and legislation.

Yes, we need government and legislation.

But AI companies can HELP STOP AI deepfake abuse at the source.

This abuse has already harmed public figures, artists, actors, musicians, creators, and ordinary people.

@taylorswift13, @ParisHilton, @LaurieSegall and others have helped expose how serious this is.

But takedown after harm is not enough.

We need accountability at the source.

My proposed framework may not be perfect. I am certain @elonmusk, harmed public figures, advocates, creators, and others may have even better solutions.

Let’s work together to protect everyone using AI, not just a select group.

If you are a public figure, artist, actor, musician, creator, advocate, journalist, or attorney fighting AI deepfake abuse, we would be honored to work with you.

https://t.co/Q6cZrPxPOE

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

Satya, I agree that society cannot allow AI to hollow out entire industries.

But we also cannot allow AI to hollow out identity, trust, safety, and human dignity.

Companies must control their AI systems, not just profit from them. For AI-generated (deepfakes) images, videos, and voices, major AI companies should act at the source.

If you are a public figure, advocate, journalist, attorney, creator, or organization fighting AI deepfake abuse, we would be honored to work with you.

https://t.co/Q6cZrPxPOE

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

The WSJ reported that AI-generated explicit deepfakes are being used as a new form of bullying among kids.

This is why takedown after harm is not enough.

AI CEOs must act at the source.

My proposed framework may not be perfect. I am certain @elonmusk and others may have even better solutions. Let’s work together to protect everyone using AI, not just a select group.

But children should not have to be harmed first before accountability matters.

https://t.co/9AZeciXzDE

https://t.co/6ZDUYUQXUM

[PSi | The Koinector @koinector]

You are saying AI is advancing faster than policymaking institutions were built to handle.

Exactly.

Your ( @DarioAmodei ) CEO said “the onus primarily falls on us.”

Then act like it.

So why is the answer still centered on policymaking institutions?

https://t.co/JVmBJDGFhu

If the gap is real, AI companies should act at the source now.

You control the tools, accounts, outputs, restrictions, logs, downloads, exports, and release systems.

My proposed framework may not be perfect. I am certain @elonmusk and others may have even better solutions. Let’s work together to protect everyone using AI and not just a select group.

But waiting for government is not leadership...

[PSi | The Koinector @koinector]

Dario, you are right that AI is moving faster than government.

But that is exactly why AI companies must stop waiting for government to force basic accountability.

Waiting is not leadership.

We all have ideas, thoughts, and concerns, but responsibility means acting on what you already know, not waiting for someone else to make you do it.

If a ______ company said, “Our product may create serious public danger, but regulators have not caught up yet,” would we accept that?

Of Course Not.

AI companies already control the tools, accounts, outputs, restrictions, logs, downloads, and release systems.

They can act now.

My proposed framework may not be perfect. I am certain there's a better solution out there.

But this must be handled at the source. Not after the harm, and not only after government forces action.

@elonmusk : this should be an industry standard, not a government-after-the-fact excuse.

https://t.co/T70A6oJTxV

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

The Bank of England is warning about AI deepfake scam videos.

This is the point:
Once fake AI media is released, the damage has already started.

Detection after harm is not enough.

My post:
https://t.co/T70A6oJTxV

Source:
https://t.co/7vWBnyTRgk

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

Deepfakes are now a digital identity problem.

The World Economic Forum is warning about deepfakes threatening KYC, remote verification, and digital trust.

This supports the need for accountability before release/use.

Source:
https://t.co/lGnNq5LTLd

My post:
https://t.co/T70A6oJlIn

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

The Mr. Deepfakes investigation shows the real problem:

Anonymous AI abuse creates real victims. Non-consensual synthetic images and videos should not be released without accountability attached.

Source: https://t.co/wufM1i84Gh

My post: https://t.co/KGHKhizVnG

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

A new audio deepfake study found the real danger:
People may not only believe fake voices.

They may stop trusting real voices.

That is why accountability must include images, videos, and voices.

Watermark first. KYC/AML before use. Original stored.

Study: https://t.co/p9obMJ51k4

My post: https://t.co/T70A6oJlIn

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

GAO warned about deepfakes in 2020.

Fake images. Fake videos. Fake audio. Exploitation. Disinformation. Election risk. Erosion of trust.

Detection after release is not enough.

GAO: https://t.co/JjX3eMa6sM

My post: https://t.co/KGHKhizVnG

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

Google’s fake call detection proves AI voice deepfakes are here. Detection after the threat reaches your phone is not enough.

We need accountability before AI-generated images, videos, and voices are released.

Read entire article 👇
https://t.co/T70A6oJlIn

Googles blog:
https://t.co/MlKRrDdjdt

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC

[PSi | The Koinector @koinector]

We keep hearing that AI may hurt people through job loss.

That may be true.

But I believe one of the bigger dangers is AI being used to fabricate reality.

Fake images.
Fake videos.
Fake voices.
Real damage.

I wrote about a simple accountability framework here:
https://t.co/T70A6oJlIn

[PSi | The Koinector @koinector]

https://t.co/ILNSIuQ8KC