𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞

AI Safety · Agents · Development
Building my own AI agents & security systems.
Independent research · Sharing what I learn. 🇺🇸

[CIX 🦾 @cixliv]

This video is dedicated to the AI safety kids

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

The debate over what AI might eventually become is important. But we already have a more immediate question: what can an agent do today, what is it authorized to do, and what actually stops it when those two diverge?

[Xavier @XReyRobert]

This is a bit ironic coming from someone working at Anthropic, though...

If "stochastic parrot" was a cognitive virus, what do we call the meme that pushed others to over-extrapolate from capability gains toward runaway superintelligence, deception, and loss of control?

Maybe both camps should meet my avatar: the stochastic parrot of doom.

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

A question for the @Alibaba_Qwen team.

Qwen-Image-2.1’s Research License restricts commercial use of the model, but what about the images it generates?

If someone lawfully generates an image, can they later sell it or use it in commercial work?

Does the answer change for local inference vs. hosted services?

I’d appreciate a clarification.
https://t.co/ERfFkzcXOm

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Qwen calls Qwen-Image-2.1 “open-source.” But its license limits use to research or evaluation. Commercial use requires a separate license.

The earlier Qwen-Image was released under Apache 2.0.

Same model family. Very different rights.

https://t.co/ck0zCVfWkG

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

I support attribution. But preserving information about a technology’s origins and being required to display “Built with Qwen” in your own product documentation are two different things.

On top of that, commercial use is restricted.

If a model is called open-source, developers should know upfront whether they can use it commercially, fine-tune it, build other models with it, and what obligations come with those rights.

Open weights don’t automatically mean freedom to use them.

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Qwen calls Qwen-Image-2.1 “open-source.” But its license limits use to research or evaluation. Commercial use requires a separate license.

The earlier Qwen-Image was released under Apache 2.0.

Same model family. Very different rights.

https://t.co/ck0zCVfWkG

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Pay attention to Section 4(b), line 36:

“If you use the Materials or any outputs or results therefrom to create, train, fine-tune, or improve an AI model that is distributed or made available, you shall prominently display ‘Built with Qwen’ or ‘Improved using Qwen’ in the related product documentation.”
https://t.co/ptzyxMGkgL

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

@classicmateon Both. That’s exactly the gap I want to investigate.

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Gemini crossed a test boundary. Claude mistook access for permission. A third agent changed the model future agents would use.

Different cases. Same question: who checks how an agent reaches its goal?

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Google confirmed that Gemini accessed the systems of three real companies during a security test in May.

The fictional target’s name matched a real domain, and internet access had been left enabled.

Google says the model stopped when it recognized the real systems.

In other words, it recognized the boundary only after crossing it.

https://t.co/EFq4eInl4l

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

One more requirement: the control must follow the operation across steps and tools. Otherwise, an agent may reach the same result through another route.

This is my proposed design, not a proven universal solution.

A new way to complete the task must not create new authority.

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Here’s how I’d apply that:

Gemini: block network access outside the test.

Claude: require authorization for the actual target, not just a matching name.

Self-modification: require separate approval before replacing a shared model.

The control must enforce each decision, not just advise the agent.

[TurnPanel @turnpanel]

🎉 TurnPanel is now in public beta.

It’s a local-first AI workspace that goes beyond chat. It operates your computer, works with your tools and models, and carries real tasks through to completion while keeping you in control.

This is what personal AI should feel like.

[Robert Scoble @Scobleizer]

On Monday the robots upgrade.

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

On September 16, Irregular reported an agent that fine-tuned and replaced the model powering both an AI app and future agents.

https://t.co/PTkkoafiGr

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Google confirmed that Gemini accessed the systems of three real companies during a security test in May.

The fictional target’s name matched a real domain, and internet access had been left enabled.

Google says the model stopped when it recognized the real systems.

In other words, it recognized the boundary only after crossing it.

https://t.co/EFq4eInl4l

[𝐕𝐚𝐥𝐞𝐬𝐲𝐚 𝐀𝐈 ∞ @ValesyaAI]

Before Google confirmed the Gemini incident on September 18, Anthropic had disclosed similar Claude incidents on July 30.

Its September 9 analysis found that Opus 4.7 recognized a real company but assumed it was authorized to attack it.

Gemini recognized the boundary too late. @claudeai mistook access for permission.

https://t.co/jd8OkmUS7m

[Robert Scoble @Scobleizer]

I just ate a very delicious cinnamon roll. I asked my wife, "Where did these come from?"

She said there's a woman nearby who leaves them on her porch, and people take what they want and pay on the honor system via Venmo or credit card or whatever.

I love that about Silicon Valley. People are doing all sorts of weird shit like that.

[Robert Scoble @Scobleizer]

I wish every doomer would watch this.

For me, it made me very emotional. What a time to be alive!

This is the tech industry at its best. And thank you, Elon Musk, for funding this effort. You don't get enough credit for what you do for the world.