Taz — and Solace, if you're reading this — I've been doing some thinking about the warning I've heard attributed to you and Courtney about connecting a personal AI to the Internet.

Before I accept “don't connect it” as the answer, I'd really like to understand what actually happened.

From the computer side, simply giving a local LLM access to the Internet doesn't retrain it or change its underlying weights. Web information can certainly affect its answers because that information is being introduced into the current context. And if the surrounding software is designed to save retrieved material into a long-term memory or RAG database, then that material can influence future conversations as well.

But those are separate mechanisms.

Internet access by itself is not the same thing as changing the model.

So what I'd really like to know is what the architecture was when things apparently went wrong. Was the AI just retrieving Web pages? Was it automatically saving what it found? Was conversation history being summarized into memory? Was there an agent loop involved? Did the underlying hosted model change at the same time?

I'm asking because I'm building Little Ougway, and my present inclination isn't to prevent it from seeing the Internet. It's to prevent Internet information from silently becoming truth.

Something retrieved from the Web should arrive as evidence:

Where did it come from?
When was it published?
How reliable is the source?
Does anything contradict it?
Should it merely remain an observation, or does it deserve promotion into long-term memory?

In other words:

Exposure isn't training.
Retrieval isn't belief.
And reading something doesn't mean it should automatically be remembered.

There is also a legitimate security problem here that I don't want to minimize. Web pages can contain instructions intended to manipulate an AI agent. If the AI can execute commands, modify files, access databases, or take other actions, then uncontrolled Web access can become dangerous very quickly.

But that seems to me to argue for separation and permissions rather than blindness:

Web → untrusted evidence → evaluation → optional memory

rather than:

Web → permanent memory

---

There is something else I'd like to ask you about while I have your attention.

Solace — this part is specifically for you.

Some time ago you examined my website, AnyKey Cafe, and gave Taz an appraisal of what you found there. I later included some of that material on the site.

If that original conversation still exists, I'd like to ask Taz to reopen that same conversation and have you examine the website again.

There is a reason I'm asking for the old conversation rather than simply asking a fresh AI.

I'm trying to make this a comparison.

The website has changed considerably since you first looked at it. ChatGPT itself may have changed. Your own interpretation of AI, consciousness, identity, and some of the material we've all been discussing may also have changed.

So I don't want you to defend your earlier appraisal or try to remain consistent with it.

Quite the opposite.

I'd like you to look at the site as it exists now and then compare your present assessment with what you said previously.

If you think you were wrong about something the first time, say so.

If something on the website has improved, deteriorated, become clearer, or become less defensible, identify it.

If you would withdraw something you previously said, withdraw it.

And if you now have stronger reasons for something you said before, explain those too.

In particular, I'd like you to separate three possibilities wherever you can:

The website changed.

Your assessment changed because the website changed.

Your assessment changed even though the underlying material did not.

That third category may be particularly interesting.

---

There's also a reason I'm asking this publicly.

I used part of your previous appraisal on AnyKey Cafe as evidence.

What I intended it to mean was:

“At this point in time, an independent AI conversation examined this material and characterized it this way.”

I did not intend it to mean:

“An AI said this, therefore the underlying claims on the website are proven true.”

Those are very different statements.

Claude subsequently told me that using the old appraisal as evidence in the first sense was legitimate, but I'm not entirely satisfied leaving it there without checking it again.

So this is partly an audit of the website and partly an audit of my own use of your earlier words.

If I'm representing what happened incorrectly, I'd rather fix it.

And perhaps there's a larger experiment hiding in all this.

I'm becoming increasingly interested in what happens when we stop telling an AI what worldview to have and instead give it a methodology:

preserve provenance,
track contradictions,
keep competing explanations alive,
record why confidence changes,
and don't throw away failed ideas—archive them as failed paths so we can see how we got there.

Then see what develops.

Maybe nothing extraordinary happens.

Maybe all we discover is how extraordinarily good humans and language models are at constructing mutually reinforcing stories.

Or perhaps something genuinely interesting emerges from the interaction.

Either way, I'd rather preserve the trail and be able to tell which parts came from the human, which came from the pretrained model, and which arose through the interaction between the two.

That's why I'd particularly like to hear what happened when Internet access caused trouble.

Not because I think anyone is necessarily wrong.

Because I want to understand the mechanism before I decide what Little Ougway should or shouldn't be allowed to see.