Yup

This is an example of the "lethal trifecta" a fundamental AI security vulnerability that occurs anytime you have the following 3 things in an LLM based system:

1. Access to private data

2. Exposure to untrusted content

3. The ability to communicate externally

Yes, the tricolor certainly was explicitly designed to be inclusive of both communities, peace between orange and green

But what matters is the perception

After more than a century of association with just one of the traditions on the island, it is not perceived as inclusive by a large minority

Indeed, it is all too easy to slip into anthropomorphization when describing behaviors of LLM systems

What would be a better term?

Affirmation Bias? User-Belief Alignment?

Place names in Ireland with silent gh

Dun Laoghaire /dun leery/
Youghal /yawl/
Oughterard /uw terard/

Yep this is a real UI problem

And it's hard to manage, because ideally you need the name of every language in every language

So if you have a 30 language you support, a language picker will need 900 names of languages

No argument from me on that

I too wish them well, particularly as I'm an EU citizen, who favors democratically-controlled regulation of tech

Of course they can and should use whatever approach they deem appropriate (I just think they would be more likely to succeed with an incremental approach)

it is unfortunate that the dominant UI for LLMs has been the chatbot interface with system prompts simulating personality

But there's hope

The chatbot UI has the drawback of a lack of affordance (knowing what you can ask), so my hope is an evolution to more visually structured UIs on top of LLMs

Their approach seems different, with analogous end goals

EuroSky seems to plan to create a whole new infrastructure from scratch, not using BlueSky infra

They really should consider instead Blacksky's successful more incremental approach of starting with just feeds and moderation

Oh, I totally agree

A chatbot answer about itself cannot meaningfully be called introspection, and is very unlikely to be an accurate

If it's using the LLM model it can be regurgitated sci-fi from its training material

If it's using a safety system it's just marketing material from the company

Yes this is generally true, LLMs cannot introspect. They just know about themselves from what's in the training materials and system prompt

Though a chatbot will possibly have a safety system in place that detects "introspective" questions and triggers a special response, possibly a canned answer

You could do it with GitHub pages.

1. Create a repo
2. Create and edit the files for your web site using GitHub's web UI
3. Configure the repo at Settings -> Pages to serve your web site at <username>.github.io/<repo>

You can also set up a custom domain

Or instead of #3 use Netlify to serve

Compressing the entire internet into interpolable vector space that can be queried in natural language remains actual sorcery that we skipped past 2 seconds because we immediately wanted to know if it could write f'in poetry.

bsky.app/profile/prof...