You're arguing against a position I've never taken.

At least on that we can agree 🙂

Yeah… LLMs don't assign truth values… they learn statistical correlations from massive text data. “Patterns of meaning” emerge from co-occurrence and context. Transformers use an attention mechanism to model these complex relationships between words, creating a structured (but not grounded) output.

On your first point... absolutely agree there needs to be a wider understanding of the limitations of LLMs, but that should involve an accurate and up-to-date explanation of how they work (and it also doesn't change the fact they can be incredibly useful). No arguments at all on your other points. 🙂

I'm not entirely sure what you're asking, but the simple answer is: because that's how they're trained. 🙂

That's an oversimplification... we're well beyond that now. While pre-training involves large-scale web data (which can include a mix of "good" and "bad" info), the data is filtered and weighed, and post-training reinforcement learning and fine-tuning pushes models towards more accurate response.

Close, but an oversimplification. LLMs aren't trained on random data... they model factual knowledge and can retrieve it with a high (and continually improving) degree of accuracy. It's true they lack actual "understanding", but they match patterns of meaning, not just words.

I had planned to bling my bike up, but typically for me I left it too late 😢.

But I'll be wearing this t-shirt!

Hey what about Brisbane @astrokatie.com! 🙂

Damn... so close 🙂

Went hunting... this is the photo I was thinking of, taken by ISS astronaut Alexander Gerst... he's tagged it as the Amazon, but it's actually the Juruá...

Ah, yep, familiar with Procreate. Nice!

IIRC, the second photo is the Juruá River, a tributary of the Amazon River? 🙂

That's pretty neat. How'd you create it? Programmatically, or by hand?

So... that'd be a no, then? 🙂

I was happy to help Filippo out and annotate the beautiful still of his image (non-annotated version sent to @apod.shinyakato.dev as I really think it is worthy).

You could do so much science communication from this one image:

- galaxies (plural!)
- aurora
- stellar evolution
- nearest system

🔭🧪

Even without the stunning colour mapping used in the Hubble or James Webb images, the Pillars are still bloody awesome 🙂. Thanks for sharing!

Hey send enough and surely at least some of them won't blow up 🙂