Post

Yeesh. LLMs do nor produce coherence, meaning, quality, etc. They generate probabilities shaped like words.

Well, how do you do it?

'probabilistic word flavoring...' basically defines AI writing.

the detector code may be open sourced but SynthID relies on a private key

Well I'm pleased to see 100% agreement among ppl I respect that most the crits of Anthropic s watermark are either hating or bollocks

The false belief in the piece seems to boil down to “any time an AI system does anything other than select the most probable next word, it is performing worse than it otherwise would be.” This is demonstrably untrue: it exactly describes running a system with the “temperature” set to 1.

Hm, point.

I think his view boils down to (1) he himself is acutely, existentially sensitive to that 0.1% difference; and (2) EU bureaucrats are know-nothing barbarians, wholly ignorant of the sacred mysteries of User Experience, motivated mainly by jealousy of American technological supremacy.

Here’s the nut graf: www.nature.com/articles/s4158…

If you look at the examples in Google’s Nature paper, their examples boosts the 4th choice, “durian”, to the same likelihood as the 2nd the 3rd choices, “lychee” and “papaya”, even though the rankings from the LLM probability distribution (PLM) were: mango 0.50 lychee 0.30 papaya .015 durian 0.05

If it were changing from 40% to 40.1%, yes, we readers couldn’t tell the difference. But there’d be no embedded signal to detect to identify the prose as generated by a specific model. The percentage changes must be more significant than that.

I am perplexed because he is clearly a talented programmer. Hasn't he ever used an RNG? It's all a function of the seed used. "choices are still made at random, but the source of the randomness is different". if you have the seed, the randomness is replicable

By my reading, Google's SynthID paper makes clear that if they limit themselves to applying the watermark to choices that are truly equivalent (per the model's PLM function), the watermark can't reliably be detected.

But the choices aren't from unranked lists. They're from ranked lists, and the only way to leave a signal strong enough to detect is to promote lower-ranked choices. The more subtle the promotion of lower-ranked choices, the weaker the fingerprint. Stronger signal necessarily means worse choices.

The same energy as buying gold plated audio cables when no one actually hears the difference.

Very interesting. Thanks for posting.

What!? Gruber not understanding a subject and writing a poor bad faith take about it? Shocked!

If you are trying to check for those invisible watermarks or hidden characters, this tool helps detect them: https://vinkius.com/mcp/unicode-normalization-and-homoglyph-detector

My read: he's not focussed on AI tech details (to a fault). Main energy is around AI altering meaning in order to watermark, which as a good writer really offends him. Maybe a bit of a purist's worry given the vast majority of terrible writing people are just fine putting out there professionally.

Tks. V helpful. A minefield to grasp.

What's unacceptable is using that shit at all.

Sourceless machine text is only a catastrophe for a culture of automatons. For the serious learner it is revealing: nothing left to trust but the work. Do you have a taste for truth? Do you know how to develop trust? Do you know how to test and verify? Do you know the limits of your knowledge?

Yes. Without understanding much about the EU ...

... and the only way he can manage to square the circle is "EU regulators are morons" I'd argue the current cookie solution isn't workable or working. Most people don't give informed consent because who can with 275 different cookies. Regulation should have mandated a browser/system based system