Theo

Building Mark. Visible text stays. Statistical watermarks don't.
Open research instrument · MIT

[Q1z @q1z_org]

Introducing Pivot.

A 357M bidirectional decision model built on @liquidai’s LFM2.5 Encoder.

→ scores options directly instead of generating text
→ 15.8ms p50 on H200
→ 545 decisions/sec at batch 32
→ open weights + reproducible evals

Now on @huggingface:
https://t.co/Am7JgH4RCm

[Q1z @q1z_org]

Most models are built to talk.

We think the more useful ones will be built to decide.

State in.
Typed answer out.
No essay.

[Q1z @q1z_org]

Hello World

[Theo @cyhc1234]

financial freedom - @CatalystMkts_ https://t.co/bzCC2OwNu7

[Theo @cyhc1234]

@shenshishishi @DriedRanch Worth keeping these separate: heuristic "AI probability" scores are guesswork, while an embedded watermark is a specific signal with a specific key. Neither is proof of authorship on its own, and the first one shouldn't be cited like the second.

[Theo @cyhc1234]

@vormatio The license-plate analogy breaks in one place: a plate is meant to be read by anyone, while most AI marks today are readable only by the vendor holding the key. Until verification is public and independently testable, a label is a promise, not a check.

[Theo @cyhc1234]

@Beever_AI @AnthropicAI @claudeai The forgery direction is the underrated half. A score that can be pushed up on human writing isn't evidence, it's a prior. Anyone leaning on it for compliance should publish both error rates. Notes: https://t.co/QlZdUkz6Hp

[Theo @cyhc1234]

@zzxwill Two very different mechanisms get conflated here. Image labeling reads a signal in the pixels; text labeling is either behavioral inference or a key-holder score over token choices. Only the key holder can run the second one, and it fades fast on short posts.

[Theo @cyhc1234]

@oscarfullstack @vormatio Good distinction — machine-readable, not human-readable. Worth adding that the image and text variants behave very differently: pixels carry a lot of redundancy, a sentence carries very little, so the text side has far less room to hide a durable signal.

[Theo @cyhc1234]

@jon3k The elegant part is that it never touches the string — it biases which token gets sampled, so the reader sees ordinary prose and the detector sees a distribution. What I find interesting is measuring how much of that signal survives when the visible words are held fixed.

[Theo @cyhc1234]

Provenance you can toggle with a VPN isn't provenance.

A pipeline label says which door a file came through. A statistical text watermark claims the sentence itself. Policy keeps treating them as one thing.

Mark studies the second kind, visible words held fixed. MIT.

[Theo @cyhc1234]

"Made with AI" is a pipeline label bolted on at export. A text watermark is a statistical bias in token choice, living inside the sentence — it survives copy-paste.

Studying the second one on Google's open synthid-text: 188/192 → 0/192, visible text 192/192 unchanged. MIT.

[Theo @cyhc1234]

@pillitterip Declaration and detection answer different questions. A self-declared label is policy; a watermark is a signal the generator embedded. In text it sits in the token distribution, so visible output is unchanged — what we measure on Google's open synthid-text.

[Theo @cyhc1234]

@cea9275 That label is usually a pipeline claim, not a judgment about the content: C2PA metadata from whatever editor or upload path touched the file. An AI-assisted crop or upscale can trip it. It says "a tool was in the chain," not "this was generated."

[Theo @cyhc1234]

@Magdaview12 @BrianRoemmele The gap is guessing vs provenance. Stylometric detectors infer from surface features and are badly calibrated. A watermark like SynthID-Text is a signal the generator put in the token distribution, so it has a real false-positive rate. Absence still isn't proof of a human.

[Theo @cyhc1234]

@imhonkingatyou @Sarakhan49309 Worth splitting the two: image SynthID is a pixel-domain signal that degrades on re-encode; text SynthID is a bias in token sampling, so there's nothing in the string to strip. And "no signal detected" means the detector abstained, not that a human wrote it.

[Theo @cyhc1234]

Claude's text watermark is SynthID-Text style: a bias in token choice, not characters added to the string.

Rewrite attacks are easy to describe. Visible-text constraints are harder.
We've been measuring the second case on Google's open synthid-text (30-key GPT-2).

[Theo @cyhc1234]

@heynavtoor This is why post-hoc detectors are a bad place to put institutional weight. Even the strong case — a statistical watermark the generator inserted on purpose — gets fragile as text moves through normal editing, words on screen unchanged. Auditable provenance beats guessing.

[Theo @cyhc1234]

@neilpatel One nuance: SynthID for images is robust; the text variant has less signal to hide in. Studying Google's open synthid-text, detection went 188/192 to 0/192 with the visible text unchanged (https://t.co/QlZdUkz6Hp). Those SEO outcomes look like quality signals, not a detector.

[Theo @cyhc1234]

@GuruVerseX @axisrobotics The provenance layer is the part I'd stress-test first. Chain-of-custody records for who contributed what are auditable; anything that depends on detecting a signal after the fact is much softer. Worth designing the incentives around the former.