The Tech Blog

In an older post we looked at how SynthID hides a watermark inside AI-generated images. Today, let us look at text watermarking and how it works. This week Anthropic says Claude models will use a version of Google DeepMind’s SynthID-Text for all text it generates.

So an LLM writes one token at a time. At each step it usually has several reasonable choices. For eg. “The regulator stayed stable even when the load suddenly _” Here changed, increased or jumped can all fit the line.

SynthID-Text uses this freedom to leave a statistical signature. A secret key, together with the recent text, generates hidden scores for possible next tokens. The sampler chooses from sensible options, and over many steps those choices become correlated with the secret pattern. These scores change with context, so there is no fixed list of “watermark words”. The same token can support the watermark in one sentence and not in another.

A detector with the matching key can later recreate the expected scores and check whether the chosen tokens match that pattern more often than chance. One sentence gives little evidence. Longer passages give more choices to test, so confidence improves. This can be added to many LLMs at the sampling stage without retraining. Each provider can use its own key and configuration. Claude’s detector would therefore look for Claude’s pattern, while another provider may use a different key or method.

The method works best when the model has freedom in wording. Creative writing offers many choices. Short factual answers and code don’t have that many options. Heavy rewriting can also weaken the signal, which is why watermark stealing(deducing patterns from text) is an active research area.

If widely adopted, text watermarking could make AI involvement easier to verify later. I am assuming this will be heavily used in academia and school work. A watermark can indicate that a compatible model likely contributed to the text, without identifying the person or chat behind it.

A few months ago around CES, I wrote about Donut Lab’s “all-solid-state” battery. I was skeptical then, but wondered whether some surface-storage or capacitor-like mechanism could explain its claims. A lot has happened since then. So I owe it to whoever reads my posts an update on that.

Donut commissioned Finland’s VTT institute to test its cells. The results show there is a real rechargeable battery here. It held charge for 10 days, so the super capacitor theory looks unlikely. VTT also measured 0–80% charging in about 4.5 minutes at 11C, although a full charge took roughly 7–8 minutes.

Then Ryan@Ziroth did a detailed investigation with more than 20 battery experts. Their strongest clue is in Donut’s own VTT data. The voltage curve looks similar to high-nickel lithium-ion, while the measured cell expansion has a distinctive shape associated with lithium entering a graphite anode(Check Pics). I think the evidence that the tested cell is lithium-based is now quite strong.

The bigger issue is what VTT has not verified. Donut still claims 400Wh/kg and a design life of 100,000 cycles. VTT has independently measured neither. Thats where the problem lies.

And 400 Wh/kg alone is no longer a science fiction. Amprius Tech publishes silicon-anode lithium-ion cells around 400–450 Wh/kg(Will deep dive on this some other day). Those cells make different compromises on power and cycle life, but they show why energy density alone cannot prove exotic chemistry.

Donut Lab rejects Ziroth’s conclusions and says it stands behind its technology. For me, I think consensus is shifting to Donut’s claims being borderline false, on a good number of items. They are still not releasing a product or any serious data to back it up. It is getting close to vaporware territory. Will see how this goes ahead.

Battery tech in EVs is moving very quickly though. Even mature chemistries like LFP are being pushed to limits by folks at BYD. That’s worth looking into soon as well if there is interest.