DentroChat

How much energy does one AI prompt use? We dug through studies: a normal prompt is about 0.3 watt hours, a long document about 2.5. A median DentroChat prompt: 0.2 to 1.3 watt hours. https://dentro.chat/blog/how-much-energy-does-an-ai-prompt-use

For scale: a kettle takes ~110-125 Wh for one litre of boiling water, that's ~90 DentroChat prompts. Driving a car one kilometer equals ~300 DentroChat prompts.

We'd show a real Wh number per message in the product, but most providers don't report energy per request yet.

CO2 doesn't depend on the datacenter but mostly on the grid. On the EU average (210 g/kWh) a normal prompt is 0.06 g. In France or Sweden it is 0.01 g, in Germany 0.1 g, in the US about 0.12 g. Same model, four times the emissions.

Our models are small mixture-of-experts: 6 to 18 billion parameters work per token, against roughly 100 billion for GPT 4o. So a median real DentroChat prompt lands at 0.2 to 1.3 Wh. Reading 41k tokens sounds like a lot, but 30k of them are usually cached.

Various sources show similar numbers: Google states Gemini median at 0.24 Wh, Epoch AI ~0.3 for ChatGPT, other papers and Sam Altman in the same range.

A long document (about 10k tokens, so 7 to 8 pages) costs about 2.5 Wh, and a long reasoning answer about 4 Wh. More tokens read and written means more electricity. Caching helps: if the same chat was read minutes before, the second run is about a tenth of the cost.

We don't often share numbers from DentroChat, but here's a rare look behind the curtain 👀 In the few months of its existence, DentroChat has handled over 30,000 messages across more than 4,500 conversations. That's close to 2 billion tokens processed!

Apart the features DentroChat offers, we believe this is also due to the trustworthiness of DentroChat as a fully European, privacy-first AI assistant. Which is something we work very hard for.

We're a small team, and these numbers are small by Silicon Valley standards (for now 😉). But they're real, they compound weekly, and they tell us we're building something people don't want to give back. If you haven't yet, go try for yourself: https://dentro.chat

2. Conversion Of everyone who starts a trial, 77% end up as paying users. This shows the trial is doing its job of providing a chance to experience the product for an extended time period, which seems to convince many users that it's worth paying for.

Both of these metrics clearly point towards DentroChat being a truely useful assistant – and in the end, this is the single most important metric for us.

That sounds very cool and it shows that there's usage. What we care much more about though, are the two other stats. 1. Stickiness On a typical day, 7 to 8 out of every 10 paying users open DentroChat again. Not "monthly active" – daily!

It's quite simple to check whether a product is truly European: 1. Headquarters in Europe 2. Parent company in Europe too 3. Subprocessors in Europe Some products tick every box (LeaderGPU, DentroChat). Most are a mixture. How to tell them apart:

Can't find the subprocessors? Email the company. Last resort, but the most current answer. That's how we learned PayNL backs up to Google Cloud. And sometimes the page is outdated: GreenPT runs their mobile apps on RevenueCat (US) and mentions it nowhere. We found it in the app's APK.

Normally it's not difficult to see whether a service is truly European. Ever used an app that screams European, then found out it isn't? Full guide: https://dentro.chat/blog/how-to-check-if-a-product-is-truly-european

Best realistic case these days: core product and main data on EU-owned machines, auxiliary tools like payment or CRM on US services. Scaleway, Clever Cloud, BunnyCDN and Lettermint work this way.

How to figure this out for a new service: 1. The headquarters usually sits in the terms and conditions or privacy policy, linked in the footer. 2. The subprocessor list is often linked there too. 3. An AI agent is good at digging out parent companies. We use DentroChat for this every day.

That means: to be sovereign, the most important step is using products with HQs in Europe. It's not a European product if it's owned by Microsoft (e.g. LinkedIn), Salesforce (e.g. Slack), or Google (e.g. YouTube), even when they have subsidiaries in Europe.

Second most important: the subprocessors. If the whole product runs on Google Cloud, like Mollie, the Cloud Act still reaches the data. Migrating to a European cloud such as Scaleway or Clever Cloud fixes that.

Most important: where the headquarters and the parent company sit. If it's the US, the company falls under the US Cloud Act. This allows US agencies to demand data of this company, even if the data is located in Europe.

Also, it can happen at any moment, that some president decides to cut off all of Europe from their IT infrastructure...

One of our founders got a rough ride in a reddit thread about AI wasting water. Funny thing: serving that thread's 6,000 views took roughly the energy of 18,000 AI prompts. We still went and read the research. 🧵

Google later published measured numbers: a median Gemini Apps prompt uses 0.26 ml of water (about five drops), cooling and idle overhead included. Datacenters in dry regions are a real issue and disclosure there is bad.

And on EU infrastructure the math flips: Hetzner in Germany (where we host) cools with air instead of water, and Sweden datacenters run on outside air all year. An EU hosted prompt is one of the lighter things we do online. Numbers: https://dentro.chat/blog/is-ai-wasting-water

The viral water number is legit research: Li et al. estimated GPT-3 uses a 500 ml bottle of water per 20-50 responses. But those figures are extrapolated, assuming hot climate datacenters where cooling is basically servers sweating. That water is gone for good.

11 Sept

Meet one of the humans behind DentroChat 👋🏽 I'm Paul, one of the founders. We read all emails and feedback here so if you have a question or an idea, write us!

50+ stars made us migrate from Github to Codeberg! 🏔️ Finding *good* European Infrastructure providers was hard, so we created a github repo earlier this year to make it easier.

😵‍💫 The issue was: promoting EU Infrastructure on a US git hosting is strange... Now that we've gotten over 50 stars and a bunch of contribution, we moved it to Codeberg, Europe's biggest Git Host for non profit code!

Check it out, and submit a PR to add good European infrastructure 🛠️ Codeberg repo: https://codeberg.org/dentroai/awesome-eu-infra New Awesome EU Infra webpage: https://awesome-eu-infra.com/

50+ stars made us migrate from Github to Codeberg! 🏔️ Finding *good* European Infrastructure providers was hard, so we created a github repo earlier this year to make it easier.

We called it "awesome EU infra" in the spirit of "awesome" repos. It lists IT providers that have a great offering (instead of just great marketing).

3 Sept

Today we tested GDPRChat. For people who don't want to wait 20 seconds for a response to "hello friend", use DentroChat instead 😉

ProductHunt now flags posts that are probably written with AI. We assume more platforms will follow, because most social media run on human interaction. It won't be only Anthropic and Model Labs marking LLM output. The platforms where it gets published will mark it too.

If all AI output was truly valuable, ProductHunt wouldn't flag it. Do you think BlueSky would benefit from marking posts and comments as "AI written"?

Either way: DentroChat is the best AI chat in Europe. https://dentro.chat #AI #Europe

The reason: AI makes it easy to generate a lot of mediocre content, whether posts, images, videos or code. Sure, it also makes high quality content faster. But for each good output, ten mediocre ones get created in the same time.

22 Aug

a few words and phrases on BlueSky that we suspect are mostly from AI models: "genuine" "it's X, not Y" "The part" "what gets me" Like the quality of the post doesn't suffer from these if it's good quality, but it makes one double check the quality of the post.

How does Anthropic's text watermarking work on a technical level? 🧵

Related: 2. C2PA This adds signed metadata to generated files, e.g. images, videos. In the metadata of a file, there's a cryptographic signature of e.g. Adobe "this comes from us" Easiest to strip: just remove the metadata of the file. However, the absence of such metadata can then be telling.

We wonder if other European LLM providers will apply such measures. We somehow doubt this will become the de facto thing, simply because it has very little real world value, and we doubt most policy makers understand it. Which, conversely, might be a reason for them to push it, see chat control.

How does Anthropic's text watermarking work on a technical level? 🧵

Furthermore: This watermarking technique only works with sufficiently long text. Shorter than a paragraph, then the statistical token occurrence can't prove anything.

Related: 1. synthID by Google for images. It works similarly: the pixel values of some pixels are slightly altered, so that a machine detectable pattern can be found. Much harder to get rid of then the text watermarking currently used by Anthropic

How does Anthropic's text watermarking work on a technical level? 🧵

In that case, also, anyone else could "forge" Anthropic generated text. E.g. create the manual for a bomb, and say it was made by Anthropic! So that shows:

watermarking is not an instrument for figuring out by which LLM provider a text was generated. It simply is a compliance mechanism as of now. If someone asks Anthropic whether some text is by them, they can confirm or reject it.

How does Anthropic's text watermarking work on a technical level? 🧵

Given that the LLM provider, in that case Anthropic, is using a secret key (and potentially secret algo) to split potential tokens into 2 groups, only they can know whether a text was generated by them.

That could be solved, if they make the secret key for watermarking public. In that case, everyone could detect whether some text was by Anthropic. But wait!

How does Anthropic's text watermarking work on a technical level? 🧵

The outcome is that the resulting text has a much higher occurrence of tokens of the watermark group.

This is watermarking method is based on a 2023 paper by Kirchenbauer et al. "A Watermark for Large Language Models,". Now it gets interesting though: who can detect such watermarks?

How does Anthropic's text watermarking work on a technical level? 🧵

During token prediction, tokens are split into 2 groups (base group and watermark group), based on some secret. Then when the alogrithm, which is the LLM, predicts a new token, one of the 2 groups, the "watermark group", has a higher percentage of being chosen.