Introducing Bert Speaks About AI — practical AI explained, without the hype

There is no shortage of AI content. There is a shortage of AI content that is specific, that cites its sources, and that admits what it does not know.
That gap is why Bert Colemont — ethical hacker and Head of R&D at Euraika — has started publishing on a channel called Bert Speaks About AI. The format is short, visual explanations of how these systems actually work and what European organisations are actually required to do about them. No breathless predictions, no engagement-farming doom, and no equations.
We are publishing a written companion to each video here on the Euraika site. The articles stand on their own — you can read them without watching — and each one embeds the video for anyone who would rather be shown than told.
What the channel covers
Three threads run through the material, and they are more connected than they first appear.
How the machines actually work. Not as metaphor, but mechanically: tokens, embeddings, attention, weights, inference, the KV cache. This matters beyond curiosity. Once you understand that a language model predicts the next token rather than retrieving a fact, hallucination stops being a mysterious defect and becomes a predictable property you can design controls around.
Where your data goes. What the US CLOUD Act does and does not permit. Why "hosted in Europe" answers a narrower question than most buyers think it answers. What Microsoft 365 Copilot can reach inside your tenant, and why the real risk is rarely the one in the headlines. This is the thread closest to Euraika's own work.
Running AI on your own hardware. Parameters, quantisation, VRAM, context windows — the practical knowledge needed to run a capable model locally, where the data never leaves the building at all.
Where to start
If you have five minutes and a governance problem, start with why AI rollouts fail without employee training. It covers Articles 4 and 26 of the EU AI Act — the AI-literacy obligation that has applied since February 2025 and that a surprising number of organisations have not registered as applying to them.
If you are evaluating a vendor, start with who can access your AI data. It ends with seven questions to put to any AI provider before sensitive information goes anywhere near them.
If you want the mechanism first, start with how ChatGPT actually works and follow it into the two local-AI pieces.
Why Euraika is publishing this
We build compliance software for European organisations working under NIS2, DORA, GDPR and the EU AI Act. That work runs into the same wall repeatedly: the regulation is not the hard part. The hard part is that the people making decisions about AI systems frequently do not know how those systems behave, and the people who know how the systems behave frequently have not read the regulation.
Nobody sells software that fixes that. It is fixed by explanation.
The videos are educational and deliberately independent of anything we sell — they cite primary sources, they name their own uncertainties, and where a claim depends on a specific configuration, they say so rather than rounding it to a slogan. That is the same standard we hold our platform to, and it seemed reasonable to hold our writing to it as well.
New videos land on the Bert Speaks About AI channel, with the written companion published here. Questions or topic requests are welcome — Bert reads the comments.


