The AS3P blog

What we learn in the field, not what the market announces.

Tested before it is written

The AI market is dominated by marketing noise. Every week someone announces a revolution, a model “ten times more powerful”, a feature that will “change everything”. Behind those announcements, the reality in the field is usually more nuanced.

Our analysis rests on more than a hundred AI projects delivered for mid-sized firms, law practices, healthcare professionals and independent IT specialists. When we write that a RAG system reaches 81.67 % accuracy with a dual schema, it is because we measured it ourselves.

Articles in English

The full blog is published in French. The articles below are available in English; the others are being translated gradually.

11 August 2026

The AI Act: how BrainDup answers it, article by article

The EU AI Act applies since 2 August 2026. What it asks for is not good intentions but evidence: logs, traces, sources, decisions. Article by article, the mechanism that produces them.

7 August 2026

Sovereign company AI: BrainDup stays with you

Marvia answers questions about the Morvan, cites her sources, and runs on an appliance you can unplug. How a fully local RAG settles the confidentiality question the cloud leaves open.

25 August 2026

Ontology and LLMs: the semantic turn in enterprise AI

Palantir, Microsoft and Google are anchoring their AI architectures to an ontology. Twenty-five years after the semantic web: what that turn demands technically, and what a sovereign KG-RAG foundation already implements.

28 March 2026

Multilingual multimodal RAG: why your AI ignores your non-English documents

A benchmark of 10 embedding models: most RAG systems understand only English and see only text.

31 March 2026

AI psychosis: when chatbots break the hold on reality

Documented clinical cases: 15 suicides, 90 hospitalisations across 22 countries. The phenomenon of building delusions together with AI chatbots.

20 August 2026

AI watermarks: regulatory transparency or life insurance for the models?

Why Anthropic and OpenAI mark their generated content: article 50 of the AI Act, the quality of the web, and protecting future models from collapse.

27 August 2026

Qwen3.8 Flash Next: the draft of Qwen4 already beats Opus 4.6

Qwen3.8 Flash Next, a preview of Qwen4, already leads Claude Opus 4.6 on all 4 benchmarks compared. Numbers, licence and limits.

18 September 2026

Taxing robots and AI: the roses and the thorns

From the ROSE levy to the 2026 Gates plan: what appeals about a robot tax, what blocks it, and why we back the augmented human instead.

25 March 2026

Tokenmaxxing: when burning tokens becomes a corporate sport

An OpenAI engineer consumed 210 billion tokens in one week. A critique of the race for volume with no added value.

20 March 2026

81,000 people said what they want from AI

An Anthropic study of 81,000 people across 159 countries: 81% say AI has already delivered, but 27% worry about reliability.

17 March 2026

Mastercard launches a virtual CFO for SMEs

Mastercard launches the Virtual C-Suite: virtual AI executives for SMEs. A revolution, or dependency?

13 March 2026

Legora: from start-up to legal AI giant

Legora, a Swedish start-up founded in 2023, went from 40 to 400 staff in a year. Valuation: 5.5 billion dollars.

8 April 2026

Anthropic did not buy Notion: what the non-deal says

A look at MCP, an open protocol that avoids silos. Why Anthropic not buying Notion says more than an acquisition would have.

10 March 2026

Claude Code Review: when AI reads the AI's code

Anthropic launches Claude Code Review, a multi-agent code review system. Result: from 16% to 54% of PRs with substantive comments.

12 March 2026

Claude generates interactive charts

Anthropic gives Claude the ability to generate interactive charts, free for all users.

13 April 2026

NVIDIA NemoClaw and Open Cloud: the trap behind the promise

A critical look at NemoClaw and Open Cloud. A complex PDF takes 68 seconds locally against 4 seconds on Claude.

The full blog in French

Eighteen articles on sovereignty, RAG, productivity and compliance.

Go to the French blog

From watching to doing

BrainDup puts these concepts into production: a sovereign AI that runs on your premises and cites its sources. We are looking for IT partners to deploy it.

Talk to us about a partnership