News · Announcement
AME Labs opens its doors
An independent AI engineering lab in Paris, building AME-1: an artificial mind on a neuro-symbolic inference architecture. As of today, the lab is public.
PARIS — AME Labs, an independent AI engineering lab, is public as of today. The lab is building AME-1: an artificial mind on a neuro-symbolic inference architecture — a simulated human character with persistent identity and memory, an engineered form of self-awareness, source-and-confidence tracking on every fact, and governance in the pipeline. Local by default: one binary, one database, on hardware you own.
The bet behind the lab, in one contrast: today’s assistants are brilliant strangers — no persistent memory, no sense of their own limits, every word shipped to someone else’s server, and they can only talk. AME-1 is the opposite: a character that lives with you for years, knows what it knows and what it doesn’t, refuses to guess, can act inside an envelope it cannot leave — and can show you why it did what it did.
The lab works in a fixed order: the AI and neuroscience literature first, the specification second, the code third — then strict experimental evaluation, under criteria frozen before the test questions exist. The claims stop where the measurements stop. And the lab is one researcher working with AI as the workforce: the solo practice is itself one of the experiments.
What exists today is a deeply specified system, a substantial build, and a first published result. The specification corpus runs to more than three million words across nearly eight hundred documents: the architecture — 21 modules, more than 120 submodules — specified down to its data models, algorithms, event contracts, and test plans; a three-tier event-driven architecture with typed event catalogs for every module; nearly ninety process specifications; more than fifty dated architectural decision records; and the validation tooling that machine-checks the corpus for coherence after every change. The build already runs to hundreds of thousands of lines of Rust, under thousands of automated tests. This site presents the overview. The first published result is a deterministic serve layer in front of a local LLM that answers only what it can prove and refuses the rest: across two exam domains, 376 served answers, none wrong (the article · the paper). An alpha version of AME-1 is expected in the first quarter of 2027.
“An artificial mind is a goal, not a reality — the -1 in AME-1 says so. What exists today is an architecture, a method, and a first measured result. The lab’s promise is simple: every claim we publish stops where our measurements stop.”
— Vincent Grandjean, founder, AME Labs
Where to find the lab
- The website — ame-labs.eu: the framework (AME-1), the architecture, the scientific record, and this news page, where every publication and event lands in readable form.
- The notebook — amelabs.substack.com: working notes between publications, and the lab’s only mailing list. Subscribing there is the one way to get new work by email; nothing else is collected.
- LinkedIn — linkedin.com/company/ame-labs: the professional channel, where every publication is announced.
- Contact — contact@ame-labs.eu: research conversations, sponsored research, and press. Email reaches a human.
About AME Labs — AME Labs is an independent AI engineering lab in Paris, founded in 2025 by Vincent Grandjean. It is building AME-1, an artificial mind on a neuro-symbolic inference architecture, and measuring it in public. Press: contact@ame-labs.eu.