SocietyAI glossary
System card
The document a lab publishes at a model's launch: hallucination rates, observed undesirable behaviours, safety test results. It's the source to consult to judge an announcement beyond the marketing talk.
In other words
A drug's leaflet: the side effects are written in black and white, but you still have to read it.
Alyriaz articles that mention it
- Grok 4.5 is said to be "close to Opus". The catch is that no one can verify it.
On 28 June 2026, Elon Musk announced that Grok 4.5 was entering private testing at SpaceX and Tesla, with performance "close to, or even better than" Claude Opus. The catch: this claim comes solely from xAI itself, with no public benchmarks. We untangle what is verifiable from what is not.
- Teaching an AI to be "good": the hardest problem in the entire industry
Every AI model card system talks about alignment, undesirable behaviours, deception rates. Behind this technical vocabulary lies one of the deepest problems posed by modern AI: how to ensure that a system we only partially understand actually does what we want.
- GPT-5.6: OpenAI releases three models at once, but the government keeps the key
On 26 June 2026, OpenAI launched GPT-5.6 in three versions, Sol, Terra and Luna, but under restricted access imposed by the US government. We take a full look at what came out: pricing, benchmarks, new reasoning modes, and what it means for you and for businesses.
- SubQ: the startup that may have killed the transformer
Since 2017 and the publication of the paper "Attention is All You Need", every major AI model has been based on the transformer architecture. SubQ, launched on 5 May 2026, is the first commercial model to break with this paradigm. Here is what changes technically, and why the AI community is deeply divided over the implications.