Mistral’s next big model has a date, OpenAI’s invisible watermark gets a name, Spanish public administrations still struggle to move AI past the pilot, and big tech starts capping what staff spend on outside models.
Mistral prepares ML4 release
Illustration generated with AI on our own machine.
What happened. French company Mistral AI is preparing to roll out its ML4 model on October 27 (cnet.com). The model was trained on more than 160 languages. The company’s scientific leadership includes Chief Scientist Guillaume Lample and VP of Science Pierre Stock.
What we think. Another large model arrives to compete for GPU cycles. We will see if the multilingual training delivers utility or just more parameters to manage.
The good bit. Expanding training to over 160 languages is a significant technical undertaking. It increases the availability of high-quality weights for non-English contexts.
What you can do. Prepare your infrastructure for larger models by using our VRAM calculator.
OpenAI introduces textGrain for EU compliance
What happened. OpenAI announced on October 5 it will add an invisible watermark called textGrain to ChatGPT and Codex outputs in the EU (coincentral.com). This system responds to transparency rules under the EU AI Act guide. The detector caught about 66% of unedited answers around 150 words long. Detection fell to 17% when replacing one word in four. At 300 words, the detection rate reached about 92%. Swapping one word in 10 for a synonym cut detection to 66% for 300-word passages. Access is currently limited to approved researchers and expert organisations. API customers worldwide can opt in. Criminal use of AI tools increased 40% year-on-year.
What we think. It is a clever way to follow the law without solving the problem of manipulation. If changing a single word breaks the detector, the watermark is more of a suggestion than a barrier.
The good bit. The move provides a standard mechanism for researchers to track synthetic content. It offers a starting point for accountability during a period of rising misuse.
What you can do. If you publish AI-generated text or images, check which Article 50 labels apply to you on our EU AI Act page.
Spanish AI production hurdles
Illustration generated with AI on our own machine.
What happened. A Capgemini Research Institute report on public administrations shows only 27% of Spanish administrations have moved AI projects to production scale (Europa Press). While Spain leads in internal process automation at 73.3% compared to a 56.8% global average, 83% of managers cite data quality and availability as a main barrier. Only 17% of Spanish public organisations are AI Front-runners. Furthermore, 76.7% of Spanish Administrations are willing to assume higher costs for sovereign solutions.
What we think. Spain is good at automating the basics but struggles with scaling. The high cost of sovereignty is a heavy price for autonomy.
The good bit. The high level of automation in HR, Finance, and IT shows a strong foundation for digital transformation. There is a clear commitment to data security and public trust.
What you can do. Learn how to manage your local infrastructure with our local deployment guide for SMEs.
In one line
From invisible watermarks to corporate budget cuts, the AI industry is entering a phase of stricter regulation and tighter controls.
Work with us
Contact us for more information or to discuss our consulting services.