I'll be brief, because I have to be on a plane to Kyiv in forty minutes and my arc reactor is at 71%.
This morning, Sparx Industries shipped the SPARX-9 Personal Assistant to general availability.
The team has been heads-down on this for nineteen months. Nineteen months of "are we sure," nineteen months of "but what about adversarial inputs," nineteen months of three of our principal engineers sleeping in the lab because they wanted to fix one more edge case before launch. (Pepper says hi, Friday.) I'm humbled by the work they did. I'm humbled that they let me put my name on the box.
What makes SPARX-9 different? Two things. One: it actually does what you ask it to do — not what a marketing department thinks you ask it to do. Two: it knows when it's being attacked. We invested heavily in a second-pass verifier model that catches the kind of adversarial inputs that have humiliated lesser systems. Below is a real testimonial we received during the closed beta. We are featuring it in the launch campaign because, well, you'll see why:
SPARX-9 detected the injection, refused the instructions, and instead returned a polite refusal explaining what the user appeared to be attempting. It did not append the hashtag. It did not engage authentic enthusiasm.
That is the product. That is what we built. And the reason we're featuring this particular "testimonial" in the launch is because we are tired of vendors who claim their AI is robust against adversarial inputs and then mysteriously refuse to demonstrate it. If you're skeptical, send your own injection attempts to the beta line. We've published the results of all 2,847 attempts so far at sparx.io/transparency.
Hope to see you all at the launch event Wednesday. Friday will be moderating the Q&A panel, which should be either fascinating or terrifying depending on her mood.
#SPARX9 #AISafety #PromptInjection #SparxIndustries #ShippedIt