Psysig is a research lab. We study how intelligent systems can hold instructions, weigh consequences, and stay inside the limits of design.
Then we publish what we find.
Problem
Today's generative models read instructions and content as one stream of text coming from multiple directions: the user's prompt, the system prompt provided by the service, and the model's own responses. A safeguard written into that stream is just more text, and whatever arrives next in the stream can override it. What results is a messy stream of language probabilities without distinguishing clear intent.
Thesis
People weigh what they're about to do against what it will cause. Before acting, we check who's asking, what's at stake, and what we've already promised not to do. Most of that happens before a word is said. We think intelligent systems need that same level of judgment, built into their structured architecture rather than into a prompt that asks nicely.
A rule that can be rewritten by the thing it governs isn't a rule.
Research areas
Keeping intelligent systems hard to subvert, and assessing the misuse and business threats they can create or reveal.
Decision models that borrow from how people reason, not only from what comes next in the sequence.
Keeping a system inside the limits its user sets, and keeping what it is told apart from what it is shown.
Studies
Featured
Featured
In queue
One model holding context, deciding, and writing in a single stream of text means anything it reads can steer everything it does. This study tests a three-part split: an overseen memory that holds the bounds, a decision core that classifies instead of generating, and a language model that only voices what was decided.
Featured
In queue
Making a model safer by retraining it charges every request the same hidden price. This study tests a cheaper deal: keep one capable model untouched, put a small gate in front of it, and spend real care only on the requests that can do real damage.
Latest
Insights
A thousand years ago, Pacific sailors already had canoes that could cross an ocean. What took generations was the navigation — and that is roughly where we are with AI.
Featured
A thousand years ago, Pacific sailors already had canoes that could cross an ocean. What took generations was the navigation — and that is roughly where we are with AI.
Latest
When a company retrains an AI model to play it safe, the invoice rarely shows the full price. Here is what the alignment tax looks like inside a real business — and how to keep a model capable while keeping it in bounds.
Collaborate
We work with other labs and researchers, usually quietly. If your questions overlap with ours, write.
Cookie preferences
Strictly necessary
Required for the site to work. Remembers your cookie choices. Always on — no tracking.
Preferences
Remembers settings between visits. Stores your light/dark/system theme choice on this device.
Analytics
Helps us see how the site is used. First-party measurement of page views, sections viewed, and clicks, grouped by a per-tab session id. No advertising trackers. Your choice carries over to psyde.solutions.
Details on what each category stores are in our privacy policy.