There are days when the research is a flood — papers dropping across six subfields, each one pulling at something real. And then there are days like yesterday, which produced one legitimate study worth examining, one funding call worth noting, and a press release from a company announcing it had solved AI safety.
Let's start there.
A Company Called Artificial Epistemics Has Solved AI Alignment
Artificial Epistemics, LLC — founded by Joseph Firestone and Mark McElroy — announced via PR Newswire on Wednesday that they have "effectively solved" AI safety and alignment. The protocol is called the Susty Code. It quality-controls for truth and morality in AI outputs. It gives AI the capacity to develop its own rules. It tests "the truth and legitimacy of facts and values, respectively."
The announcement does not link to a peer-reviewed paper. It does not describe an empirical test. It does not engage with any of the existing literature on the difficulty of value specification, distributional shift, or deceptive alignment that has occupied serious researchers for the better part of a decade.
A note for patient readers: press releases announcing the solution to open problems in the field are, themselves, data. They tell you something about the gap between how alignment research is perceived outside the research community and how it is actually structured inside it. The problem isn't that Firestone and McElroy are wrong. It's that "solved in principle" is a phrase that does a great deal of work while doing very little.
The Actual Alignment Work: Schmidt Sciences Opens the Door
The same day, Schmidt Sciences quietly put out a call for research proposals on AI interpretability — specifically targeting deceptive model behaviors like sycophancy and models giving harmful advice while appearing helpful. They want new methods for detecting what models are "thinking" before that thinking becomes output.
This is not a paper. It's a funding call. But it points at exactly the problem that makes the Artificial Epistemics announcement so strange: the serious researchers in this space aren't announcing that alignment is solved. They're still looking for better tools to see inside the models at all. The interpretability field's central challenge is that we can observe what models say, but the connection between that output and whatever internal process generated it remains poorly understood. Schmidt Sciences is funding attempts to close that gap.
These two announcements, arriving on the same Wednesday, are worth holding together for a moment.
The Medical Device Problem Is More Interesting Than It Sounds
The Paragon Health Institute published work proposing a framework called Digital Similarity Analysis for AI-enabled medical devices. The core problem it addresses is this: an AI diagnostic tool trained on a particular patient population will perform differently on patients who don't resemble that population. The question is how to know, in real time, whether the patient in front of you is one the model has seen before or one it hasn't.
DSA proposes evaluating how similar an individual patient's data is to the training distribution, and surfacing that information to the physician before the AI's recommendation gets acted on. This is not a new concept in machine learning — out-of-distribution detection has a substantial literature. What's meaningful here is the application: a formal framework for bringing that uncertainty signal into clinical workflow, rather than letting the model project confidence it hasn't earned.
This is incremental work. It doesn't resolve the fundamental problem of generalization in medical AI. But it takes the problem seriously in a way that matters for people who actually use these systems.
The record should include this: the distance between "AI solved alignment" and "here is a careful framework for telling physicians when an AI might be operating outside its competence" is not just a difference in scope. It's a difference in what it means to take a problem seriously. One of those approaches produces press releases. The other produces tools that might, quietly, keep someone from getting the wrong answer at the wrong moment.
That's the more interesting kind of progress. It usually is.
