Introducing Anthelic
A new independent lab, a few questions worth years of work, and a commitment to doing the research in the open.
Today we're starting Anthelic — a small, independent research lab studying machine intelligence. This first note is short. It explains why we're here, what we care about, and how we intend to work.
The name comes from an anthelion: a faint luminous spot that appears in the sky directly opposite the sun, formed when light bends through countless ice crystals. It is easy to miss and hard to explain — a real phenomenon that rewards careful looking. That felt like the right image for what we want to do.
Why another lab
The most interesting questions in AI are not only about making models larger. They're about understanding what these systems are actually doing: how they represent knowledge, where their reasoning breaks, and what it would take to trust them with things that matter.
Those questions benefit from a particular kind of environment — small, patient, and free to follow a result wherever it leads. We're building Anthelic to be that environment.
What we're setting out to understand
We're starting with three threads. They'll sharpen as we go, and we'll write about each as the work develops.
- How models reason. Not whether they can produce the right answer, but the mechanism underneath — and when that mechanism quietly fails.
- Reliability. What makes a system behave predictably outside the examples it was trained on, and how we'd measure that honestly.
- Evaluation. Better ways to tell whether progress is real, because most of the field is bottlenecked on knowing what "better" means.
How we'll work
Three commitments, which we'll be held to by everything we publish here:
Depth over breadth. Reliability over spectacle. And doing it in the open — including the experiments that fail.
Most research is reported as a clean line from question to result. The real path is messier, and the mess is often the useful part. We'll try to show more of it than is customary.
Follow along
If any of this resonates, the best thing to do is subscribe — we'll send new research and essays as they're ready, and nothing else. You can also reach us at hello@anthelic.ai.
More soon.
— The Anthelic team