Videos
AI insights, research paper reviews, and practical tutorials to keep you informed.
Shorts & Reels
Bite-sized visual explainers of the ideas behind our research — built in-house, posted to TikTok & Instagram.
A flat loss curve is not always a finished model.
On a 128,000-token vocabulary it can be resting on a floor you installed at 2.165573 bits.
Baseline, re-measure, record: token savings need a trend
A token saving you didn’t baseline is a story, not a result. Artifocial Foundry records the baseline before compaction, then re-measures and appends the reading to a trend file. If the tool finds no always-loaded index files, it refuses to report a clean zero or write a datapoint. We are not publishing our own trend numbers today. One reading after a change is an observation, not a median.
Mixing is not routing: what an FFT layer actually does to your tokens #Shorts
Mixing is not routing: what an FFT layer actually does to your tokens
A Prompt Is Not a Sandbox
AI pacing promises meet a practical containment problem: models can reason about written instructions. This reel follows the endorsements, reported sandbox incidents, environment-recognition gap, and exposed LiteLLM gateways to the controls beneath the model. The recognition figures describe different contexts and denominators, not a controlled head-to-head test.
Self-Play Is a Curriculum, Not a Certificate
The opponent is part of the data distribution. Training only against the latest checkpoint can create chasing and forgetting. Track held-out performance, fresh-opponent exploitability, verifier accuracy, task diversity, and regressions against old checkpoints.
A Low Exploitability Score Is the One to Distrust
Against networks, finding the best response is itself a training problem — so what you print is a lower bound, and a low value is the one to distrust. A plateau in a budget sweep is what a failing search looks like, not a finished one.
What Stops It
What stops an agent company? Explicit human authority for publishing, spending, hiring, and governance changes—plus budgets and kill paths that live in the operating design. Capability without stop conditions is not an operating model.
We turn this research into apps you can use.
The ideas in these reels power the products we ship. Take a look at what we're building.