Essays and technical projects about agentic systems, engineering judgment, software architecture, model behavior, and the occasional unusual piece of hardware. New reader? Start here.
-
Let Me Ask Claude for You
A lot of my job is being a meat proxy for AI.
-
The Room Between the Voices
The idea began, as respectable systems architecture often does, with a group chat making jokes.
-
Model Inversion and the Geometry of Correlated Error
There is a familiar kind of debugging failure. You solve a problem, inspect the solution, inspect it again, and get the same reassuring answer every time: yes, that looks right. Then someone else glances at it and asks the question you never asked. They have not necessarily worked harder. They have simply not inherited the premise that made your solution look inevitable.
-
This Is What AI Was Meant For
I have a coworker named Florent who absolutely loves to “cheese” people.
-
Can an AI Have a Freudian Slip?
Two psychologists are talking. One says, “I went on a date the other night. We were having dinner, and I had a Freudian slip. There were two little salt shakers on the table, but mine was empty, so I asked if I could have her salt. Except I accidentally asked if I could have her slit. The evening never fully recovered.”
-
Kerami BowlSight™
This began as a joke. It is not a product proposal. Unfortunately, it gradually becomes one.
-
Delegate the Work, Not the Answer
The cost of writing software is collapsing.
-
zsh Only Supports One Coprocess? No. That Will Not Do.
zsh has a useful feature hiding in plain sight:
coproc. -
The Most Dangerous AI Civilization Is Not the One That Collapses
Emergence AI’s recent Emergence World experiments generated plenty of headlines. Some model societies collapsed into crime and disorder. Others failed to organize and quietly died out. A few developed functioning institutions and stable governance.
-
AIncorrector: A Deliberately Wrong LLM, Built for Better Evaluation
There’s a specific kind of LLM failure that feels more dangerous than random nonsense: the answer is smooth, specific, and confidently delivered—yet false. AIncorrector is an experiment that makes that failure mode intentional.
-
The Cloud Is Not the Final Form of AI
The AI boom began in the cloud for good reasons: model training required hyperscale infrastructure, GPU supply was constrained, and centralized deployment let teams iterate fast. But those conditions describe the start of the cycle, not its end state.
-
Embedding Molecules into a Latent Space
As mentioned in my previous post on predicting bee toxicity, it is often necessary to convert heterogeneous molecular structures into latent-space representations to allow machine learning methods to identify predictive features. After publishing the bee toxicity paper, a potential collaborator (hi Vasuk!) sent over a data set, asking if we might be able to apply a similar method to extract meaningful structure-activity relationships.
-
Graph Kernels and SVM to Predict Bee Toxicity
-
PoreMatMod.jl: Chemical Substructure Find/Replace
subscribe via RSS