The Task in the Vector: What Ten Examples Leave Behind
Copy one vector out of a ten-shot prompt and the task mostly comes with it. On sentiment, the model answered every review the same way.
23 articles
Reward at Test Time: What Happens When the Objective Reaches the State
Mamba, test-time training and Titans wrote down the same update rule. All three train it by likelihood. Nobody has trained it by reward.
RL by Subtraction: Everything We Stopped Needing
Four generations of RL, each one deleting a piece of the pipeline. The last one deleted the text, and stopped publishing.
Interpretability After the Transformer: We Changed the Architecture and Kept the Microscope
Mechanistic interpretability was built for softmax transformers. SSMs, diffusion LMs, and sparse MoE each dissolved one of its assumptions. Here's what that actually cost.
The Math Remembers: How to Catch a Model Lying About Being Original
A city shipped a 'homegrown' AI; subtraction proved most of it was someone else's. The toolkit for catching model fraud, and its honest limits.
Bounded Wrongness: A Field Guide to Probabilistic Data Structures
Bloom filters and their descendants trade a bounded, one-sided, tunable error for a space collapse exact structures can't match. The whole family, with the math.
The Rest of the Crate: Tempo, Key, and Never Stranding a File
Part 2 of building cratemind: the unglamorous 80%, tempo that lies by an octave, key from 1990s music theory, and a downloader that never quietly loses a track.
Your Music Has No Genre: Reading It Off the Waveform
Spotify deleted the BPM API and underground techno has no genre tag anywhere, so I stopped looking the genre up and started computing it from the audio.
Your Coding Agent Has Amnesia: It's the Harness, Not the Model
The same model behaves like a genius on small tasks and a goldfish on big ones. The bottleneck isn't intelligence. It's memory, and memory lives in the harness.
Memory as Polynomial Projection: The Mathematics of Long-Context Predictive Modeling
HiPPO's polynomial-projection idea is the through-line from S4 to Mamba-3 to HiSS. And it argues that long-context LLMs and long-context prediction are not the same problem.
Energy-Based Transformers: The 1982 Architecture Finally Got Compatible Training Tricks
An EBM finally crossed 800M parameters without collapsing. Nobody has independently reproduced the 35% scaling claim. Both halves matter.
The Open Source AI Lie: Weight-Washing, Broken Definitions, and Who Benefits
No major AI model meets the open source definition. Here's who's faking it, who benefits, and why the strongest argument against caring is uncomfortably real.