Blog
When a Researcher Told ChatGPT: “Treat It as Euler’s Method”
From a brief mathematical prompt to a NeurIPS Spotlight: a record of scientific exploration with ChatGPT.
What Is RLCD? The Secret Behind Jev
RLCD is a calibrated, schema-conditioned extension of pairwise reward modeling: Bradley–Terry becomes Plackett–Luce, and the reward model becomes Jev's typed decision interface.
SparseLeaf: Synaptic Sparse Fine-Tuning
Low rank is not sparsity. SparseLeaf constrains which connections can change while preserving high-rank updates, turning parameter-efficient fine-tuning into synaptic plasticity.
How Fugu Is Implemented — A Technical Inspection
A technical inspection of Sakana Fugu as a learned orchestrator: a small routing policy over frozen frontier models, adapted with SVF, a linear head, and evolutionary refinement.
Why MoE-LoRA Is Slower Than MoE and Dense-LoRA: A Mathematical View
MoE-LoRA is slow not because LoRA adds many FLOPs, but because sparse routing turns low-rank updates into many tiny, dynamic, low-reuse matrix multiplications.
CiteClaw: Reference Infrastructure for Agentic Research
CiteClaw combines citation lookup, PDF metadata, Zotero, BibTeX, CSL, batch processing, MCP tools, and an agent skill into reference infrastructure for research agents.
Macaron-V1-Preview: 749B MoL Agent Model post-trained from GLM5.1
Macaron-V1-Preview is a 749B Mixture-of-LoRA Agent Model post-trained from GLM5.1.
Agent-Own Loop Harness: When SDD Runs Itself
A small practice report on turning OpenSpec from a human-attended workflow into an agent-owned loop harness.
Harness Engineering: Building the Shell for the Ghost
For the past two years, a lot of applied AI discourse has orbited around the model itself: bigger models, better reasoning, stronger…
Triquetra for LoRA's entangled knobs: When LoRA rank changes, how should learning rate move?
Why LoRA learning-rate transfer across ranks is not universal? The key factor is how alpha scales with rank: different scaling rules lead to different rank-to-learning-rate transfer stories.
From Pointers to Footnotes: Why Agents will be “Reference-First”
If you’ve built anything non-trivial with LLMs, you’ve felt the temptation: just paste more context. Another chunk. Another thread. Another…
Agent Experience First Design (AXFD): Designing Systems for a Non-Human User
For the last two decades, “user experience” meant human experience: interfaces, flows, copy, and conversion.
On the eve of the Vibe Computing era: Why agentic coding feels like a 1980s terminal — and why that’s the point
On the eve of the birth of the Vibe Computer
An Intuitive Guide to Three Bayesian Optimization Acquisition Functions
A plain-language explanation of Knowledge Gradient, Expected Improvement, and Upper Confidence Bound in Bayesian optimization.