# Dec 22, 2025

## 2025 marked a transformative year
- A year of vibes

- **2025 marked a transformative year** for Armin Ronacher, as he transitioned from traditional programming to utilizing **Claude Code** and other AI tools, leading to a significant shift in his workflow and productivity.

## Structured outputs create false confidence
  
  - **Structured outputs degrade response quality** by forcing models to prioritize format over accuracy, leading to errors in data extraction and reasoning, as demonstrated with receipt parsing examples.

## ONNX Runtime and CoreML May Silently Convert Your Model to FP16

- **ONNX Runtime (ORT) with CoreMLExecutionProvider may convert models to FP16**, leading to altered predictions; to maintain FP32 precision, specify the model format as MLProgram during inference session creation.

## Universal Reasoning Model (53.8% pass 1 ARC1 and 16.0% ARC 2)

- The **Universal Reasoning Model (URM)** enhances universal transformers by integrating short convolution and truncated backpropagation, leading to significant performance improvements in reasoning tasks.

## [R] EGGROLL: trained a model without backprop and found it generalized better

- **EGGROLL** demonstrates that optimizing **NDCG directly** using evolution strategies can outperform traditional contrastive loss methods, achieving a **22% improvement** in validation scores despite a lower training score.

## Why “negative vectors” can't delete data in FAISS – but weighted kernels can

- The **Negative Vector Deletion Experiment** demonstrates that inserting negated vectors into ANN indices does **not** achieve deletion, while operator-based data structures (OBDS) can accomplish true O(1) unlearning through **destructive interference**.

## [R] Universal Reasoning Model

- The **Universal Reasoning Model** achieves **53.8% pass@1** on ARC-AGI 1 and **16.0% pass@1** on ARC-AGI 2, indicating a significant advancement in reasoning capabilities compared to previous models.

## Continuously hardening ChatGPT Atlas against prompt injection

## [P] A memory efficient TF-IDF project in Python to vectorize datasets larger than RAM

## Learning What to Write: Write-Gated KV for Efficient Long-Context Inference

- **Write-Gated KV** introduces a novel approach to **KV cache management**, predicting token utility to enhance long-context inference efficiency, achieving **46-57%** reduction in memory usage and **3.03-3.45×** speedups on the Llama model.