Checkpoint
Checkpoint on Beyond Market Intelligence: a running collection of 3 stories we have gathered and hand-picked because they are worth your time. Every post here touches on checkpoint in some way — the news, the analysis, the deep dives, and the occasional surprise find. Acme AI is the next-generation, AI-powered spreadsheet platform built to replace Excel and redefine how analysts, data scientists, and enterprise teams work with data. New stories are added to this page as we find them, so check back if you want to keep up with what is happening around checkpoint, or subscribe to the RSS feed to get them as soon as they are published. Browse the collection below, or head back to the homepage to see everything Beyond Market Intelligence is covering right now.
![Survival of the Fitted: Qwen3.6-27B’s Jacobian lens reads and steers Qwen3.8-27B with zero refitting [R]](https://preview.redd.it/49qcp6szzkjh1.jpeg?width=640&crop=smart&auto=webp&s=3a9f5c5e0c028e8d98994f2b827853953eb7d529)
Survival of the Fitted: Qwen3.6-27B’s Jacobian lens reads and steers Qwen3.8-27B with zero refitting [R]
Recent research demonstrates surprising stability in interpretability lenses across model updates. Specifically, a Jacobian lens fitted to Qwen3.6-27B effectively steered Qwen3.8-27B, a subsequent version, with zero refitting. This study, detailed in a new Hugging Face dataset, reveals that transferred lenses maintain their ability to identify latent entities, even exhibiting improved performance at mid-depth layers. The findings suggest a measurable transferability of these instruments, potentially streamlining monitoring pipelines and reducing the need for constant refitting. Explore the full dataset and analysis here: [https://huggingface.co/datasets/ec75hash/jacobian-lens-
Transformers are famously bad at arithmetic, so I set one's weights by hand (no training) and it multiplies with 100% accuracy [P]
Researchers have demonstrated a surprising feat: achieving 100% accuracy in arithmetic calculations within a Phi-3 transformer model, entirely without training. By meticulously hand-crafting the model's weights to implement a grade-school multiplication algorithm, they’ve created a functional three-digit calculator—and extended it to support up to 12-digit multiplication via Hugging Face checkpoints. This experiment highlights a stark contrast in performance compared to frontier models, revealing limitations in their ability to handle precise calculations.
I built a compiler that turns computation graphs into the weights of a vanilla transformer — no training anywhere [P]
Explore a novel approach to transformer architecture with TorchWright, a compiler that generates transformer weights directly from Python computation graphs – eliminating the need for any training. This innovative system, detailed in a recent post on ood.dev, allows users to define algorithms independently of the learning process, producing standard Phi-3 checkpoints compatible with vanilla Hugging Face. See how this achieves expressiveness within a transformer, building upon work like RASP while prioritizing accessibility and a stock architecture.