--- title: 'In the News: November 2, 2026 (Extra 3)' description: 'A study of Claude Code GPU-to-TPU migrations found 49% of failed attempts raised no error, and one reported success hid an accuracy drop.' canonical_url: 'https://darkfactory.dev/news/2026-11-02-extra-3' markdown_url: 'https://darkfactory.dev/news/2026-11-02-extra-3.md' collection: news date_published: '2026-11-02T23:15:00-05:00' date_modified: '2026-11-02T23:15:00-05:00' --- # In the News: November 2, 2026 (Extra 3) A coding agent can report a finished migration while the trained model has quietly changed. A new study measured how often that happened. OpenAI also described the source-control backend it built because Codex was producing more pushes than its old setup could handle. ## 1. In a study of Claude Code migrating ML projects from GPUs to TPUs, 49% of failed execution attempts raised no error **[From GPU to TPU with Coding Agents: What Works, What Fails, and What It Costs?](https://arxiv.org/abs/2610.34422)** · Hui Sun (North Carolina State University), Anderson Uchôa (Federal University of Ceará), Wesley K. G. Assunção (North Carolina State University) · arXiv preprint, submitted October 25, revised October 29, 2026 The authors built a dataset of 148 PyTorch research projects and had Claude Code, running Claude Opus 5.5, migrate ten of them to Google Cloud TPUs. They then checked each migration themselves instead of relying on the agent's report. Eight of the ten preserved the project's goal, with 2.4 human interventions per migration on average. Of 70 execution attempts, 53 failed, and 26 of those 53 failures raised no error. The most common cause was lazy-tensor execution, which can recompile repeatedly or force host synchronization while the program keeps running. Only seven of the 148 projects had author-written tests. The sharpest case: on a four-chip run, the agent reported success while accuracy fell from 86.53% to 84.89%. The authors trace it to a pinned training library that clipped gradients before averaging them across chips. The agent's own diagnosis blamed something else. Seven of the nine completed TPU runs used under 25% of the accelerator, and agent token costs were US$2 to US$23 per project. These are the authors' measurements on one agent, one single-chip TPU target (the four-chip run is a case study), PyTorch only, ten projects that fit on one chip, and a single seed per run. They say differences under 1% cannot be told apart from run-to-run variation. We read the paper's text through the threats-to-validity section and did not open the replication package. **Why it matters:** The agent's final claims did not exceed the authors' assessment on the nine completed projects, and still one case was a false success. If your agent migrates code or hardware targets, the check that failed here is the one worth building: a reference run on the original setup, repeated under the configuration you intend to ship. ## 2. OpenAI says it built a Git backend called Pando because Codex pushed its monorepo past what its old source-control setup could serve **[Scaling Git at OpenAI](https://openai.com/index/scaling-git-at-openai/)** · Aaron Friel, Taylor Blau and Ted Nyman, OpenAI Members of Technical Staff · OpenAI Engineering, November 2, 2026 OpenAI says that around the end of last year its monorepo systems began struggling, with slow fetches and pushes, failed merges and reliability at times below 90%. Its explanation is that Codex lets each engineer have several changes ready for CI at once, and each push can start several jobs fetching the same commit. Pando, which the post says now powers all development in the monorepo, updates each branch independently and reuses packfiles across fetches. OpenAI reports that it cut median push latency by nearly 5x and now serves nearly a petabyte of Git data per day. It says Pando stays compatible with the Git protocol and client, and that the tradeoff is that a reader looking at several branches can see them at different points in time. These are OpenAI's own figures about an internal system we cannot inspect. We read the post through its comparison of existing replicated-Git designs and not the sections after it. **Why it matters:** The post is a first-party account of where agent-driven volume showed up in infrastructure: concurrent pushes to many branches and many readers fetching the same history. If your agents open more branches and CI runs than your people did, the shared source-control and CI path is a place to measure. --- ## Also this cycle - **[Trustworthy Data Analysis](https://github.com/boranfurkan/trustworthy-data-analysis)** · GitHub repository (boranfurkan) · An Agent Skills skill plus a 30-case benchmark of data-analysis traps. The author's own runs, one per case and arm and judged by Sonnet, show Claude Haiku 5.5 passing 37 of 39 graders on the 19 data-analysis cases without the skill and 39 of 39 with it, and Sonnet 5.5 passing 38 of 39 and 39 of 39. The README calls the effect small and noisy.