---
title: "AI engineering, November 7, 2026"
description: "The Forward Pass daily issue, November 7, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-11-07
updated: 2026-11-07
---

# AI engineering, November 7, 2026

The Forward Pass daily issue, November 7, 2026. Source: https://theforwardpass.net/archive/daily/2026-11-07

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_Thinking Machines benchmarks Inkling 2 Small, REINFORCE Zero speeds training 2.40×, Google tackles distracting TODOs, OpenRouter’s Auto Code matches Claude Opus 5.5 at 74% less per task_

Google found coding agents chasing bug tickets that were already closed. In Google's study, 41% of unprompted bug detours chased closed or otherwise dead-end tickets.

Meanwhile, Thinking Machines reports Inkling 2 Small scored 30.6 on AA Index v5.0, versus GLM-5.3-Flash's 27.8. REINFORCE Zero's authors report 2.40× synchronous training throughput with comparable quality.

If you read one thing today: REINFORCE Zero. Its released training code and kernel-alignment recipes let you test asynchronous RL in your own pipeline.

## Thinking Machines reports Inkling 2 Small leads JobBench at 70.5%

**Top News** · 284 HN points

Inkling 2 Small is Thinking Machines’ second-generation model for coding and autonomous work. Thinking Machines reports a 30.6 AA Index v5.0 score, ahead of GLM-5.3-Flash at 27.8 and MiniMax-M3.1-Flash at 29.1.

Here's what it offers:
- Inkling 2 Small scored 70.5% on JobBench, the highest result in the company's comparison.
- Coding scored 72.1% on DeepSWE v1.1 and 86.3% on SWEBench Pro v2 (Hard).
- The MoE has 594 billion total parameters, with 17 billion active. It accepts image inputs and supports variable effort.
- You can use it in Tinker Playground, or fine-tune and sample it on Tinker.

One catch: Initial access uses closed weights. Thinking Machines plans an Apache 2.0 weight release after external safety testing and red-teaming.

Sources: [thinkingmachines.ai](https://thinkingmachines.ai/news/inkling-2-small/)

## REINFORCE Zero reports 2.40× training throughput with comparable model quality

**Top News** · Reported · vllm.ai

REINFORCE Zero is an asynchronous RL recipe. In the reported comparison, REINFORCE Zero kept reasoning quality close to synchronous training. The authors report 2.40× training-step throughput over 12,000 updates on 16 GPUs.

Here's what changed:
- Training took 68.55 hours, versus 164.77 hours for the synchronous run.
- The 1.5B reasoning comparison reached 48.80% macro pass@1 at lag 2, versus 48.91% for synchronous training at 6,000 updates. At lag 8, asynchronous training reached 48.53%.
- The recipe uses plain REINFORCE with group-relative rewards, without a value model or sample-rejection gates.

One catch: You need aligned rollout and training kernels, plus unclipped token-level importance sampling to correct policy-weight lag.

Try it: `hijkzzz/REINFORCE-Zero` for the paper and code, `NVIDIA-NeMo/labs-molt` for training code, and `Inferact/zero-mismatch-rl` for alignment recipes and model-specific fixes. Also available via Amazon Web Services, releasebot.io

Sources: [vllm.ai](https://vllm.ai/blog/2026-11-06-zero-mismatch-async-rl)

## Google links agent bug detours to 105% more wait time

**Top News**

A TODO pointing at a bug can pull a coding agent away from your request. Google's engineering study associates these unprompted detours with 105% more wait time, after controlling for task and developer factors.

Here's what Google found:
- Detours were associated with 70% more steps and 77% more tokens per session, alongside the longer wait.
- Closed or otherwise dead-end tickets accounted for 41% of unprompted detours. Another 57% involved open bugs outside the developer's request.
- Cleanup agents inspected bug, file, and code-change history before proposing removals. The work removed thousands of obsolete references over six months, with more than 95% human-review acceptance.
- Google recommends presubmit analyzers, corpus scanners, and automated agents to catch and remove obsolete TODOs.

One catch: The measured effects are associations tied specifically to inline TODO(bug/...) and FIXME(bug/...) references.

Sources: [developers.googleblog.com](https://developers.googleblog.com/agent-distraction/)

## Signals

1. [OpenRouter’s Auto Code matches Claude Opus 5.5 across five benchmarks at 74% less per task](https://openrouter.ai/blog/announcements/auto-and-auto-code-routers/)
2. [AMD releases PrimusClaw, an open-source agent runtime for Kubernetes clusters with Instinct GPUs](https://rocm.blogs.amd.com/software-tools-optimization/primusclaw/README.html) · Official · amd.com
3. [Zyphra releases ZYFA, running FP8 attention up to 1.76x faster than AITER on AMD MI355X](https://www.marktechpost.com/2026/11/06/zyphra-zyfa-fp8-flashattention-amd-mi355x/) · Reported · marktechpost.com
4. [Nous Research releases Arete-27B, a Qwen3.8-27B model post-trained to reinforce Western values](https://nousresearch.com/arete)
5. [Vercel’s agent-desktop lets AI agents control macOS, Windows and Linux apps without screenshots](https://github.com/vercel-labs/agent-desktop) · 155 Stars today
6. [WeLM-Image-Turbo generates 1024×1024 images in 4.98 seconds on a single NVIDIA H20](https://huggingface.co/papers/2611.07392) · 39 Upvotes