---
title: "AI engineering, October 26, 2026"
description: "The Forward Pass daily issue, October 26, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-10-26
updated: 2026-10-26
---

# AI engineering, October 26, 2026

The Forward Pass daily issue, October 26, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-26

> 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.

_Inception v3 beats frontier LLMs on skin-cancer diagnosis, China proposes controls on 23 AI technologies, NVIDIA mass-produces photonic switches, IQuest Research releases 320B IQuest-Q1_

A skin-cancer classifier can run in your phone's browser. Inception v3 uses WebGPU without uploading patient images.

Meanwhile, a CCTV-affiliated analysis reported DeepSeek and MiniMax's overseas-export registrations as China proposed controls on 23 AI technologies. NVIDIA put its Spectrum-X photonic switch into mass production using TSMC's COUPE platform to ease copper bandwidth bottlenecks.

If you read one thing today: the Inception v3 evaluation. The head-to-head results give you a concrete basis for comparing Inception v3, GPT-6 Astra, and Claude Fable 5.1 on the tested skin-cancer task.

## Inception v3 hits 62% skin-cancer accuracy, beating GPT-6 Astra and Claude Fable 5.1

**Top News** · 502 HN points

If you are choosing a model for skin-cancer image classification, a task-specific CNN beat two frontier multimodal models in this evaluation. Inception v3 scored 62% on first-choice diagnosis, compared with 42% each for GPT-6 Astra and Claude Fable 5.1.

Here is what makes the comparison useful:
- The test used 200 biopsy-confirmed MILK10k lesions, each with clinical and dermoscopic images, and six possible diagnoses.
- Accepting second-choice diagnoses raised the scores to 173 cases for Inception v3, 134 for GPT-6 Astra, and 138 for Claude Fable 5.1.
- The CNN was fine-tuned on public ISIC images, with same-lesion and same-patient images excluded before training and no test-set tuning.
- Inception v3 can run in a mobile phone browser using WebGPU without uploading patient images.

One catch: this was a constrained, task-specific diagnostic test, not evidence of general clinical performance.

Sources: [hyper.ai](https://hyper.ai/en/stories/1786dbecdb7427e1a4e20b85fba1ab94)

## China proposes export controls on 23 AI technologies, including agent systems

**Top News** · Reported · cna.com.tw · translated from Chinese

If you deploy Chinese models containing restricted technology overseas, China is proposing export controls on 23 AI technologies. The proposal covers large-model training methods, agent systems, and computing-power clusters.

Here's what changed:
- The Ministry of Commerce and Ministry of Science and Technology opened public consultation on October 16.
- The proposed requirements apply to overseas deployments of AI models containing restricted technology.
- DeepSeek and MiniMax completed overseas-export registration and compliance procedures for open-source releases and overseas model licensing, according to a CCTV-affiliated media analysis published on October 22.
- A securities analyst described the registrations as clarifying the overseas route for general-purpose open-source models while sensitive core technologies face strict controls.

One catch: the export-control revisions are a proposal, not a finalized rule.

Sources: [cna.com.tw](https://www.cna.com.tw/news/acn/202610250093.aspx)

## NVIDIA puts Spectrum-X photonic switches into mass production using TSMC COUPE

**Top News** · Reported · ctee.com.tw · translated from Chinese

As bandwidth demands rise, copper links lose more signal, and boosting transmission power adds heat. NVIDIA announced in August that its Spectrum-X Ethernet photonic switch had entered mass production using TSMC's COUPE platform.

Here's what changed:
- Co-packaged optics (CPO) puts optical-electrical conversion and the switch chip on the same substrate, shortening the electrical path.
- Fiber connects inside the server, with the design intended to reduce power use and improve transmission speed and bandwidth.
- TSMC's COUPE stacks electronic and photonic integrated circuits, then co-packages them with the switch chip.
- The supply chain includes TSMC for silicon-photonics processing, SPIL for packaging and testing, and Foxconn for complete switch assembly.

One catch: optical components are heat-sensitive, but nearby chips generate substantial heat. Temperature changes can shift optical wavelengths and refractive index, so CPO needs micro-scale thermal isolation and management for stable operation.

Sources: [ctee.com.tw](https://www.ctee.com.tw/news/20261025700264-430501)

## Signals

1. [IQuest Research releases IQuest-Q1, an open-source 320B CLI agent model with 15B active parameters](https://hyper.ai/en/papers/2610.30949)
2. [RVL-CDIP audit finds training duplicates or near-duplicates in 35.2% of document classification test samples](https://arxiv.org/abs/2610.32126)
3. [StreamEval releases open-source benchmark for interactive video generation with 300 cases across five tasks](https://arxiv.org/abs/2610.28402)
4. [LattePanda launches Sigma 2 server with Core Ultra X7 358H, up to 64GB RAM, 136.5GB/s bandwidth](https://www.techpowerup.com/353798/lattepanda-launches-lattepanda-sigma-2-the-embeddable-ai-single-board-server) · Reported · techpowerup.com
5. [NVIDIA OpenShell runs AI agents in kernel-enforced sandboxes and formally verifies policy changes](https://github.com/NVIDIA/OpenShell) · 44 Stars today
6. [Hurston-1 generates Southern American and African-accented speech from text, including 8 kHz G.711 phone audio](https://huggingface.co/destined/hurston-1) · 56 HN points
