Skip to content

Explore Watch Read Tips Quiz Motion About Search Subscribe

Menu

Explore Watch Read Tips Quiz Motion About Search

Quiz of the week, 2026-09-27

The week two labs shipped models and a factory read twelve characters

Ten questions from this week in AI and the factory floor. One is a match-up. Every answer is a fact from a piece on this site or a public source, and every explanation says where it came from.

10 questions1 match-upsAbout 4 minutesEvery answer sourced

0 of 10 answered

Question 1 MES and the floorBefore any circuit is built on it, a bare silicon wafer is given one thing that becomes its only identity for the next three to four months. What is it? A radio chip embedded in the edge A serial number cut into it by laser A printed barcode sticker A plastic tag clipped to the carrier

A laser scratches a serial number into the wafer itself. Nothing stuck on or clipped to it would survive what comes next, so the mark has to be part of the silicon. Read the piece

Question 2 MES and the floorSEMI M12 is the standard for the human-readable mark on a wafer. How many characters does it specify? 18 32 12 6

Twelve. Its close relative SEMI M13 specifies an eighteen-character version, and SEMI T7 adds a two-dimensional data-matrix code on the back surface of a 300mm wafer. Many wafers carry both a human-readable row and a machine-readable code, on opposite sides, so one can cover for the other. Read the piece

Question 3 MES and the floorWhy do wafer identification marks become hard to read as a wafer moves through the line? The marks fade in daylight The process steps themselves cover, etch and polish the mark away Operators rub them off by handling the wafer The laser that cut them slowly reverses

The steps that build the chip are the steps that erase the name. US patent literature states that markings "may disappear or become unrecognizable" after repeated deposition and chemical-mechanical polishing, and that they also degrade through masking, etch and edge bead. Nobody did anything wrong. Read the piece

Question 4 MES and the floorTrue or false: when a reader cannot resolve a wafer's mark, the tool picks the most likely identity and carries on so the line keeps moving. True False

False, and the opposite is the point. No name means no recipe, so the batch stops at the loadport. That is the system working, not failing. What happens next is a decision the plant makes: hold and confirm the identity against the record, or send a person to read the worn edge with a magnifier and type it in. Read the piece

Question 5 The numberRoughly how many process steps does a leading-edge logic wafer pass through, and over how long? 500 to 1,000 steps over three to four months About 5,000 steps over two years About 50 steps over a week About 200 steps over one month

Five hundred to a thousand discrete steps and roughly ninety mask layers, over a three- to four-month cycle. The same six-step loop of clean, deposit, pattern, etch, clean and measure runs 80 to 120 times on the same wafer. Read the piece

Question 6 AI this weekOn 22 September 2026 OpenAI released two new models, nineteen days after its GPT-6 Astra flagship. What were they called? GPT-6 Nova and GPT-6 Vega GPT-7 Sol and GPT-7 Luna GPT-6 Mini and GPT-6 Turbo GPT-6 Sol and GPT-6 Luna

GPT-6 Sol and GPT-6 Luna, tuned for cost and speed rather than raw capability. Sol runs $2 per million input tokens and $10 per million output; Luna runs $0.10 and $0.50. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol. Read the piece

Question 7 AI this weekAnthropic shipped Claude Opus 5.5 on that same day, 22 September 2026. What was its headline claim? The first model with no usage limits A model trained entirely on synthetic data Previous-flagship-level performance at about 40% lower cost on a typical workload The first model able to run offline

Anthropic said Opus 5.5 performs at the level of Claude Fable 5.1 on most work, with about a 40% reduction in typical workload costs and output speeds up over 30%, priced at $4 and $20 per million tokens. Read the piece

Question 8 AI this weekOne day after those two launches, on 23 September 2026, Cognex launched the In-Sight 1750 Series. What does it do? Reads wafer and panel ID marks the process has degraded, without operator intervention Schedules which wafer runs next Inspects finished chips for cracks Writes the laser mark onto the wafer

It reads identification marks, including degraded ones on challenging substrates that would otherwise need a person to step in. Its chief executive framed it as combining the company's AI with decades of semiconductor expertise so manufacturers can scale production without sacrificing traceability. Read the piece

Question 9 AI this weekAlso reported this month: DeepSeek's founder Liang Wenfeng called one bet "one of the company's biggest" and said it must succeed. What was it? Abandoning open weights entirely Moving its training runs to the United States Opening its own chip factory Training its next models on Huawei and other domestic chips

Training on Huawei and other Chinese-made accelerators. Reports describe plans to deploy at least 160,000 Huawei Ascend 950DT chips at a new data centre in Inner Mongolia, with deliveries expected between the fourth quarter of 2026 and the first quarter of 2027, supporting a two-trillion-parameter model now and an eight-trillion-parameter one later. Source

Question 10 AI this weekMatch each September 2026 release to the company behind it.

GPT-6 Sol and GPT-6 Luna

Claude Opus 5.5

In-Sight 1750 wafer reader

Ascend 950DT training chips

Two frontier models on the Tuesday, a wafer reader on the Wednesday, and behind all of it a fight over whose silicon the models train on. To be fair on the reader: it is an improvement, not a breakthrough. The previous generation already read faint marks, and read rates in this field are vendor-published rather than independently benchmarked. What is interesting is that a problem this basic still supports a product line. Read the piece

Your result

Share on LinkedIn

Earlier weeks

2026-09-20The week the AI boom ran out of power 2026-09-15The week the AI companies asked to slow down 2026-09-12The week the dashboard turned out to be a copy

FabSpeak

One story a week from AI and semiconductor manufacturing, taken apart at the layer that decides whether it ships. No sponsorships, no vendor endorsements, no invented numbers.

Sections

AI on the factory floor

MES & shop-floor systems

FabSpeak

Elsewhere

© 2026 FabSpeak. Independent. Every number on this site comes from the piece it belongs to.