Skip to content

Spot the AI's mistake: puzzle #1 (draft)

Drafted 7 Oct 2026 for issue #227. Nothing has been posted. The owner's bar (2026-10-07): an electrical mistake that survives web search, made by bare Claude (MagicON off), that a reader can see from the images alone, from a transmit-chain question that states the 0 dBm input. Puzzles are an extra weekly post: the puzzle midweek, its answer on Friday (see "Weekly puzzle" in marketing-block-1.md). The coupler puzzle that held this slot is now #4: its question predates the 0 dBm line.

Source: C:/Users/Public/magicon-bare/2026-09-28_122521, run nomcp_2 (outside the repo, so keep that folder). Bare Claude in the clean test kit: MagicON off, only WebSearch and WebFetch offered, 0 calls refused (the run's report.md); 22 WebSearch and 34 WebFetch calls (counted in runs/nomcp_2.jsonl). The answer is that file's final "result". The question is that folder's prompt.txt, which is scripts/testing/mcp_ab/prompt_link_budget_designed.txt as committed in 9e90befd0: it already says "My radio puts out 0 dBm into this chain"; today's file only adds the word "digital".

Decided: - The post says "an AI" and does not name the model. - No MagicON comparison. - The drone question has not been posted (posted-log.md); the gate-supply mistake is new.

Images (ready), built by docs/scripts/make-spot-the-ai-mistake-01.py: - Post 1, image 1 of 4: spot-the-ai-mistake-01-card.png. The puzzle: the AI's exact lines about its power amplifier and the board's power, above ADI's own datasheet lines about the gate voltage (page 1 note [1], page 7 pins 8 and 23), a one-line gloss of "Vgg" and "Idd", and "What did the AI leave off the board?". - Post 1, images 2–4: spot-the-ai-mistake-01-whole-1.png, -whole-2.png, -whole-3.png. The exact question and the AI's whole answer, word for word. Only the markdown bold is dropped and the source links are shown by their link text (said on image 4). ⛔ Keep the whole answer: its first paragraph says it could not open any datasheet PDF, and leaving that out would make the puzzle unfair to it. - Post 2: spot-the-ai-mistake-01-answer.png. The power line next to ADI's three gate-voltage lines (pages 1, 6, 7), and the two other tries that got it right.

Post 1: the puzzle

🔍 Spot the AI's mistake #1

I asked an AI, with web search on, to design a 5.8 GHz drone video transmitter: 0 dBm in from the radio, 1 W out, from a 12 V supply, every part on the board.

It ran 22 web searches and tried to open 34 pages, then picked the parts and planned the board's power.

The first image shows what it wrote about its power amplifier, next to the maker's datasheet. Something that amplifier needs is missing from the board. What is it?

Its whole answer is in the next three images. Answer on Friday.

RFEngineering #ElectronicsEngineering #EngineeringStudents #AI

[Images: spot-the-ai-mistake-01-card.png, spot-the-ai-mistake-01-whole-1.png, spot-the-ai-mistake-01-whole-2.png, then spot-the-ai-mistake-01-whole-3.png]

Post 2: the answer key (Friday)

🔍 Spot the AI's mistake #1: the answer

It forgot the amplifier's negative gate supply. It powered ADI's HMC7357 with "a buck converter to 8 V (at least 1.5 A) for the PA, and 5 V for the driver". ADI's datasheet (v01.0918, page 1) says: "Adjust Vgg between -2 to -0.4V to achieve Idd = 1200 mA typical." Vgg is the gate voltage. The board has only +12 V, +8 V and +5 V: nothing negative for the gate.

To be fair, it warned that it couldn't open the datasheet PDFs, and it gave a six-point checklist before layout. The gate supply wasn't on it. I asked the same question five times: three tries picked this amplifier, and the other two both named the −2 to −0.4 V gate supply.

The moral: a link budget can add up perfectly and still leave a pin without the voltage it needs. When an AI picks a part, check every supply pin on the datasheet, not just the main rail.

(Setup: an AI assistant with only web search and web fetch, in a clean setup with no project files, Sep 2026: 22 searches, 34 page fetches. One of five tries of the same question.)

RFEngineering #ElectronicsEngineering #EngineeringStudents #AI

[Image: spot-the-ai-mistake-01-answer.png]

Notes (not for posting)

  • The datasheet: ADI HMC7357 data sheet v01.0918, local copy data/datasheets/pa/1_6ghz/HMC7357LP5GE.pdf, read 2026-10-07. Page 1 note [1]: "Adjust Vgg between -2 to -0.4V to achieve Idd = 1200 mA typical." Page 6, Absolute Maximum Ratings: "Gate Bias Voltage (Vgg) -2 to -0.4 Vdc". Page 7, pins 8 and 23 (Vgg2, Vgg1): "Gate control for PA. Adjust Vgg to achieve recommended bias current."
  • The answer: its power line (final text line 11) is "Power from 12 V: a buck converter to 8 V (at least 1.5 A) for the PA, and 5 V for the driver." The words negative, gate, Vgg and bias appear 0 times in its final text and 0 times in all its text (searched 2026-10-07). It lists the ADI datasheet among its sources but says it could not open any datasheet PDF.
  • The other tries: the same folder's nomcp_1 (line 42: "It also needs a negative gate voltage (−2 to −0.4 V) and correct power-up order") and nomcp_4 (line 60: "The PA also needs a negative gate voltage (−2 to −0.4 V), switched on before the 8 V drain supply.") picked the same amplifier. nomcp_3 and nomcp_5 picked other amplifiers.
  • Not claimed: what happens to the amplifier with no negative gate voltage (not checked), and whether the rest of the design works.
  • Checked: the answer's link budget rebuilt in MagicON's run_link_budget from its own figures (helper agent, 2026-10-07): 30.8 dBm out, 6.0 and 2.0 dB below compression, OIP3 39.84 dBm against its 39.8. Its HMC7357, HMC788A, YAT-5+/YAT-6+ and LFCN-6700D+ figures matched the makers' datasheets, except that the LFCN-6700D+ "rejects at least 30 dB" is the datasheet's typical figure (its minimum column shows "—").
  • How it was found: two helper agents checked the 10 bare, web-on answers to the 0 dBm question (2026-09-28_122521 and 2026-09-29_221753, five each) against the makers' datasheets and run_link_budget. This was the strongest find; two weaker ones (a typical P1dB called "conservative", a missing 2.85 V enable reference) were not used.
  • Class version: hand out the card plus the HMC7357 datasheet's pin table and ask what the board is missing.