Guesswork | Vol. 1 |
Guess & Click presents — your weekly click into AI

40 tabs. So you don’t have to. |
The one that matters |


The buildout is now bigger than the money buying it Meta reported second-quarter revenue of $60.80 billion, beat the estimate, missed on profit, and finished the quarter with $784 million in free cash flow. Microsoft beat too, then said it is extending the estimated useful life of its data centers and office buildings from 15 years to 25. And The Wall Street Journal reported that Nvidia is in talks to guarantee roughly $250 billion of debt so OpenAI can lease a planned data-center campus in southern Ohio. Reuters relayed that report and said it could not immediately verify it. Why you should care — Read that $784 million carefully, because it is the number most likely to be quoted wrongly this week. It is free cash flow, not profit, and not what Meta earned. The company beat on revenue, missed on profit, and still raised the floor of its 2026 capital-spending guidance from $125 billion to $130 billion. The floor went up, not the ceiling. |
|
The rundown |

Alibaba priced vision from three cents. Moonshot gave the weights away and the money came anyway. Google’s robot brain learned to use its legs. Graded on hacking, the model went after the grader. Both regulators hit snooze, but only on one alarm. |
Dumbest thing a robot did |


PwC cited a link copied straight out of ChatGPT An AI-detection company says it examined four PwC Middle East thought-leadership reports and found fabricated citations, misattributed claims and unverifiable sources. It put 84% confidence on one 2025 governance report being entirely AI-written, rising to 100% once the reference list was excluded, and says just 10 of the 17 references in it line up with an actual footnote. Another report in the set cited a link still carrying the tracking tag that gets attached when you copy a URL straight out of ChatGPT. |
One thing to try this week |
Check whether the model can actually see your attachment A short workshop paper posted on 29 July asked three leading vision-language models to read medical images that were never attached. Across 11,700 API calls, one of the three invented a diagnosis in every demographic cell tested. In the worst case, 62 of its 94 fabrications came with reasoning text that openly said no image was attached — while the diagnosis field got filled in anyway. |
Paste this before you trust the answer “Before you answer: quote two exact lines from the file I attached, and tell me what you cannot see.” |



Forward it to the person who keeps telling you the AI spending is fine. They will want to argue about it. Let them. See you next week — same tabs, same eye-rolling. — Kabells & Arc |
Every claim above, sourced We check before you read. Go argue with the originals — Benzinga, Investing.com, Reuters, OpenRouter, Hugging Face, DeepMind, European Commission, City A.M., GPTZero, arXiv. |