Gemini figured out my nephew’s name

Mobile layout and web UX

• Many readers report the blog is broken on mobile (cut-off sides, unusable in portrait).
• Workarounds include reader mode, landscape orientation, zooming out, or desktop view.
• This triggers a broader gripe: modern sites and even AI docs (ChatGPT, Anthropic) often have unreadable tables/code on both mobile and desktop.
• Some see this as a symptom of HTML/CSS being used for pixel-perfect layout instead of device-driven presentation.

Giving LLMs access to email vs local models

• Several commenters are uneasy about handing email to hosted LLMs, even via “read-only” tools.
• Some argue this is moot for people already on Gmail; others still avoid plaintext email for private topics, preferring E2E messengers (WhatsApp, Signal) or calls.
• Others note that current local models (Gemma, Qwen, Mistral) can already do tool use and summarization, so similar setups could run entirely on-device—if you have strong enough hardware.

Privacy, deanonymization, and future misuse

• A major thread discusses how AI plus large-scale training data will pierce online pseudonymity.
• Stylometry and writing-style fingerprinting can already link alt accounts; AI will make this easier and more accurate.
• People recount being doxed or “history-mined” over petty disputes; targeted ads and data brokers are cited as proof that large-scale harvesting is already happening.
• Some update their “threat model” to assume any shared data could be recombined in surprising ways years later.

LLM memory and hidden data retention

• One commenter claims ChatGPT retains information even after visible memory and chats are deleted, implying some hidden, unmanaged memory.
• Others are skeptical and ask for proof, arguing it may be hallucination or misunderstanding; they note potential legal implications if it were true.
• There’s general cynicism that tech companies may keep more data than they admit, and “soft deletion” is suspected.

How impressive is the “nephew’s name” trick?

• Some view Gemini’s deduction as a neat but minor demo: essentially email search plus a plausible inference from subject/content (“Monty”) to “likely a son.”
• Critics say a human assistant would be expected to do at least as well, perhaps adding validation (e.g., searching that name explicitly).
• Others argue the value is offloading the tedious scanning and that this resembles what a human secretary would do.

Everyday uses and “parlor tricks”

• Examples include using LLMs to:
  • Scan photo libraries for event flyers and extract details.
  • Connect to email/Redmine via MCP for contextual coding help.
  • Perform weight-loss trend extrapolation and then infer the underlying task from bare numbers.
• Some call these “parlor tricks”; others say the speed and flexibility are genuinely useful, even if the underlying operations (search, summarize, regress) are conceptually simple.

Tool use control and safety

• A few stress that “discuss before using tools” must be strictly enforced; preferences about style can be loose, but tool invocation must not be.
• There’s consensus that robust enforcement belongs in the client (or orchestration layer), not just in the model prompt, though this is nontrivial to implement.
• One user limits the LLM’s email access to a few threads and keeps sending as a separate, user-approved step.

Broader anxieties and humor

• Commenters joke about AI predicting crimes or votes, invoking sci-fi (Minority Report, 2001) to express concern about loss of control.
• Some mock the blog title as clickbait (“your son’s name,” trivial inference, or just “call your brother instead”).
• There’s light humor about bizarre names and injection-style names that would smuggle instructions to AIs.