AI Horseless Carriages

AI Writing Style, System Prompts, and “Voice”

• Many commenters agree current AI email features (e.g., Gmail/Gemini) produce verbose, generic, “HR-drone” prose that feels worse than just writing a line yourself.
• There’s strong interest in exposing system prompts and letting users define style rules (“short, direct, no fluff”, per-recipient tone, etc.), or training on past emails to match one’s natural voice.
• Some point out that base models and fine-tunes already can write very human-like text; they suspect big companies deliberately keep outputs stilted for safety, branding, and comfort.
• Others worry that AI impersonating users is deceptive and socially corrosive, favoring explicit “sent on behalf of X by an assistant” labelling.

Usefulness vs. AI Hype

• Many see most product AI integrations as “AI slop”: meeting recaps they don’t read, banal “activity summaries” (e.g., fitness apps), or longer rephrasings of trivial emails.
• Widely acknowledged high-value uses: code assistants, deep search over complex docs/standards, transcription (Whisper), some RAG-based knowledge retrieval, and niche classification/extraction tasks.
• Several note that LLMs make previously hard ML problems (small-data extraction, categorization) commercially viable, even if they’re not flashy.

Meeting Notes, Summaries, and Accuracy

• Experience with AI note-takers is mixed to negative: missing key decisions, inverting conclusions after flip-flop discussions, or producing summaries that are meme-worthy but unusable.
• Long debate on accuracy: some argue AI only needs “good enough” for attendees; others insist that unreliable records are worse than none, especially for legal or historical reference.
• There’s concern about discoverability: detailed, auto-generated minutes may be subpoenaed and chill honest discussion.

Rethinking Email and Agents

• Many like the article’s concept of AI that reads and triages email, proposes labels/actions, and drafts minimal replies, versus tools that “beautify” text.
• Skeptics argue that for most short emails, prompting is slower than typing; supporters counter that non-native speakers or anxious writers gain a lot from tone-safe drafts.
• Some envision a future where agents negotiate directly (rescheduling, status, simple decisions), with humans mostly reviewing/overriding rather than composing.

UX, “Horseless Carriages,” and Future Directions

• Commenters see current chat-style interfaces and bolt-on buttons as transitional “horseless carriages”: AI wrapped around existing workflows instead of rethinking them.
• Desired next steps: new UI paradigms beyond chat, persistent learned behavior, robust tool invocation (calculators, calendars, CRMs), and products designed around AI from the start rather than AI as a feature.
• There’s ongoing tension between enthusiasm (“this feels like a new paradigm”) and skepticism about cost, reliability, privacy, and whether many AI features solve real problems or just chase hype.