I Don't Like LLMs
Strong reactions to Martin Fowler’s “I Don’t Like LLMs” essay highlight a split between seeing large language models as indispensable tools and finding their “LLM voice” and confident errors grating or untrustworthy. Many commenters report heavy daily use for coding, search, and personal tasks despite deep skepticism about accuracy, personality, and the values embedded by their creators. A recurring theme is that LLMs are best treated as fallible machines rather than pseudo-people, with some arguing that anthropomorphized chat interfaces and enshittified web search are pushing users toward shallow, homogenized thinking.
Perception of LLM “voice” and personality
- Many dislike the default “LLM voice”: verbose, hedged, self-congratulatory, faux-remorseful, and “executive-assistant-like.”
- Some say this tone triggers anxiety or irritation more than the content itself.
- Others argue the voice is easily fixed (“just prompt it differently”), while counterexamples note persistent verbosity even with strong instructions and “concise modes.”
- Debate over anthropomorphism: some never feel these systems are people; others feel the interfaces intentionally invite that illusion, which they find manipulative or creepy.
Usefulness vs unreliability
- Widespread acknowledgment that LLMs are highly useful for coding, research discovery, tax help, contracts, etc.
- Many balance “very helpful” with “I don’t trust factual output at all; everything must be verified.”
- Some treat them as advanced search/linters and avoid reading their prose to prevent absorbing wrong information.
- Annoyance at needing to constantly refute incorrect answers, especially professionally.
LLMs as tools, not coworkers
- Several commenters frame LLMs as tools or compilers, not agents to “chat” with.
- Best experiences come from terse, imperative interactions: “issue commands, get artifacts,” not life advice or deep discussion.
- Some explicitly avoid memory/“personality” features; others exploit memory for convenience (e.g., workouts, notes) without feeling it makes the system “know” them.
Trust, safety, and agents
- Strong skepticism about autonomous agents controlling infrastructure or databases; recommendation to keep AI away from destructive commands and use deterministic “circuit breakers.”
- Disagreement over whether talk of “agent swarms taking over” wrongly suggests independent AI agency vs shorthand for humans misusing tools.
- Concern that focusing on sci‑fi scenarios distracts from real current harms: bad actors using AI to do bad things more efficiently.
Search, information access, and the web
- Some feel traditional search and the web are so degraded by ads/SEO that LLMs are now the only practical way to extract information.
- Others counter that better search engines, blogs, communities, and even libraries still work if one puts in the effort.
- LLMs are praised as discovery tools for academic papers and cheaper software alternatives, sometimes even replacing paid apps.
Cultural and meta-discussion
- Debate over whether current models reflect a particular startup/brogrammer culture and value system.
- Questioning of “influencer” opinions about LLMs, especially from people perceived as no longer hands‑on.
- Frustration at snarky “gotchas,” mockery of site outages, and rising toxicity in discussions about AI, alongside participants who defend blunt disagreement as normal.