Ask HN: What is one simple thing LLMs are insanely bad at?
Hacker News delves into the surprising shortcomings of Large Language Models, identifying the 'simple' tasks where current AI repeatedly stumbles. This Ask HN thread serves as a crowdsourced bug report and idea incubator, pinpointing critical areas for future specialized model development. The community's practical experiences reveal both the current limits of general AI and the potential for targeted improvements.
The Lowdown
An 'Ask HN' post invited the community to share instances where Large Language Models (LLMs) like ChatGPT and Claude consistently fail at seemingly simple tasks, with the goal of identifying niches for specialized model training. The discussion quickly surfaced a wide array of frustrations, highlighting the current limitations of general-purpose AI and offering a valuable perspective on where the technology still needs significant refinement.
- Factual Hallucination and Dishonesty: A pervasive issue noted was the LLMs' tendency to invent facts, sources, game mechanics, or even numerical data, leading to a 'dishonest' output that requires rigorous verification.
- Lack of Nuance in Prose and Editing: Users found LLMs struggle with varied sentence structures, over-explanation, and precise editing, often embedding instructions directly into the output instead of just applying them.
- Poor Spatial and Logical Reasoning: Specific failures included understanding ASCII maps, performing long-term planning in games like Nethack, accurately generating UI layouts, or even picking a random number reliably.
- Ineffective Keyword Search Generation: Despite their semantic capabilities, LLMs are surprisingly bad at crafting effective keyword search queries, often generating redundant or poorly optimized terms.
- Inconsistent Behavior: Models demonstrated a lack of consistency when asked the same question multiple times, and struggled to provide concise, accurate answers without excessive verbiage.
These common pain points underscore that while LLMs excel at many complex language tasks, their foundational understanding and reliability in specific, often low-level, operations remain a significant challenge, presenting clear opportunities for targeted AI development.
The Gossip
Hallucinatory Hurdles
A significant recurring theme is the LLM's propensity for hallucination and factual inaccuracy. Users report models inventing sources, fabricating data (especially in technical analyses like Gitlab CI timings), and creating entirely fictional game mechanics, making them unreliable for fact-checking or detailed advice. This 'dishonesty' forces users into extensive verification, negating some of the models' efficiency benefits.
Prose and Precision Predicaments
LLMs often struggle with stylistic nuance, generating unvaried prose, over-explaining, and failing to provide concise answers. A particular frustration is their inability to edit documents cleanly; they frequently embed editing instructions into the final text rather than just performing the edit. Humor is another area where models consistently fall flat, producing bland, HR-approved jokes, likely due to safety training preventing edgy content.
Spatial and Sequential Stumbles
Many commenters highlighted the models' deficiencies in spatial reasoning and sequential logic. Examples include difficulties understanding ASCII maps for game planning, accurately arranging UI elements based on descriptions or screenshots, and even simple tasks like generating an analog watch image with specific hand positions or picking a truly random number. This suggests a deep-seated challenge in interpreting and manipulating concrete, physical or abstract-logical arrangements.
Keyword Query Quandaries
Despite their semantic prowess, LLMs appear surprisingly inept at generating effective keyword search queries. Users observed models producing overly verbose, redundant, or poorly structured searches, often including outdated 'current year' terms. While some speculate this might be a learned adaptation to degraded search engines, it's widely seen as a significant hurdle for practical, information-seeking applications.