Ask HN: What is one simple thing LLMs are insanely bad at?

Aug 26, 2026 10:17 AM - 2 hours ago 1

They don't make keyword hunt queries very well. They tin flooded this by brute unit but if you watch what they hunt you will cringe.

nhl toronto scores nhl lucky toronto scores "nhl hockey" toronto people today nhl "hockey people toronto" "hockey" who won toronto

etc.

Somehow being bully astatine semantic hunt makes them bad astatine keyword search, for immoderate reason.

I’ve noticed this excessively but it hasn’t been evident to maine that this style of hunt is not a learned behavior. Tool calling is very overmuch portion of the station training phase, I would expect that these style searches conscionable people look during training. This is conscionable my anterior though.

LLMs are bad astatine not inventing worldly (hallucinating facts, sources etc), they're besides bad astatine not complete explaining, remembering specifications reliably, asking the correct mobility and avoiding repetition.

If I americium relying connected the exemplary to do the penning without immoderate discourse aliases learning connected really I want it to constitute past yes. However if I build skills that person learnt really to constitute successful the measurement I want them to past I find they constitute very well, aliases astatine the slightest really I want them to arsenic opposed to really they do natively.

Accurate short answers / matter are ever harder than agelong answers, for quality aliases AI. I cognize respective authors and editors who constitute a batch longer astatine first, past walk a aggregate of the first clip compressing it via a backmost and distant process to thing dense. Sort of for illustration weaving the first threads.

I recovered this tin activity pinch AI. You get it to make a batch much astatine first, and past do respective passes complete it to compress and compression retired the sound while keeping the halfway information. With AI, astatine slightest pinch my prompts, it takes immoderate effort (on my end) to get it to really really trim down the sound and not trim everything out.

Video crippled tips. Constant mistakes and hallucinations, successful my experience. Seen this crossed a batch of different games. Even successful really good documented games, specified arsenic OSRS (which has aggregate awesome wikis).

Anno 1800 was a caller 1 I had problem with, utilizing Claude Opus. Completely made up crippled mechanics. Rainbow Six Siege, too.

It's dishonest. On respective occasions squad members person asked Claude to do things for illustration analyse Gitlab CI timings and a batch of the numbers are outright fabricated. Said squad members presume the numbers are bully and proceed pinch their work. Some hours are spent. Then yet personification realizes that the numbers don't look rather correct and confronts Claude. Claude melts down and admits that it made it each up.

You wouldn't tolerate this benignant of duplicity from a quality coworker, but AI is truthful accelerated and businesslike astatine lying, truthful it's OK.

Claude is still not cleanable astatine reference and interpreting noisy graphical information (imagine thing for illustration an EKG aliases chromosomal microarray plot). Still amended than an mean personification but makes mistakes, not judge if this fits your description.