ChatGPT now fails at most basic tasks
By CrystalScout894 · ISSUE · About ChatGPT
Of course, the videos of ChatGPT failing to count to a hundred or keep a stopwatch are more forgivable, when one understands that’s not quite much how LLMs are built and operate. Then again, the biggest competitors do not fail at these tasks.
Moreover, in the past months I’ve found ChatGPT has gone from bad to worse for pretty basic things. It’s now failing to identify images (pics ranging anywhere from,say, hardware, to writing, to species-identification). It’s failing to provide me with links for basic search prompts, even when instructed directly. I’ll give it It frequently provides the wrong answer between two options, on painfully obvious problems. (Then usually does the classic “you’re right! My mistake, it’s actually…” reverse.) I do a lot of historical research and it now consistently gets wrong event/people/idea names when I provide them (e.g. I recently gave it the shorthand of a famous SCOTUS case and it started referring to an obscure legal case from a small Asian country by the same name.)
I find it fails at these extremely basic tasks consistently now. It feels like as Claude has become super genetic and Gemini highly intelligent, ChatGPT has gotten measurably worse in terms of output, especially in rudimentary logic and processing. (Id say the exception is codex, but it’s still nothing magical compared to cowork.)
Getting frustrating that it’s reaching a point where my good old fashioned brain (ugh!) is more consistent and efficient and reliable/verifiable lol
82 upvotes · 6 comments
Comments (6)
StormNinja586: It seems that they are routing their best inference capability somewhere
HappySpark80: I recommend always using "Thinking", which as of today is 5.4, as it performs significantly better than the Instant model (5.3) in my experience.
CrystalScout894: Forgot to mention image generation/modification, list creation with constraints (often inserts errant items), remembering instructions/dialogue earlier in chat, and probably more
JordanReed: I’ve had it give me answers taken from questions I asked in other chats.
The example: it was listing slot options. The first option was “double pop”. Searched it, couldn’t find it… then I tried… wait a minute. It then listed some obscure mini game that I was asking about from the other chat as a sl
SolarSurfer919: yeah the SCOTUS case thing is wild to me, that kind of named entity confusion is exactly, what you'd expect from a model that's been quietly degraded or rerouted to a cheaper inference tier. i've noticed the same flip-flop behavior on historical stuff where it confidently corrects itself to the wron
ThunderPulse5561: Is Claude or Gemini better atm?
More discussions about ChatGPT on HonestUse