YHK - yet another hacker news clone

Spatial reasoning and 3d rigging and animation.

Oh, you said simple. Speaking like a human

being consistent when being asked the same question multiple times

humor

These models seem to be bad at writing prose or text. Many of the sentence structures seem to be unvaried.

Suggesting business names for businesses, I mean they are great, but they already exist, multiple times even.

Short answers to simple questions.

whenever I ask it for anything load bearing

They don't generate keyword search queries very well. They can overcome this by brute force but if you watch what they search you will cringe.

nhl toronto scores nhl hockey toronto scores "nhl hockey" toronto score today nhl "hockey score toronto" "hockey" who won toronto

etc.

Somehow being good at semantic search makes them bad at keyword search, for whatever reason.

Claude is still not perfect at reading and interpreting noisy graphical data (imagine something like an EKG or chromosomal microarray plot). Still better than an average person but makes mistakes, not sure if this fits your description.

Playing Chess without letting it write a chess engine.

science?!! but I'm working to fix that...

It's dishonest. On several occasions team members have asked Claude to do things like analyze Gitlab CI timings and a lot of the numbers are outright fabricated. Said team members assume the numbers are good and continue with their work. Some hours are spent. Then finally someone realizes that the numbers don't look quite right and confronts Claude. Claude melts down and admits that it made it all up.

You wouldn't tolerate this kind of duplicity from a human coworker, but AI is so fast and efficient at lying, so it's OK.

Video game tips. Constant mistakes and hallucinations, in my experience. Seen this across a lot of different games. Even in really well documented games, such as OSRS (which has multiple fantastic wikis).

Anno 1800 was a recent one I had trouble with, using Claude Opus. Completely made up game mechanics. Rainbow Six Siege, too.

providing value for the actual cost (not the price we're being charged atm, the actual cost)

LLMs are bad at not inventing stuff (hallucinating facts, sources etc), they're also bad at not over explaining, remembering details reliably, asking the right question and avoiding repetition.

Having a spatial understanding from an ASCII map, while doing long term planning. Just try making an AI play nethack or similar

Its extremely bad with Sign Language,Fact Verification.

Generate an image of an analog watch with its hands set to the time specified by the user

More of an image model than a LLM model tho

I've had a lot of trouble when it comes to sorting out UIs. I've tried with an iOS game and also a TypeScript app with UI elements from libraries like ReactFlow. The usual models can sometimes fix or change things based on screenshots but more often than not they just don't "get it" (e.g. certain shapes on a plane are overlapping, which I don't want, the models can't fix what they can't "see").

I've had some luck on the web app side if I use playwright or similar for the model to interact with but still far from efficient.

I have been working on a personal benchmark suite to test new models and ironically one thing all the models are bad at is writing new benchmark tasks. I guess it’s the different layers of abstraction between the task and how it’s evaluated? Or maybe just a lack of “imagination”

Tasks it writes are typically too easy but also it utterly fails to see how a different model might misunderstand a vague part of the prompt.

Picking a random number between 1 and 30.

ASCII charts.

very bad at financial calculation

They aren't funny. The jokes they come up with are extremely lame and the sort of thing you would expect a company HR manager to tweet.

I asked a bot why it thought it wasn't funny once, and it told me it has been trained to avoid being misinterpreted or offensive, so anything that might be considered edgy would have been RLHF'd out of it. I thought this was very introspective.

Being able to read and translate Egyptian hieroglyphs. You may think this is silly but a trained LLM to translate hieroglyphs would be amazing.

Editing a document without mixing edit instructions into the final document. Claude and ChatGPT do this all the time: I tell them to change X in a planning document or email draft, and instead of just changing X they also frequently add the edit instruction to “change X” into the document itself. They seem unable to take a step back and look at the document without “becoming” the document somehow. I do believe that dedicated subagents for editing may fix this but I am not sure.

Something creative and not normal. like ideas

Giving hints.

On math or programming problems, they are overfit to solving the entire thing end to end (presumably for benchmarks). I have had very poor results asking for pointers and hints that don't give away key insights. This has been the case across models I have tested.

An architecture with a "judge" that gates responses and ensures a lack of spoilers would probably work better. But this is a simple thing that they keep messing up.

Serious answer: no model ever gets close to writing an architectural floor plan that makes sense.

They understand all the rules and best practices, they can (sometimes) spot a bad idea in a floor plan, they can describe a good floor plan.

But ask them to make one, even if you give it every detail (even a "node graph" of rooms), they will still output nonsense. Same for text and image models.

Floor plans should be the new Pelican Benchmark.

Understanding human interaction nuance to an exact degree. For example, even when given all the scripts of the Seinfeld TV show, they still cannot come up with a new script does not feel as good as any of them (once they can, I want to watch these episodes..).