[tessierashpool9]

Asked for a solution of a photographed Ubongo puzzle: https://gemini.google.com/share/f2619eb3eaa1

Gemini Pro neither as is nor in Deep Research mode even got the number of pieces or relevant squares right. I didn't expect it to actually solve it. But I would have expected it to get the basics right and maybe hint that this is too difficult. Or pull up some solutions PDF, or some Python code to brute force search ... but just straight giving a totally wrong answer is like ... 2024 called, it wants its language model back.

Instead in Pro Simple it just gave a wrong solution and Deep Research wrote a whole lecture about it starting with "The Geometric and Cognitive Dynamics of Polyomino Systems: An Exhaustive Analysis of Ubongo Puzzle 151" ... that's just bullshit bingo. My prompt was a photo of the puzzle and "solve ubongo puzzle 151"; in my opinion you can't even argue that this lecture was to be expected given my very clear and simple task description.

My mental model for language models is: overconfident, eloquent assistant who talks a lot of bullshit but has some interesting ideas every now and then. For simple tasks it simply a summary of what I could google myself but asking an LLM saves some time. In that sense it's Google 2.0 (or 3.0 if you will)

[dktp]

Deep research, from my experience, will always add lectures.

I'm trying to create a comprehensive list of English standup specials. Seems like a good fit! I've tried numerous times to prompt it "provide a comprehensive list of English standup specials released between 2000 and 2005. The output needs to be a csv of verified specials with the author, release date and special name. I do not want any other lecture or anything else. Providing anything except the csv is considered a failure". Then it creates it's own plan and I go further clarifying to explicitly make sure I don't want lectures...

It goes on to hallucinate a bunch of specials and provide a lecture on "2000 the era of X on standup comedy" (for each year)

I've tried this in 2.5 and 3. Numerous time ranges and prompts. Same result. It gets the famous specials right (usually), hallucinates some info on less famous ones (or makes them up completely) and misses anything more obscure

[geon]

I tried asking for a list of the most common gameboy color games not compatible with the original dmg gameboy. Chatgpt would over and over list dmg compatible games instead. I asked it to cross reference lists of dmg games to remove them and it ”reasoned” for a long time before it showed what sources it used for cross references, and then gave me the same list again.

It also insisted on including ”Shantae” in the list, which is expensive specifically because it is uncommon. I eventually forbid it from including the game in the list, and that actually worked, but it would continue mentioning it outside the list.

Absolute garbage.

[tessierashpool9]

I mean, isn't that a little ridiculous? Aren't those language models already solving complicated exam questions and mathematical problems?

[geon]

According to the creators, the models are on a phd level of intelligence, but they can’t get the simplest thing right.

[tessierashpool9]

Overselling is only the tip of the iceberg. The real problem is that a lot of managers base their decision to introduce language models into business processes on cutting edge Pro edition demos, but what is, of course, actually used in production is some cheap Nano/Flash/Mini version.

[DANmode]

Too easy.

[Workaccount2]

LLM's are bad at anything with images.

There is something fucky about tokenizing images that just isn't as clean as tokenizing text. It's clear that the problem isn't the model being too dumb, but rather that model is not able to actually "see" the image presented. It feels like a lower-performance model looks at the image, and then writes a text description of it for the "solver" model to work with.

To put it another way, the models can solve very high level text-based problems while struggling to solve even low level image problems - even if underneath both problems use a similar or even identical solving frameworks. If you have a choice between showing a model a graph or feeding it a list of (x,y) coordinates, go with the coordinates every time.