News for hackers. Stuff that matters.
Re:Gemini Ultra Released by rvnx on Thursday February 08, 2024 @07:49AM (#39301328)
Proof of Gemini cheating: https://i.imgur.com/eYJDFjS.png

Answer about cherries falling from the sky...

(there is no question or context beforehand, this is the first question of the chat)

Re:Gemini Ultra Released by alphabetting on Thursday February 08, 2024 @08:13AM (#39301547)
That's a bummer. I just made this one up which GPT-4 failed and Gemini got right but definitely need to do more testing to see what's being taken from online and what is actual reasoning.

https://i.imgur.com/3sNr3LW.png https://i.imgur.com/EIj0nZg.png

Re:Gemini Ultra Released by nycdatasci on Thursday February 08, 2024 @08:42AM (#39301795)
Great example! Thanks for sharing. I tested the same prompt using a variety of GPT-4 model versions and it consistently got the answer wrong. When I continued with "use algebra/formulas to confirm your answer" it corrected the initial mistake, but I'm surprised that it's not doing that out of the box.
Re:Gemini Ultra Released by mewpmewp2 on Thursday February 08, 2024 @09:03AM (#39302002)
I tried through API with latest GPT-4-Turbo-Preview and this is what I got:

------

Let's break down the information given:

The school is two blocks north of the pool. This means if you are at the pool and move two blocks north, you'll arrive at the school.

The convenience store is one block south of the school. Therefore, if you start at the school and move one block south, you'll reach the convenience store.

Based on the first two statements, the relative positions from north to south are: School, Convenience Store, Pool.

Thus, the third statement saying "The pool is south of the convenience store" is indeed true, based on the positions provided by the first two statements.

-----

But it starts weirdly messing up, when I asked it to explain it using coordinates.

---

Let's use a simple coordinate system to represent the positions of the school, the pool, and the convenience store. Let's assume each block corresponds to one unit in this system. We'll start with the pool as our reference point at coordinates (0,0). Remember, moving north increases the y-coordinate, while moving south decreases it.

The school is two blocks north of the pool. If the pool is at (0,0), moving two blocks north puts the school at (0, +2).

The convenience store is one block south of the school.

If the school is at (0, +2), moving one block south puts the convenience store at (0, +1).

Given these positions:

The pool is at (0, 0).

The school is at (0, +2).

The convenience store is at (0, +1).

Therefore, instead of the pool being south of the convenience store, the pool is actually located to the north of the convenience store if you consider "north" to mean increasing y-coordinates. The convenience store, being at (0, +1), is southward relative to the school but not to the pool. The pool, being at the origin point (0,0), is south of both the school and the convenience store in terms of their y-coordinates. Thus, the third statement, "The pool is south of the convenience store," is false according to the given coordinates.

---

It gets coordinates and everything right, but still somehow messes up North and South.

Re:Gemini Ultra Released by Alifatisk on Friday February 09, 2024 @06:07AM (#39313636)
You know you can share conversations right?
Re:Gemini Ultra Released by aaronharnly on Thursday February 08, 2024 @09:09AM (#39302066)
Nice one! Gemini got it wrong in my (one) test:

https://g.co/gemini/share/4bae8ca3dd0a

Re:Gemini Ultra Released by microtherion on Thursday February 08, 2024 @10:30AM (#39303054)
Gemini happened to get the yes/no answer right, but the reasoning it gave is completely wrong, so it should not get any credit for it.
Re:Gemini Ultra Released by rvnx on Thursday February 08, 2024 @07:56AM (#39301394)
In addition just realized it thinks Apple is only one syllable.
Re:Gemini Ultra Released by dev1ycan on Thursday February 08, 2024 @08:22AM (#39301624)
I genuinely believe LlMs have for the most part peaked already just like the previous iterations of ML and even brute forcing.

So much money being poured into AI startups just for them to have to resort to cheating to improve their results a tiny bit and claim advancements.

Re:Gemini Ultra Released by snapcaster on Thursday February 08, 2024 @09:05AM (#39302031)
I'm confused on how this is "cheating" isn't it just getting the answer wrong?
Re:Gemini Ultra Released by rvnx on Thursday February 08, 2024 @09:14AM (#39302122)
It's answering with "cherries", though "cherries" were never mentioned anywhere in the question since the task was to choose between "apples" and "pears" this time,

and not "cherries" and "pears" like the example found on the internet.

Re:Gemini Ultra Released by deltaburnt on Thursday February 08, 2024 @09:39AM (#39302420)
I agree with who you're responding to. Cheating, to me, would imply that there's some sort of hard coded guiding to the LLM. This just seems like typical LLM hallucinations?
Re:Gemini Ultra Released by sp332 on Thursday February 08, 2024 @10:25AM (#39302986)
It's cheating because it has memorized the answer to the puzzle instead of using logic to solve it.
Re:Gemini Ultra Released by room500 on Thursday February 08, 2024 @01:48PM (#39305865)
I thought that is essentially what LLM's do? They learn what words/topics are associated with each other and then stream a response.

In some ways, this is proof that Gemini isn't cheating... It is just doing typical LLM hallucination

Re:Gemini Ultra Released by sp332 on Thursday February 08, 2024 @02:45PM (#39306652)
Well, sometimes. Sometimes not. https://arxiv.org/abs/2310.17567
Re:Gemini Ultra Released by justanotherjoe on Thursday February 08, 2024 @11:20PM (#39311340)
Llm's can also do some exploring based on combinatoric play of learned elements.
Re:Gemini Ultra Released by block_dagger on Thursday February 08, 2024 @02:12PM (#39306182)
Your concept of cheating is simply how LLMs work.
Re:Gemini Ultra Released by esafak on Thursday February 08, 2024 @06:17PM (#39309144)
It is not. LLMs do not just memorize; they also extrapolate, otherwise they would be useless. Just like any ML model.
Re:Gemini Ultra Released by ajross on Thursday February 08, 2024 @09:42AM (#39302458)
I don't understand the leap to "cheating" either. LLMs aren't abstract logic models; they don't promise to reason from first principles at all. They give you an answer based on training data. That's what you want them to do. That they have some reasoning features bolted around the inference engine is a feature companies are rushing to provide (with... somewhat mixed success).
Re:Gemini Ultra Released by bloopernova on Thursday February 08, 2024 @09:57AM (#39302632)
Thank you for answering a question I had half formed in my head.

Do LLMs have logical rules built in? What makes them different to a very advanced Markov chain?

Are there any models out there that start from logical principles and train on top of that?

(Apologies for poor understanding of the field)

Re:Gemini Ultra Released by ajross on Thursday February 08, 2024 @10:02AM (#39302704)
> What makes them different to a very advanced Markov chain?

Really nothing. There's some feedback structure in the layers of the model, it's not just one big probability table. But the technique is fundamentally the same, it's Markov, just with the whole conversation as input and with billions of parameters.

Re:Gemini Ultra Released by og_kalu on Thursday February 08, 2024 @09:29PM (#39310723)
>Do LLMs have logical rules built in?

Handcrafted by humans ? No

But it's still possible to learn such rules from the data in an effort to complete the primary objective (predicting the next token)

Re:Gemini Ultra Released by justanotherjoe on Thursday February 08, 2024 @11:37PM (#39311457)
There's no logical rules built in at all. But Transofmers architecture is specifically trained to learn combinatoric play and rules of engagement from the data, so it can extrapolate and do cool, new things, that are not in the training data. In a way, you give them a chess board, the rules of the game, and then it can play. You don't teach them every possible board state. What's interesting is with significant amount of parameters it seems to encode more and more abstract and human-like understanding of the 'elements' at play and the 'rules of engagement' on top of them.

Edit: Not native. I'm not sure 'rules of engagement' is the correct english term here.

Re:Gemini Ultra Released by bloopernova on Friday February 09, 2024 @08:34AM (#39314686)
I understood you just fine, your English is great!

Thank you for the explanation. It seems like the LLM "plays" to learn? That's very cool, thank you again.

Re:Gemini Ultra Released by justanotherjoe on Thursday February 08, 2024 @11:25PM (#39311374)
This is not hard to understand. LLM can solve never before seen logic puzzles. This specific one proves that it HAD encountered this before, proving it was not doing anything emergent, but just basic remembering. Worse, it's not even reading the prompt correctly.
Re:Gemini Ultra Released by summerlight on Thursday February 08, 2024 @02:03PM (#39306073)
With some shallow understanding on how those models work, this looks much more like usual hallucination likely due to sparse data around "Tabitha" and "Cherry" which makes a single training data much more representative. If you try some common names like "Emily" or "Sarah", it will just do the job.

If you're trying to make the case that it's cheating because it is already in the training set, then you probably should come up with different questions. This is machine learning 101.