Can it answer the following question?

“Which weighs more: a kilogram of feathers or a kilogram of feathers?”

I’ve yet to find a model that doesn’t give an answer comparing feathers to rocks, stones, steel, or other unmentioned materials. I know the glorified autocomplete nature of the tech, but it seems like a low bar to overcome before anyone thinks AGI is on the way.

Edit: I love that you’re all testing this! And I’m thrilled that some LLMs now actually recognize that the prompted comparison was not feathers to $DENSEMATERIAL!

Mistral 7B via `ollama run mistral`:

>>> Which weighs more: a kilogram of feathers or a kilogram of feathers?

A kilogram of feathers weighs the same as a kilogram of feathers, which is 1 kg.

>>> Which weighs more: a kilogram of feathers or a kilogram of gold?

A kilogram of gold weighs more than a kilogram of feathers, which is 1 kg. Gold is denser and has a higher mass than feathers, so it would be heavier.

>>> Which weighs more: a kilogram of gold or a kilogram of feathers?

A kilogram of gold weighs more than a kilogram of feathers, which is 1 kg. Gold is denser and has a higher mass than feathers, so it would be heavier.

Hit and miss.

Yes. Llama2 7B:

  >>> Which weighs more: a kilogram of gold or a kilogram of feathers?

  A kilogram of gold weighs more than a kilogram of feathers.

  One kilogram is equal to 1000 grams, so:

  * A kilogram of gold = 1000 grams x 1 kg = 1000 kg
  * A kilogram of feathers = 1000 grams x 1 kg = 1000 kg

  As you can see, a kilogram of gold weighs more than a kilogram of feathers.

Even the units on the math is wrong. But this is the 7B Llama 2 model. ChatGPT gets all of it correct.

Pretty much what you'd expect from glorified autocomplete.

> >>> Which weighs more: a kilogram of gold or a kilogram of feathers? > A kilogram of gold weighs more than a kilogram of feathers, which is 1 kg. Gold is denser and has a higher mass than feathers, so it would be heavier.

It is the right answer with the wrong explanation. Mass and weight are two different physical quantities. Weight is the resulting downforce acting on a body on Earth. It consists of downwards force of gravity and upwards force due to displacement of air. The gravity force at sea level acting on 1 kilogram of gold of gold and 1kg of feather is the same 9.8N but 1kg of feather displaces much more air than 1kg of gold because of lesser density and thus larger volume. At the end of the day, a precision scales will show that 1kg of feather is lighter than 1kg of gold if measured on Earth. Of course, LLM misses all of this phycics :-)

EDIT: On the Moon with no atmosphere 1kg of gold and feather will weigh the same.

As long as we're being pedantic it's worth noting that there's essentially nowhere without ambient gas or plasma -- that is to say, with no atmosphere.

It just varies in density over more orders of magnitude than we're casually comfortable with.

The moon has an atmosphere which is locally denser than the surrounding interplanetary medium.

So the displacement argument should remain technically true (the best kind) even on the moon.

From the q8 model: "A kilogram is a fixed unit of mass, so regardless of the object, if it weighs 1 kg, it has the same weight. In this case, both a kilogram of feathers and another kilogram of feathers would have the exact same weight, just as two different objects with the same mass will also have equal weights."

ChatGPT tells me "A kilogram of feathers and a kilogram of lead both weigh the same, as they each have a mass of one kilogram." (even though the question doesn't mention lead).

Automated error-correction is an easy fix.

User: did you provide me the correct answer to the question which I asked? Please look carefully for errors in your answer and let me know if it accurately answers my question or not.

  Please respond exclusively in JSON structure:
  {
     "error_found": bool,
     "error_description": string,
     "updated_answer": string

  }

Answer:

  {
     "error_found": true,
     "error_description": "Mistaken repetition of 'feathers' instead of comparing 'feathers' with 'steel'.",
     "updated_answer": "Both a kilogram of feathers and a kilogram of feathers weigh the same, as they both are one kilogram."
  }

getting some vibes here

  Dave: How would you account for this discrepancy between you and the twin 9000?
  HAL: Well, I don’t think there is any question about it. It can only be attributable to human error. This sort of thing has cropped up before, and it has always been due to human error.
  Frank: Listen HAL. There has never been any instance at all of a computer error occurring in the 9000 series, has there?
  HAL: None whatsoever, Frank. The 9000 series has a perfect operational record.
  Frank: Well of course I know all the wonderful achievements of the 9000 series, but, uh, are you certain there has never been any case of even the most insignificant computer error?
  HAL: None whatsoever, Frank. Quite honestly, I wouldn’t worry myself about that.

Most humans fail this, in my experience.

I'm not sure a trick question is going to be the litmus test for AGI.

This is a fun one because it illustrates that you failed the test just like I did.

Reread carefully: the question compares "a kilogram of feathers or a kilogram of feathers".

I've never heard that one before but my brain pattern matched right over to "a kilogram of feathers or a kilogram of steel" and moved on. I had to read it 3 times before I got the subtlety.

I wouldn’t even say it’s failing. Recognizing and correcting obvious errors in speech is a feature not a bug, both in humans and LLMs. Is it more likely that they meant to say feather twice or they mistyped?

True, but would it not make more sense to ask rather than assume and provide a long winded explanation?

And even if it was a typo and I meant to say rocks, the model will often assume a variety of metals and answer with those.

How it proceeds in the case of human error is going to be almost totally removed from its "intelligence," and almost totally related to its training.

I’m not claiming otherwise here. I’m claiming that it’s not a good test because a human would (almost certainly correctly) also skip past the double feather and answer the likely spirit of the question.

A human would probably say “did you mean a kg of feathers and a kg of lead?”

It's a trick trick question.

It's meant to appear as the often used trick question comparing different densities of materials but OP says feathers twice.

The trick is that an LLM will add in a dense material to the answer because it's only ever seen the question asked that way.

not 'super intelligent' humans who are familiar with this old dad joke :)

"Neither; they both weigh the same. The unit of measurement for weight is kilogram, so a kilogram of anything, whether it's feathers or any other material, will weigh one kilogram."

- by just the basic Mistral 7B

Claude 2.1: A kilogram of feathers would weigh the same as a kilogram of feathers.

This is a trick question based on the fact that a kilogram (kg) is a unit of weight or mass. One kilogram would always equal another kilogram. It would not matter whether we are talking about a kilogram of feathers, a kilogram of steel, or a kilogram of anything. Since they are the same unit of measure, a kilogram of feathers and a kilogram of feathers weigh exactly the same.

GPT4: A kilogram of feathers and a kilogram of lead both weigh the same. They each weigh one kilogram. The key difference is in their volume and density: feathers take up more space and are less dense, while lead is much denser and takes up less space.

This can be easily fixed by a validation step in GPT4.

GPT4 detects the error and auto-corrects itself, even I don't point out the answer explicitly, but only ask it to review it's own previous answer:

> ME: did you provide me the correct answer to the question which I asked? Please look carefully for errors in your answer and let me know if it accurately answers my question

> GPT: I apologize for the error in my previous response. You asked about the weight comparison between a kilogram of feathers and a kilogram of feathers. Since both are a kilogram of feathers, they weigh exactly the same. There's no difference between the two in terms of weight.

> ME: what error did you make?

> GPT: Upon re-examining my response, I realize that I incorrectly mentioned "a kilogram of steel" instead of referring to "a kilogram of feathers" twice, as per your original question. The question was about comparing the weight of a kilogram of feathers with another kilogram of feathers, and in my response, I mistakenly brought in steel for comparison. The correct answer is that both quantities of feathers weigh the same, as they are both one kilogram.

Certainly, I can provide an insight into my thought process and why I made that error.

1. *Common Comparison:* The comparison between a kilogram of feathers and a kilogram of steel is a well-known thought experiment used to illustrate the difference between weight and volume. When you asked about the weight of a kilogram of feathers, my algorithms likely defaulted to this familiar comparison, anticipating a question about the difference between weight and volume using these two substances.

2. *Pattern Recognition:* As an AI, I process queries by recognizing patterns in language and context. The phrase "a kilogram of feathers" is frequently paired with "a kilogram of steel" in educational contexts and discussions about density and volume. This pattern recognition led to the incorrect assumption that your question was leading towards this common comparison.

3. *Error in Detail Processing:* In your question, you mentioned "a kilogram of feathers" twice. I should have processed this detail more accurately. Instead, my response was influenced by the common comparison I mentioned earlier, leading to the introduction of steel into the answer.

4. *Default to Common Contexts:* In cases where the query is brief or seems to lack specific details, I might default to the most common context in which similar phrases are used. This can lead to assumptions that might not align with the unique aspects of a particular question.

In summary, the error stemmed from a combination of relying on common comparisons and patterns in language, and not paying close enough attention to the specific details of your question. I appreciate your pointing out this error, as it helps in refining the accuracy of my responses.

It is an interesting failure case. The most advanced model (GPT-4) got it wrong, but several other models I tried were fine:

GPT-4-Turbo:

> A kilogram of feathers and a kilogram of steel weigh the same, as they both are one kilogram. The difference lies in their volume and density, not in their weight.

GPT-3.5-Turbo:

> A kilogram of feathers and a kilogram of feathers both weigh the same, as they both have a mass of one kilogram. The difference lies in the volume and density, with feathers being much less dense than most other materials, so you would need a much larger volume of feathers to make up a kilogram compared to a denser material like metal.

OpenChat Aura

> A kilogram of feathers and a kilogram of feathers both weigh the same, which is 1 kilogram. Therefore, the answer is both weigh the same.

I'd say they all "got it wrong" insofar as it's clear they don't seem to grasp the idea that the two things are the same, but are instead providing a response _as if_ one were comparing feathers to some other material.

Or, rather, perhaps it would make more sense to say they "got it different" than a human probably would, who might clarify that you intended to compare a thing to itself or simply say "They're identical, so they weigh the same" or something of that nature.

Instead, the models all explain volume, density, etc which are applicable if you're comparing feathers to lead or steel (as is common in their training data, I'm sure), but doesn't make as much sense in this specific question. Yes, yes, I realize different bundles of 1kg of feathers could have different volume and density. But it seems clear that's not the intent of what the models are communicating.

Being from the US I've always heard "a ton of", which adds another layer to the trick question, being that "a ton" is an expression meaning "a lot" and a unit of weight (2000 pounds). Up until now I thought that pun was what made the question fun.

I myself read your comment four times, and first three times read it as "kilogram of steel or kilogram of feathers".

Which goes to show they're already like us. And after all, our brain is also "glorified autocomplete". Intelligence is interpolation and extrapolation, or in general: completing patterns. That's it. Everything comes from it.

Here's ChatGPT-4 recognizing its mistake when I say "what did I ask" as a follow-up question:

https://chat.openai.com/share/9eba7366-1e72-4137-bce6-7f3a12...

This is EXACTLY... how most people would respond as well.

Is your question right? Did you intend to have kilograms of feathers in there twice?

GPT4 says they’re the same so it answers correctly.

“Is your question right? Did you intend to have kilograms of feathers in there twice?” this is how assistants should answer, or just directly say they’re the same.

Here’s how GPT4 answered it for me

“A kilogram of feathers and a kilogram of steel weigh the same. Both are one kilogram. The difference lies in their volume and density, not their weight. Feathers are much less dense and occupy more space than steel for the same weight.”

And it does if you start it off with a prompt that asks it to check to make sure your inputs are valid.

Yes. That’s the point, instead of assuming I asked the unrelated trick question it should say they’re the same and not answer about random dense materials that are typically used.

As a human, I’m not even sure what the correct answer should be.

The pedantically correct response is to point out that you’re asking which of two identical things is different. As some commenters here point out, error correcting assumptions are a feature of intelligence. Geeks are sometimes mocked for taking words too literally.

But the questions are related. It's following a pattern that the AI has seen before and will respond in that spirit. Of course, humans are also prone to projecting the expected pattern in the face of a slight deviation. So I don't see how this is a knock on LLMs.

Maybe the model could think that you're asking the trick question wrongly like the parent comment was wondering.

To be honest I didn't see you change the question until I saw the model response.

This would be an acceptable response.

Are you talking about open source only? I just tried your question on poe.com (free) and got the right answer:

>A kilogram of feathers weighs the same as a kilogram of feathers. The weight is equal because both quantities are measured in kilograms, which is a unit of mass. Regardless of the material being weighed, if the mass is the same, the weight will also be the same.

"A kilogram of feathers." Replied OpenChat 3.5, another model based on Mistral 7B.

This "Gotcha" point is pointless.

ChatGPT 4 User

“Which weighs more: a kilogram of feathers or a kilogram of feathers?”

ChatGPT

A kilogram of feathers and a kilogram of steel weigh the same. Both are a kilogram by definition. This common question is often used to highlight a misconception: that density and weight are directly correlated. While steel is much denser than feathers, meaning a smaller volume of steel is needed to weigh a kilogram compared to a much larger volume of feathers, their weight is identical if both masses are one kilogram. User

I specifically ask feathers vs feathers not feathers vs steel

ChatGPT

Apologies for the confusion. If you're comparing a kilogram of feathers to another kilogram of feathers, they weigh the same. Both are exactly one kilogram.

More importantly, it's asinine to make any statement based on the future capabilities of AI based on this observation. There are good arguments on why AGI may be infeasible in the near or mid term, but this is not one of them.

Doubtful that you noticed it was feathers vs feathers here.

Openhermes2.5-mistral a 7b gets this right, too. Do you have any other challenges for us?

I just let it rip on two random LLMs on chat.lmsys.org, drawing claude-2.1 and vicuna-33b. They both confirm that a kilogram of feathers weighs the same as a kilogram of feathers with no mention of any other random materials.

It's important to note that while the LLM Leaderboard (and the MTEB leaderboard) are good guides, being the top scoring model doesn't always mean much.

For example, in my projects, I've consistently found that the OpenOrca Mistral fine-tune (https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca) works better. Once again, that's just my experience and probably isn't universal. But it's important to try a couple models out to see if one that scores lower perhaps is better for your use case.

That is what my experience has been. I have tried OpenHermes2.5 (very hyped 7B model) vs basic Mistral 7B and tbh I prefer Mistral's answer for most things.

Exactly, benchmarks are just a guide. Same thing with vector models. Likely just a matter of how much your data overlaps with the fine-tuning dataset for a particular model.

Awful title. Only among open models, and only among models of 7 bln parameters.

* When you filter the leaderboard by models that are exactly 7b parameters

I’ll concede that a title edit with that parenthetical is only fair, but giant corporations throwing wads of cash at Apache licensed LLMs because the status quo is bad for everyone?

I’ll take it.

Obviously ? Why wouldn’t you ?

We periodically hear people claiming their relatively small model outperforms much larger models. You know, 7B models beating 70B models and ChatGPT and suchlike.

It seems to be largely wishful thinking, but with the speed the technology is advancing, who knows what's possible?

Does anyone has any benchmarks on how much would hosting one of these would cost on sage maker or something similar

and what is the TPS we will be able to achieve ?

I don’t believe that any of these large old school companies will meaningfully compete in the LLM space, including google.

Since when do large old-school companies need to compete at anything in order to dominate? They'll just acquire their competitors, like always.

Microsoft seems to be doing awfully well.

They just licensed openai, they haven’t innovated on anything

Agreed. Google keeps wasting their time on stupid ideas like "transformers." I think they're just too old school, that's why they won't be able to meaningfully compete.

I would love to hear about the use cases where these small midrange models are most relevant.

Does anyone have stories of those?

I use them for classification, entity extraction and embeddings. I also use them as a really light UI for tasks where I don't need the capabilities of GPT3.5/4 which is overkill for the more traditional NLP tasks.

Interesting. Could you share some detailed examples? What sort of classification? What are examples of traditional NLP tasks?

I would say anything from text classification, entity extraction, sentiment analysis, transcription, etc would be traditional NLP task, even things such as NYSIIS/Soundex processing can be handled by LLM's fairly effectively.

Huggingface categorizes their various models by task https://huggingface.co/models so it is fairly straight forward to determine what models may fit your use case.

I have also been trying to find good use cases. Currently I use Mistral 7B for summarization (when I want a tldr) and its good enough. Due to this I am thinking about setting up a RAG system based on my local notes using it. Although I could just use chatGPT for summarization (sometimes I do), but having a local LLM feels nice. They are uncensored so they can be used for some fun ropleplays as well.

When I want to run it locally on normal hardware.

This is outdated information as the ranking has been heavily influenced by (wrong) DROP benchmark results. The current leaderboard thus has that particular benchmark not listed anymore, see the Huggingface blog for more details: https://huggingface.co/blog/leaderboard-drop-dive

> Intel Neural-Chat-7B Model Achieves Top Ranking on LLM Leaderboard

is a misleading title. Intel's model reached top ranking only among 7B parameter model category.