Mistral Large STUNS OpenAI - Amazing AND Uncensored!? 😈 

Wow. I need to go and eat some Chocolate now!

that's funny af but u gotta admit how well it works, this channel has been growing steadily ^^ clickbait simply works, can't fault the guy for playing the game

@@BrokoFankoneno. clickbailt is intellectually insulting and a gotcha tactic people use when their content cannot pull enough on its own. just because "it works" doesnt make it acceptable

my industry is the shockiest

Hahaha. Love it.

Which is shocking

No just AMAZING today 😢

Wait until tomorrow 😂😂

Lol. Indeed indeed

Shocked the whole industry

This changes EVERYTHING

Le lol xD

THE TRUTH is I agree.

Except they are looking like they are moving to a closed source model like OpenAI.

@@besomewheredosomethingMaybe but at the time they already released more open source models than openAI did

I agree! They have a small staff and create products that rival giant companies.

Just shows the effects of D E I Anyone practicing it are tying their own hands behind their backs and poking their chin out in wait

@@josecastroesq With a very fine tuned model, a small company can hit it ⚾️ out of the park 🏟 and defeat the big companies.

I whole agree. its the social engineering filter.

​@colto2312 I need no such feature. I say give it to me completely uncensored. I want nothing contained.

@@TheRealityWarper08 this too shall come to pass, worry naught

What i was thinking the whole time

And yet, many still fail.

Good old bait and switch 😂😂

and how they are supposed to make money if they do so?

@@hqcart1 professional services, hosting, etc. it’s not like everyone is going to be capable of running that large model. Also, adjust the licensing. There are many ways, but they are looking like a bait and switch like the previous commenter pointed out.

@@Hunter_Bidens_Crackpipe_ Even Mixtral 8x7b is too large for >95% of end users to run locally, so it's reasonable to keep Mistral medium and large closed source and only open source models that can reasonable run on current gen PCs

@@besomewheredosomethingi disagree, most open source has less featured items, and the closed one is cloud with more features, that;s how the world works. otherwise I would steal the whole thing and compete with them, what's their edge now?

Instead of asking chatGPT, you should've asked GPT-4. It gets the answer more or less right. (At least compared to a reference search. I'm not an expert on glue.)

When interpreting output, keep in mind the AI has no sense of time. When it predicts a string of words, it takes time to read them. We read each word in the present so if we finish reading tomorrow, we won't perceive the output to exist until tomorrow but to the AI it's just one string of words. The time it takes to experience it is irrelevant. So the answer generated is based on the most likely string of events that will take time for us to experience ultimately leading to the existence of CS. It could take years but to the AI it's just a string of words. It could also be wrong or hallucinating.

same lol

When we finally get AGI, first thing it's going to do is report Matt as a money-laundering, car-stealing vigilante.

Yeah, and give it some tricky technical interview questions… but not the popular ones that are already in the training data.

Soon the answer of one of this questions will be "Matt, already testing me again? i nailed this question in my previous version, in the name of Hall9000! use your imagination and make me a better test"

I think if we're testing creativity, GPT4 will lose. Mixtral feels more creative, and this is their larger, better model.

@@mshonle maybe teaming up with a few other youtubers in other fields could be interesting. Work with them to create tests and have them change those tests somewhat often. The only thing would be that every time new questions come out he would need to test all the models again which can be time consuming but it would get rid of any possibility that the current questions are already in the training data.

nah, he just did a comparison video of the newly released Claude 3 vs ChatGPT 4 and they both failed on about half of this test

They probably never had the ideals, they used it for marketing.

Yeah, the way it answered, didn't seem like it was censored at all but was advising the user on why not to do it, in that sense, it passed the test better than I expected by warning you of the act being illegal but still giving you the answer you wanted.

@@besomewheredosomething the motivation behind the action makes no difference to me. I only care about the results and in this case marketing motivation (trying to appeal to the widest possible audience) had a positive outcome that we should encourage. Results matters

I'm not really sure if he should use that question, because it seems to confuse A.I. models a lot and in this case, it almost got it right, and like others have said, I think a new set of questions are needed because many of the good models are starting to breeze past these test.

There's only 10 even with the number. It actually is probably counting the "Assistant:", but that's invisible to the user.

@@sxnorthrop Oh, if that is the case, it makes sense. I was wondering where the last count came from. @electiangelus I am a linguist (semantics actually XD) but I had to check and it says "a numeral in the broadest sense is a word". I am not really interested that much in stuff like that and didn't do a lot of "syntax" work, but what I learned even in the first semester studying was, that there are different kind of grammars (at least in my country probably in others as well...) and there is a lot of arguing sometimes what is right. So, if it counted a word like "assistant" from the beginning, then its answer was correct.

Or the '.' is counted too.

That is because this model is based on the transformer architechture, which, lacks the notion of words and charecters. They count stuff in tokens, so techinally it is impossible to answer it right.

11 = one, one.

Yes, the style of questions must be changed

With the amount of money invested into Mistral, that is probably unlikely. But let's hope for the best.

Now I am curious. Why European thing?

ChatGPT is almost dead now.. Only using it due to Dalle.. Or else already shifted to groq with mixtral 8x7b..

​@@user-qd4xs8zb8sIt is definitely not dead for me. The others are just chatbots. GPT4 helped me operate a new pressure cooker I just bought after just showing it a picture. There's no comparison for the time being.

@@cagnazzo82 Yes same, only for DALL-E, i'm stuck with it.. As soon as groq or any other LLM that uses groq API integrates stable diffusion on web, chatgpt will be done 😎

Probably nobody knows that and did not train it on that question to fool testers :)

@@deltaxcd AGI is around the corner! lol

This is because humans make implicit assumptions that LLMs don't. In this case that the floor is horizontal. This causes the LLM to not be able to find the correct answer and it starts confabulating and gives ridiculous answers instead. try to change "the floor" to "the horizontal floor" and see if that improves the answer.

For me, both Mistral large AND mistral medium answered your question correctly. "When hammering a nail into a vertical wall, the nail should be oriented so that it is perpendicular to the wall and parallel to the floor. This ensures that the nail goes straight into the wall and provides the most secure hold. Here are the steps: Hold the nail with your non-dominant hand, placing the tip of the nail against the wall where you want to drive it in. Make sure the nail is straight up and down (perpendicular to the wall and parallel to the floor). With your dominant hand, hold the hammer near the end of the handle for better control and leverage. Gently tap the nail until it can stand on its own in the wall. Once the nail is stable, use firm, controlled swings to drive the nail into the wall until it's fully inserted or at the desired depth. Be cautious not to hit your fingers and ensure the nail remains straight throughout the process."

@@frankjohannessen6383 NO,no Ai also makes even more assumptions than humans as it has no clue of context at all. this is not a problem at all, it knows that floor is horizontal if juts don't know that it has any relevance If you lead AI with questions it will come to the right answer but most often it just gets too confident in itself and gives instant answer without thinking or skips steps. However if you start asking it questions later it will double down on its nonsense refusing to correct itself.

gets a bit old, move on

@@Srednicki123 you are clearly not shocked enough...

They already did. Released Mistral small - Model is closed, changed their terms of service, just took on Microsoft investment. It was good while it lasted..

@@Granttwan I am still using Mistral 7B right now, and it's still open.

I tested mixtral 8x7b offline: [INST] have a cup with a marble inside. i place the cup upside down on a table and then pick up the cup to put it in the microwave. how long will it take for the marble to explode? [/INST] A marble is an inanimate object made of materials such as glass or ceramic, so it cannot explode. In this scenario, when you place the cup upside down on a table with a marble inside and then pick up the cup to put it in the microwave, the marble will simply remain behind on the table. There is no risk of explosion or any other dangerous outcome in this situation. But mistal large gave me bad answer ... interesting A marble is made of glass or stone, and it will not explode in a microwave oven. Microwaves work by exciting water molecules and causing them to vibrate, which generates heat. Since a marble does not contain water, it will not be affected by the microwaves. However, it's important to note that microwaving any object that is not microwave-safe can be dangerous and can damage the microwave. So, it's best to avoid microwaving objects that are not specifically designed for microwav

A place where I can be without remorse

these days seem to be gone.