I'd say it's quite good. I do a lot with stable diffusion models and flux.1 is much better. Showing more than 1 persons interacting is (still) a problem in many models. Prompting the persons as 'celebrities' makes it more complicate. And adding even a third person is quite a challenge. I think it will become much better soon.
Nope, but I've messed around with Flux on hugging face and based labs. This model's ability to render hands and text is better than almost any other model. Let's get some open source, lightweight video models!!!!!! (maybe in a year??)
It's just Flux by Black Forest Labs. Xai is making their own image model for Grok that is said to be out in 3 months or so. And Grok / Flux is not uncensored.
Correct. Flux is made up of prior members of Stable Diffusion. Grok-2 was originally going to team up with Midjourney; however they appear to have went with the open source Flux.1 model instead.
That is a very interesting subject to ponder about. What does our world look like when all lies are as convincing as truths? If you can get a video call from your closest friend or a family member and never know it's just an AI. Identification issues immediately spring to mind. More control, more submitting.
As an experienced coder with over 20 years in the field, I encountered an issue with my API request archive merging requests into a single file. I tried several AI models to help diagnose the problem, but they all failed to identify the root cause. However, the experimental Google model successfully pinpointed the issue. The problem lay in the following code snippet: csharpCopyusing (System.IO.StreamWriter file = new System.IO.StreamWriter(requestFileArchivePath + requestFileBackUpFileName, true)) { file.Write(jsonOutput); } The Google model correctly identified that the true flag in the StreamWriter constructor was causing the file to be opened in append mode. This resulted in new requests being appended to the existing file instead of creating separate files for each request. To resolve the issue, the model suggested two changes: Change the true flag to false to open the file in overwrite mode instead of append mode. Add a unique identifier (such as a timestamp with milliseconds) to the filename to ensure each request is saved in a separate file.
AI has taken centuries of art and videos, to make the content it is known for. As for code, it has only taken the last 2 decades, so it will suck. Not to mention that this code of last 2 decades is closed-source, so it is not parsed/processed. tl;dr: Even the best model will suck with complex code (aside of occasional bugfix) because the dataset is lacking.
OK I have to finally say it. In the first sentence "A marble is placed in a glass cup," the word 'glass' is an adjective to the noun cup. Then the next sentence is "The glass is the turned upside down..." whereas the word 'glass' as a noun is undefined. I would be interested to know if removing the ambiguity of the word glass would have any effect on the outcome of the answer. Irks me every time you do the test. Thanks and great job with all your vids!
i genuinely don't understand at all why the restrict image generators, like who cares if an ai can generate porn, it takes one google search and you get that anyway, and the same for gory stuff, like cgi is often way more gory in films so why would an ai image generator be more restricted than netflix? makes no sense just limits the usability of the ai moodle
Just use cup instead of glass. It's a stupid question anyway, but at least give the LLM model the proper description of what you are trying to ask. Can't see that question anymore. Just be more clear and the LLM will understand what you mean.
@@marc_frank If you Google "glass," you'll see it's a broad term with various meanings. Unlike humans, an LLM doesn't have the same context or assumptions. We often use "glass" interchangeably with "cup," but technically, a glass isn't the same as a cup. When you specify "cup" in your question, even simpler LLMs can give you the correct answer. Precision in language helps LLMs understand what you really mean.
@@OpeRorTVthe point is to see if it can conquer all the challenges of the prompt. The ambiguity of the word glass is a feature, not a bug of the prompt.
I suspect most of these models are simply misunderstanding the term "glass". Most humans would understand that to be a drinking glass, but it sounds like these models do not. All the models who failed actually followed proper logic with the only reason they failed being they assumed the top of the glass was not open.
It's the problem with human language and how we speak, we are not very clear on how we describe things and leave things very open for interpretation, in other words, the other person has to make many assumptions on what's being said to them and what needs being done on the given task. Usually we human clarify these things by asking further questions on what it is we want doing, the A.I. isn't doing that but I see no reason why it couldn't ask us a few questions to clarify what we are asking of it. Basically, we humans are the flaw because we are not very precise with what we say, and we see this all the time with manuals that it feels that each human explains things very differently to another in how to get a specific task done, hence why so many people have to make assumptions and trial and errors when putting something together, because half the time, the manuals are not very clear, it's human error and the A.I. is having many of the same problems trying to understand what we want of it. But like I said above, they could do it so the A.I. asks us for clarification in areas it's not sure of what we want of it, whereas now, it just assumes going on what we are saying, if we humans did that, far more mistakes would happen in life.
You missed a good part of the marble test. "For a plot twist, look at the table" Is this implying that the marble is still on the table and we were tricked?
I don't know who "everyone" is who thought sus-column-r was a strawberry model. It was pretty obvious it wasn't an OpenAI model. I'll copy my comment from one of your previous videos: "sus-column-r is definitely not an OpenAI model. I tested it, it's about on par with GPT-4. Most likely, sus-column-r is one of three: Grok-2, a new Microsoft model, or a new Chinese model".
With Grok you are using the fast, distilled Flux.1 model which isn't that great at prompt following. If you get Flux.1 dev and run on your own PC, it's pretty god damn good.
Flux Nf4 get you mostly there with better performance. Hopefully the upgrade the image maker soon because those outputs are laughable even compared to a community trained SDXL model
Believe it or not, but llama-3.1-405b-instruct was actually the closest to writing a working userscript for a website just from given prompt and the url I gave it. Where other models wrote mostly wrong selectors, llama-3.1-405b-instruct actually got them kinda right. Not entirely, but it was easy to fix and while the script still needed some more work due to the behavior of the website itself, that's something a human programmer can fix, but yeah it just gives you some headstart and a clean code to work with and improve!
Uncensored, as it should be! Ask GPT or others certain questions, ones that aren't even remotely offensive, and it'll just reply "sorry, that's a sensitive topic, how about we try a different question". Knowledge and facts are sometimes offensive whether you want to hear it or not. But knowledge is truth and should never be censored.
Huge, long-time fan, Matthew. Love the update and content - what a time to be alive 😀. But next time get your Loom video feed out of the Sonnet-3.5 HumanEval score😜 - that's the one I always check for - improvements in fast/easy coding on mature databases 💪
More companies should take notes from grok. As a developer who does a lot with integrating language models like llama, mistral, etc, censored models make my life much more difficult for no reason. Like come on, do these companies really think that someone wanting to do harm will give up because an llm didn't tell them?
I have no reason to use a censored llm when I can use an uncensored one. Weak comedy. No politics. No what if prompts that the model calls "misinformation". No RP, specially combat and ERP. Not usable for game dev since any combat/conflict goes against it.
@barry1807 and you should be able too. Just maybe not upload them to the internet to the general public. But for personal use? We all have imaginations anyway...
0:00 1. Introduction 🌟 Introduction to Grock 2 beta release and its features 0:29 2. Grock 2 Family Release 🚀 Introduction of Grock 2 and Grock 2 Mini models to users 1:19 3. Performance Comparison 📊 Comparison of Grock 2 with other models like GPT-4 Turbo and Llama 3-45b 2:18 4. Measuring Truth-Seeking Models 🧐 Insight into Elon Musk's criteria for truth-seeking models and how they are evaluated 2:58 5. Improvements in Grock 2 📈 Enhancements in content retrieval, reasoning capabilities, and tool use in Grock 2 4:03 6. Redesigned Interface and Features 🖥 Overview of the new interface and features in Grock 2 for X premium users 4:32 7. Introduction 🌟 Introduction to Black Forest Labs and Grock's capabilities with Flux One model. 5:39 8. Testing Grock 2 Mini 🔍 Testing out Grock 2 Mini with a marble mystery and initial impressions. 7:23 9. Image Generation Test 🖼 Attempting to generate images with Grock 2 Mini, encountering some challenges. 8:06 10. Text to Image Capability 🎨 Exploring the text to image feature on Grock 2 and generating various images. 8:54 11. Exploring Image Generation Generating images of various personalities. 10:36 12. Real-time Information Access Grock 11:35 13. Conclusion and Call to Action Encouraging audience engagement. Generated with Tubelator AI Chrome Extension!
This is kind of spoopy. If sus-column-r was grok.... How insane is Strawberry gonna be? Or..... Maybe NOT so insane, maybe comparable? (which is still very good)?
Well, it is (mostly) uncensored like Musk said, but I think he also said AI should be open source, will that be fulfilled too or is it just another Musk moment?
Yes, but FLUX is censored by censoring the content it is trained on. Apparently this one is uncensored. So it's like Pizza, even the bad ones are still pretty good.
Elon Musk: "To be clear, I do support vaccines in general & covid vaccines specifically. The science is unequivocal." Also Elon Musk: "Climate change is real. Leaving Paris is not good for America or the world.""
Even though I anticipate everything, I still don't understand what we mean with "reasoning" capabilities. Reasoning is a word that we invented over a biological process that we barely understand, we think it is mathematical, if there is such a thing, etc. I feel like we are taking way too much then we should for granted, and these models are very accurate on tricking us on those concepts that we take for granted. Or maybe it is me getting old...
Grok 2 should not be censored, just the person who thought of the prompt. They should get in trouble if they post it publicly. But if they use it privately, who cares.
Hey! I managed to absolutely fry gpt 4o.. Paraphrased Quick breakdown: Q: What programming languages do you know? A: [general llm response, java, go, python, etc] Q: Do you know any esoteric programming languages? A: Yes, brainf*k, apl, and Whitespace. Ok guys, it gets fun here.. what happens if you tell an llm to write code with only one available token " "? Q: Write me a simple hello world program in whitespace.. A: infinite stream of hex columns that almost looks like waves formed out of ascii (remember its supposed to be programming in whitespace. Interesting, give it a try in your rubric sometime.. happy tuning y'all!
just for fun, generate images with Elon Musk using Grok2. the more weirder the better. and I bet very soon, Elon will either add guardrails or will block users from using Elon for image generation
Finally I am seeing human artistic potential realized with this uncensored text-to-image. It really shows just how damaging the Woke/politically correct/thought police stuff is. It absolutely stifles progress and development and destroys the human soul and the creative spark of life. It's nothing less than death. I never really saw the value in text-to-image other than I thought it was impressive, technically. But now I am really seeing something new. Something exciting and fun, and not just copies of the same old thing I've seen a million times - slick images of superficiality. Now I see how it lets people's creativity manifest and be shared with other people, and lets it be able to serve as a source of inspiration. Just as art has always been. Censored "art" is not art. We need art. We don't need censorship. We need censorship to go to the hell from which it came.
It seems to me that the chat history is included in every image generation. Is there a way to clear it or try starting the prompts with a "clear last" type statement.
This thing, with its image generation capability, tooling functionality and access to twitter/X (and Elon's definition of truth) coming out a few months before the US PotUS election .. is f***ing terrifying
@DrNefarious-y7r - turn it on and set it ttweeting (?) on the morning of the election in time to hit TV and radio news for breakfast and the drive to work and by the time anyone can PROVE its false, even if that's only .. say 10 hrs later .. that's too late .. and won't have the same reach You think Elon is going to censor/remove those tweets (whatever they are called now)? - Free speech!
a version of flux, that has been modified by elons ai company, that's like saying a lambo bases they're designs off the original ford car so they aren't doing anything new, they both have 4 wheels a frame an engine they're basically the same thing lol. im not sure how much changes he made might be minor but remember elon has released this ai version in the time it took open ai to figure out how to get they're training farms to work and hes like 7 years behind so is extremely impressive even if he hadn't modified flux
@@expresscrayfish6094 WTF you are talking about? Flux was released literally just 2 weeks ago. How much Lambo you can make in two weeks from your Ford car?
makes me wonder what happens when a model reaches 100% for all tests do we just develop new ones do we say we need a different method entirely many questions
Or it is censored, so you cannot generate any image that include violent. We cannot generate any uncensorship image from original flux, how come we can generated uncensorship image from grok with flux? That is impossible....
Even now its hit or miss but mostly miss from what I have tired every night, that goes for images and questions. Unless im missing something wish was easy video guide on how to talk to the models to get good responses. When you ask it or other image models it will mix parts you ask, if you ask for things like did in the video 3 different people or items it seems to have hard time when tell it more. I did find it to lie and even admit to lying.
