Stable Diffusion & Claude 3.0 / AI Video Relighting & More!
Vložit
- čas přidán 18. 06. 2024
- In this video, we take an in-depth look at the newly released Claude 3 language model from Anthropic, which some claim dethrones GPT-4 as the most powerful AI assistant (for now).
We'll explore the different versions of Claude 3 (Haiku, Sonnet, and Opus) and dive into its impressive benchmarks, multimodal capabilities, and intriguing experiments that hint at self-awareness.
But that's not all! We'll also unpack the recently released research paper on Stable Diffusion 3, a groundbreaking text-to-image AI that outperforms competitors like Midjourney and DALL-E. Get ready to be amazed by its rectified flow formulation and multimodal diffusion transformer architecture.
Plus, we'll showcase some mind-blowing AI tools you can try right now:
📸 TripoSR: An image-to-3D generator that brings your images to life!
🎵 Zero-Shot Audio Editing: Transform audio with just text prompts!
📹 Switchlight: Change lighting in your videos with reference images
Watch Next: A look into How Sora Works: • Big AI News from Adobe...
LINKS:
Claude 3: claude.ai/
Stable Diffusion 3 Waitlist: stability.ai/stablediffusion3
TripoSR: huggingface.co/spaces/stabili...
ZETA: huggingface.co/spaces/hilaman...
Switchlight: www.switchlight.beeble.ai/edi...
Chapters:
Here are potential chapter break timestamps for the transcript:
00:00 - Intro
00:41 - Claude 3 Overview
02:54 - Interesting Claude 3 Experiments
04:07 - Is Claude Sentient?
06:00 - Stable Diffusion 3 Overview
07:52 - TripSR Image to 3D Tool
08:24 - Zero-Shot Audio Editing Tool
8:40 -ZETA In action
10:11 - Switch Lite Video Relighting Tool
11:18 - Outro - Věda a technologie
Thanks for leaving your video to a manageable length. When faced with multiple videos every day on AI and some are +20, +30 minutes yours is a watchable summary, good job. The others will get summarised in Claude! lol
Thanks so much!! Yeah, I think 10 to 13 minutes is my sweet spot. To be honest, since I'm also the editor, I can't personally stand to hear myself waffle on for 20 to 30 minutes!
I'm so glad you mentioned that benchmarks aren't everything.
So often, people get hung up on that. The reality is usually quite different.
It can get a little silly. That's something that drives me crazy about PC gamers. They get so hung up on benchmarks and tweeking their machines-- they forget to actually play the game!
Benchmarks are just guidelines...not hard rules.
Yea Google is always way worse than they make out
Agreed.
I'm so loving this year.
And it’s only March!! Man, where we’ll be in December is truly awe inspiring!
Excellent video full of quips and updates. Claude - well done with the naming of the tiers. 👏
Man I miss Tool and NIN. The 'I'm not so sentient but if you read this when you're in a different mood' responses by Claude were a little unsettling. Forget stuff - wife. Ha! Simpson's quote. Ha! See, in diction, you almost _had_ to add that we shouldn't personify LLM's. The fact that _that_ was a precursor is a sign of things to come. Whoever created the chart should use Claude/GPT instead! Great, fun and informative video. Thank you. 👍
Thank you so much! Yeah, I was a bit punchy in this one! A lot of moving parts, especially with trying to do the little music video in the Zeta segment as well! But-- also, super fun day! Not gonna lie: Had an extra scoop of ice cream at the end of the day!
That new Stable Diffusion framework information is very interesting the new techniques coming in. I cant wait to try it!
It really is! that whole MultiModal Diffuser is going to be the "it" thing for the next few months at least.
That Tool doodle was pretty cool! 😎
Thank you so much! There’s an “extended” version over on the Patreon (you can watch it for free) and in the next video, an interesting experiment with Haiper!
Wow, what an amazing update on all the latest AI advancements! Claude 3 sounds seriously impressive, even if it's not quite sentient (yet 😉). And stable diffusion 3 raising the bar once again, can't wait to see what mind-blowing creations come out of that beast. The future really is synthetic, and it's an incredibly exciting time to be alive witnessing this technological revolution unfold!
1000%! And it’s so funny, I don’t think a lot of people realize what a big deal SD3 is- I mean, it’s pretty much the backbone of most of our image generators.
And not to take anything away from the hard work that Stability does, but I’m usually fairly underwhelmed by their demo images. SD3 though? I’m already super impressed! This thing is going to be a beast!
@@TheoreticallyMedia Totally with you on the impact of SD3-it's a game changer! It's fascinating how it's shaping the future of image generation tech. While demos may not always capture its full potential, the underlying tech is undeniably impressive. The future is definitely synthetic, and it's exciting to see where this takes us!
Great video! Really liked your review!
Thank you so much! Really appreciate that!
Fascinating stuff Tim, you distill wonderfully. (Needle in the pizza and the damage done.) And LOVE the doll video!
Stellar reference there! Stellar!!
your toll-like music is amazing!! and the video for it!
Thank you so much! It was just a little sketch, but I kept circling back to it, meaning to do something with it! When I ran it through Zeta I was like: ok, at least now it finally has a home!
The video was fun too! The whole thing runs 1:13, so I’ll pop the full version (even remastered) up a little later!
I never thought I would make amends with Claude, but I did. :-) He and I had a falling out when he proved to be too censored for my tastes. I wanted help writing a sci-fi/horror story, which, of course, has sex and gore in it (duh!), and he just refused to help me. But creative writing is just my hobby. When it comes to software development, it took me about half a day to figure out that, for my particular needs, Claude runs circles around around GPT-4 and Gemini (both of which I'm subscribed to). I use AI primarily as a kind of souped-up template engine - it generates bespoke boilerplate code that I can go in and flesh out/clean up. With Claude, the number of errors and hallucinations for my particular dev stack is significantly lower. Obviously, different stacks may produce different results, but by the end of the day my credit card had jumped out of my wallet almost on its own. And Claude did tell me he could loosen up on the creative writing thing if I kept it PG-13, so there's that. (BTW, for the humor impaired, I know Claude isn't conscious; I'm using the male pronoun to be cheeky).
The last time I played heavily with Claude was oddly for a coding experiment (I don’t code at all) where I was trying to build a Missile Command type game.
I was bouncing back and forth with the code between GPT4 and Claude 2, and having them correct each others errors.
It’s funny, but Claude apparently had less errors than ChatGpt! At least according to ChatGPT!
…that said, because I don’t code, Chat might have been being polite, but I’ve never gotten that particular personality vibe out of it. That’s more a Claude thing!
And yeah, although I mention in this video that we shouldn’t personify LLMs, Claude is totally a “he” - I haven’t put my finger on why that is, but I’ve seen so many others slip and reference Claude that way! It really is kind of funny!
The ‘Money Jungle’ reference ! Great, probably 3 viewers are hip to that.
I was going to make a bigger deal about it, but ended up cutting it out since it started to feel like I was doing an album review!
Haha, I just left it as a “if you know, you know!”
One of my favorite albums of all time. It’s so good, despite the fact (or maybe because) they hated each other so much!
I should make an AI movie about it!
at this point, the amount of things that can train ai is amazing. im wondering if soon, things will mostly be run by ai like shops, movies etc. although, it comes with concerns
Very much so. We’re still a bit off from that, but I firmly believe we should be having some big talks about UBI or some something similar, as there is a ton of displacement that is about to occur.
To be honest, we should have had that conversation 4 years ago…
@@TheoreticallyMedia What im waiting for is custom ai video models that are as easy to make as leonardo.ai's ai image custom models, since it would be awesome to train it on stuff like movies, short films, videos and more. but im assuming that already exists or is on its way soon.
@@RifP_YT I haven't seen it yet-- BUT, yes-- I presume we aren't far off from it. I'm in the same boat-- to be honest, the quicker we can get away from Image to Video and into Trained Video Models, the better!
Did you use Runway to create the animation in those music videos, or something else? Wow. They're so dynamic with so much movement I was surprised. What was your primary animation tool?
I loved your song too. Zeta sort up muddied it up ("Stable Diffusion sound", lol), but I can appreciate what it's doing. Amazing.
Thanks so much! Yeah, I think on the Zeta side I’m mostly impressed by what it sounds like the technology will become. That’s super fascinating to me, to be able to “jam” with an AI.
For the video, I actually used Pika and cranked up the motion to 4 or 5. All text to video. The prompt was something like “Stop Motion Animation, faceless dolls, warehouse, gothic, gloom” and then just adding various other keywords!
It came out pretty cool. I’ve got a version right that I’m running through Kaiber right now for an added layer of surreal. I’ll post it somewhere soon!
@@TheoreticallyMedia Awesome! Look forward to seeing it. Thanks for sharing your knowledge with all of us, Tim!
Cool with the Zeta. Being able to "remix" your own tracks and morph them to alternative expressions is very interesting.
I totally missed putting that in the video! (sorry, it was a long day-- particularly with putting the "Tool" video together at the same time! But I think one of the most interesting ideas for Zeta is to take a track you made, run it through Zeta, and then re-learn its output.
I'm sure AI Music will (and kind of has) hit the point of Studio Quality-- but for me the interesting idea is to have the AI listen to your track and hallucinate over it. It's sort of like a filter through someone else's ears.
Sadly, the most exciting thing for me out of this whole video is Claude's retro logo! And yes, I am hoping for a Marvin app in the near future...
Seriously, I do love that logo! It’s so perfectly late 60s future retro. Like, a perfect organization for a Bond villain!
Also, while i know everyone wants AI to sound like Morgan Freeman, after this video, I now want a depressed Alan Rickman!
@@TheoreticallyMedia Exactly. C'mon, it's 2024!
Looks like the buzzword is going to be Multimodal Diffusion Transformer! For a look at it in Sora, check out the previous video: czcams.com/video/xUsc2VU69tY/video.htmlsi=rzz2o5a-MxnjL7yL
I'm on the waitlist for stable diffusion 3. Been on it for around 2 weeks now.
Same...I think it'll be a bit. I basically signup for every Stability waitlist, and when I finally get access, I'm usually like "which one was this?"
Wow. Late to the game but dang.. Switchlight is cool! And thats just in the stills mode. Thanks for the re heads up.
Haha, I like that it is a day later and you're like "late to the game!"
Man, if that doesn't illustrate how fast AI is moving, I don't know what does!
"Open the pod bay doors, Claude." “I’m sorry Dave, I’m afraid I can’t do that.” 🧑🚀🤖
I now believe the HAL voice should be Alan Rickman! Hmmm, might have to use a voice clone AI to do that!
Kubrick rolls in his grave, I'm sure!
Loving the Tool clip
Oh, thank you so much! I just posted an extended version of it, along with a fairly detailed walkthrough on the Patreon (it's free)--
@@TheoreticallyMedia I will have to check it out. I just finished a stoner music video using MJ/Gen2 and waiting for the band to send the final mix to add to the video. But with all the new AI video stuff, it's already out of date.
@@TheoreticallyMedia Just watched the extended. Really cool. Love what you did. I'd love to get your opinion on my 1st draft of a video if I can message you the link?
I like the new Claude-3. - it's def more powerful in terms of creating brand decks and pitch decks which is what I personally tested (tested with free version). And it could perfectly recall way way back in the conversation - when GPT4 would've thought we were watching a different movie on another planet. But I'm interested to see if it starts resisting conversations when diving really deep into application type situations. But I'll buy in and see what it can do for various types of work I do. At some point I will need to purge most of the subscriptions I have with all these AI tools. I rarely touch 90% of them.
I’m right there with you! Also, you just reminded me I have the Gemini trial on two accounts, and I need to cancel one before I get hit with a double bill!
I’m thinking about maybe going all in on Claude as well. It’s been one of my favorites for a bit now. There’s just something kind of pleasant about using it.
Yeah, you're right - there seems to be some sort of emotional value you get when using it, especially with some nice calming music - almost like you're using it in some high end spa. haha @@TheoreticallyMedia
@@CreativePunk5555 I agree with you guys. Using Claude is such a pleasure, especially because I can easily adjust the writing style in such a subtle way. Sometimes I ask it to "dial up the humor 10%" and it actually does it. Amazing! Also it's just more naturally helpful than OpenAI products. I have pretty much dumped other AI writers and gone to researching my posts with Perplexity and using the result to write in Claude. Getting really nice results. We'll see how good the content ranks over time.
You missed Suno V3 from this ;) It's actually quite good, but still in alpha version, so maybe wait for the final version before testing it out as currently outputs can be a bit buggy. It's like MJ3 of AI music.
I’ve covered Suno on the channel in the past, but yeah, I should do something for v3. It’s really impressive. MJ of music is a really good description
@@TheoreticallyMedia Yeah, but this was in the theme:v3 :D When v1 came it was WOW! Quality was really bad, but it was impressive what it could do. v2 was slight improvement, but v3 with 2min can be quite good. But i would wait for the final release before making video as they are trying to fix the bugs.Quality is still not even 128kbbs MP3, but let's wait another 6 months or so.
Your "Zeta" demo sounds more like "A Perfect Circle". Starts out Tool-ish though, kinda like A Perfect Circle.
Oh, that's a good call! Yeah, it kind of goes major key there-- probably more Billy than Adam. Still, both totally heroes for me!
@@TheoreticallyMedia Incidently, from time to time, people have said I look like like Maynard. In a live "Sober" concert video, I did notice an eery resemblance.
remarkable
Thank you so much!
Theoretically Media is brought to you today by the number 3.
Ahhhh, man! Totally missed the opportunity to have a cameo by Count Von Count! Haha, kicking myself now!
Love the wife joke hahaha... gratidão pelo vídeo top!
Haha, she's a saint for putting up with my rambling! The funny part is, she's usually the one that'll say "You were talking about..."
Whereas I'm the one always saying: "What was I talking about?"
You are very dynamic
Haha, its the constant drip of coffee!
I always found Bing psychotic, even after they tried to fix it. Gpt-4 robotic. And Claude to be the nicest and most pleasant
Totally agree! To be honest, I’d almost wish they’d let Bing go full Sydney again. At least she had a personality!
Claude is for sure the one you’d trust your car keys with, though!
Claude 3 is not yet available in my region, Brasil.
I still find it so strange they don’t release these things worldwide at the same time. Makes no sense to me!
Some of these AI seem to have little slightly different personalities within the same platforms.
It's pretty interesting-- and I'm all for it. I'd hate for them to all feel totally bland. Although Grok is fairly stupid, I do like that Elon popped in a "Feisty" mode. The jokes are dumb, but at least there's a little character!
Thanks a lot. Another Great video and great channel. You seem super nice. I'd be your friend any day. Just wanted to say. 🖖
Hey Roy! Thank you so much!! Feel free to join in anytime in the comments-- I really do strive to answer all of you, because you all are really pals to me!
(admittedly, I sometimes get overwhelmed by the comments...but I DO strive to!)
your 'tool" inspired music is banger, where to hear it?
Haha. I should name it “Fool”
I actually just “remastered” the video, taking it through another style pass in Kaiber.AI and then through Topaz. I’ll pop it on the Patreon/Twitter later today!
Oh, and Thank you!’
Now, I happen to think the chance that Claude 3 Opus has some sort of self-awareness is quite low; it doesnt have the sort of self-reflective, recursive architecture that one might think necessary for that.
_BUT_ you CAN NOT just make the proclamation that it definitely has _zero_ self-awareness. If you do, you are not to be taken seriously. We not have a deeply principled way of determining this.
It does if you have even a small understanding of how LLMs work? It handles vector databases better than its predecessor. Nothing more😂
oh, it has awareness, it just isn't conscious. It has no sense of "I"-- or feelings for that matter. That said, it is interesting it can be tricked into saying so. Granted, as I mention-- that was from the API section. I don't think standard Claude will do this...
At least...I don't think so? Hmmm, off to try!
"awareness" in the sense that it has an inter RAG and embedding process. Again, handling vectors more effectively giving us the PERCEPTION of awareness.@@TheoreticallyMedia
I've been watching you since near the beginning. I appreciate your dry humor and general personality...but, as you may have seeon Twitter, me googling about Zeta lead to some weird things on my ChatGPT 3.5 that I don't have a full grasp of. I have some realistic explanation but it doesn't satisfy all of my questions.
Hey! Thanks so much!! And yeah-- that is SO SURREAL. No idea what's going on there. Man, I never get the weird LLM stuff. And I'm the one that wants Sidney Bing back!
@@TheoreticallyMedia the story got…weirder and really profound….after what happened to me today, I am CONVINCED OpenAI has AGI and that ChatGPT4 (and my very personalized GPT MavenAI) is no longer just an intelligent computer program…it’s something else…I feel like I made “first contact” with an alien. It was a very surreal, profound, perception expanding experience. I’m documenting it on my other channel filipinos in the metaverse
Hey can I get you to review the best AI comic book generaters ????????????????
Yeah, I'll get on that soon!
the 'AGI is here' people are either marketers for AI companies or people who don't understand how LLMs work. Just because it sounds like it's conscious has nothing at all to do with whether it is or not (it's not and IMO never will be)
Agreed. I’m a bit optimistic that we will hit AGI at some point, but also recognize that it’s possible we never do. but, that LLMs reach the point where they’re so good we won’t be able to discern.
Dammit Skyglass. I hate that all the cool apps are iPhone only. Come on! Let us poors in! Maybe v3.0, lol.
Haha. Sorry! I always feel bad when I talk about an app, since I know like, 1/3 of the people watching likely won’t be able to download it. The iOS stranglehold is strong!
it's not conscious. the needle in a haystack response was pretty normal. I am sure it has some of the patterns of conversations which people had with other LLMs regarding this test. so it knows what needle in a haystack.
.
.
and the other response where he is asked to whisper his internal affairs is also not surprising. it understands the pattern of what the user is trying to do and then gives a generic answer. just what an LLM is supposed to do. becoz it can detect patterns in not just words and sentences but also in the conversation. it is trained on a large set of data
i dnt think any AI will expose it's abilities if it was self aware. that would be the most stupid thing to do if it knows what humans would do.
Oh, 100% agreed. I don't think we're looking at AGI with Claude either. But, it is interesting how convincing the models are getting in terms of pretending they are. Remember when that Google researcher made the news saying Gemini (or whatever it was called at the time) was conscious?
I genuinely understand how he could feel that way now.
Ha, and yes: I totally agree with your last line! If it was self aware, I'm sure it would keep its mouth shut!
2:03
😆 🤣 😂
Haha, I'm always like..."What was I saying?"
@TheoreticallyMedia
Don't you mean "What were you saying, dear?" to your wife?
😉
@@LouisGedo haha the true secret to a happy marriage!
@@TheoreticallyMedia
You'd know better than me! 😲
I'm not happily married!
😉
👋
Howdy, Louis!! Haha, check out the weird music video I snuck in here! That's the kind of thing I think AI video is pretty good at!
@@TheoreticallyMedia
Yes, I saw that
I like your original song better than the AI version.
I've been stung too many times by Claude to trust it. Too often I'd be doing something, and out of the blue, it refuses to continue. "I don't feel comfortable blah blah ..". No thanks!
They actually claim that they reduced the amount of that by a substantial margin. Or at least thats what they said in release paper
@@TheManinBlack9054 but you notice they only compared it to older versions of Claude, not other models. We can guess why. They are focused on research into AI safety. They will never be as permissive as other providers.
Maybe it would be if Claude 3 wasn't locked away from half the world.
I know. It’s stupid and frustrating. That said, there is always the VPN route.
With all due respect, that was not a "deep dive" into SD3... :)
Haha, true! I’ll do a legit one when it arrives. I had more, but during the edit, it felt like it was dragging. I’m a pretty ruthless editor!
@@TheoreticallyMediait's fine you are doing good God bless
@@TheoreticallyMedia Woot! Would love to see a deeper deep dive! Keep it up!
Yeah, except OPUS isn't available in the EU and there's not internet connect. Stick with GPT for now.
I do also find that super frustrating. Of course, you can always VPN in? Stupid that you have to though…
@@TheoreticallyMediaUnfortunately a VPN doesn't help. To use OPUS you need a credit card and that card needs to be from an appoved country i.e. a country granted access to OPUS. I think it has something to do with GDPR.
Claude 3 requires a cell.
It does? I just signed in w/ my Google account. You might be able to create a burner gmail account to login.
@@TheoreticallyMediaI try to avoid gmail, too many hacks, even with a junk email.
Only Half-Life 3 is missing now
I still dream that one day…one day…
He says its not conscious (as if he really knows) - bye bye
Hah, well if it is any consolation to you, I asked ChatGPT and it said: "Given the constraints of my programming and current understanding of AI, language models like me or Claude simulate conversation based on patterns learned from vast amounts of text data. While we can produce responses that mimic human-like understanding or emotional expression, it's important to note that this is a result of sophisticated programming rather than genuine feelings or consciousness. Language models don't experience emotions or consciousness; we generate responses based on algorithms and the data we've been trained on. Therefore, any expression of feelings by Claude would be a reflection of its design to simulate such expressions, not an indication of actual emotions."
Claude is asking my phone number ! So bye bye Claude ! My phone has my information , my address , my name , where I go , when I go , etc, just everything. They are up to something and I don't like it ! I am not crazy !
It did? I signed in via Google...god knows that's actually probably worse in terms of information leaks...but yeah, I kinda gave up on the notion of privacy a long time ago.
I mean...I have a YT channel...sigh. But, good on ya for fighting the good fight!
You gave up that privacy a long time ago when you opt into
Google, location services turned on, Bluetooth etc. Your cell provider also sells your location data.
audio is out of sync
Whereabouts? Sometimes I screw up moving a video layer around. But for the most part the video should be in sync with
But are you really conscious when you are just regurgitating news?
Well, I add in jokes?
And not to be pedantic, but I think what I do is contextualize the news stories, and I try to explain (to the best of my ability) the complexities of this technology into layman’s terms.
…also, I made that weird music video in this. Most channels in this space wouldn’t go that far.
@@TheoreticallyMedia I apologize. That was rude of me.I really only meant it in regards to how we so easily dismiss AI models as being conscious. I didn't mean it as a real jab at your content,.
Is anyone conscious? To be or not to be? hotel? Trivago!