Udio, the Mysterious GPT Update, and Infinite Attention
Вставка
- Опубліковано 10 кві 2024
- It’s been a strange 48 hours in the world of AI, with the ‘ChatGPT moment for Music’ from Udio, that has reminded millions of what AI is capable of, and papers from Google that show that models can give infinite attention to text but we also got befuddling updates from OpenAI that suggest that not all is smooth sailing. We’ll begin with the quirky new tool on Udio.com and how musicians are reacting to it, then cover the strange manner of the release of GPT-4-Turbo with Vision and quickly touch on Mixtral 8 x 22b and Command R+ before turning to a fascinating new ‘Infinite Context’ paper from Google. One of the authors worked on Gemini, but that may or may not be relevant…
www.assemblyai.com/?...
AI Insiders: / aiexplained
Udio Intro: www.udio.com/
/ 1778045322654003448
‘The Site Is ****ing Down’ / 1778093021378089240
Musicians React: / udio_ai_music_generati...
Investors: www.udio.com/about-us
Will.i.am: iamwill?lang=en
suno.com/
Mixtral 8 x 22B and Command R+ Benchmarked: huggingface.co/mistral-commun...
LIveCodeBench Leaderboard: livecodebench.github.io/leade...
Majorly Improved: / 1777772582680301665
MATH Benchmark: / 1777926220132626753
Function-calling Usable with Vision: / 1777769463258988634
GPT-4 Turbo Vision Benchmarked on GPQA: / 1778463039932584205
Hassabis Chafes: www.theinformation.com/articl...
Robot Football Simulation Paper: www.science.org/doi/10.1126/s...
Video: • Watch agile mini human...
Udio Origin Story: www.theinformation.com/articl...
Leave No Context Behind: arxiv.org/pdf/2404.07143.pdf
Manaal Faruqui: scholar.google.co.uk/citation...
Gemini 1.5: storage.googleapis.com/deepmi...
Llama 3 Coming: www.theinformation.com/articl...
AI Insiders: / aiexplained
Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
Hey look the only good ai UA-camr posted
Not true but thank you anyway! :)
Who else would you recommend?
bro you're not kidding. Most others are just topical coverage
@@aiexplained-official but it feels that way. You put quality before quantity, and not rushing stuff out. You arent like "STUNNNED" "SHOCKED" "SURPRISED" "CHANGES EVERYTHING" with then a very modest update.
I concur the best tho
I wish there were channels like yours for every major topic.
Ah, that's a lovely thing to say Julius :)
100% agree! Well said Julius!
Hi, musician here…
I was very scared in the past, knowing that AI would come for music eventually as it does for all things. But I’m now actually very relieved to see it exist. Despite the music being great and basically indistinguishable from human music, it doesn’t change the reasons I do music. Music is an expression, a communication, a meditation, a spiritual journey. As far as I’m concerned the corporations behind pop music don’t produce anything different from an AI. Let the world burn and I’ll still be playing music. And until we’re all physically plugged into the matrix, live music will live on
And I will listen to it, and love it.
We still want to hear your music, any human's music, if it's good. The only way AI music affects you is economically. But that's no reason to stop doing what you love and possibly creating a masterwork that others can appreciate. It's an economic problem we will eventually fix
As a musician myself too I've been having fun messing around with AI music to mix with my own things. I use it mostly as inspiration or take elements of something I'm liking to add other elements, etc. It isn't much different for me than hearing music to gain inspiration, except I can take some parts of it too. I been trying to make some jungle, beatcore, glitch beat stuff and it helps getting a lot of sounds and cool beats i can incorporate with my arrangements.
As a listener to music, It will destroy the awe, inspirational skill/technique and emotional captivation I've had with bands and music I love. I have no desire to listen to machines, but soon there will be no way to know the difference as AI art has proven. It is not just about potential irrelevance of creators, but for the entire culture itself.
Yeah, I think we’re probably going to have to majorly restructure our entire economy pretty soon. Every job is threatene, the very nature of what it means to do work is getting redefined. I’m still studying, but we live in a new world every month nowadays. When I finish my studies, we’re going to be in a different place entirely. Everyone should be very worried about themselves. If anything, the stability that music provides is spiritual respite in these very meta human times
It is so insane. If someone would have told me 20 years ago what we can do today, I would have been so excited and amazed. And now we literally live in the future. Next 20 years are gonna be wild.
Even just 5 years ago LLMs like GPT-4 and music models like Udio would’ve seemed like sci-fi
In 20 years, the year 2024 will seem as distant to us as the 14th century feels today. We are expected to achieve AGI within the next 2 to 5 years. Once we have it, that’s when the real progress begins.
We live in a gold rush, but that will quickly become a dystopian nightmare with extremely hypnotising entertainment and population control. In France, they talk about limiting data to 3Gb/month/person "for environnemental reasons".
Misuses of ai will be the reason to implement tracking systems, and if you don't adhere to those, you will basically not exist (why a bank would give an account to a person who has something to hide ?).
Fights against totalitarism and population brainwashing will be wild.
Live in the present
@@MrSchweppes progress for who? We don't know how to control AGI or robustly align it with human preferences (like the preference that humanity shouldn't be destroyed, even though it's existence would be an obstacle to basically any goal).
I'm an elementary teacher and played with Udio a bit yesterday (lots of hiccups due to servers being overwhelmed, though) and your suggestion that teachers could create little songs to summarize the lesson is exactly what I had in my head.
Essentially the same concept (bespoke, personalized, AI-generated education) was central to Neal Stephenson's 1995 novel The Diamond Age: Or, A Young Lady's Illustrated Primer. Fun, fantastic read that feels more relevant than ever.
@@JohnVance aw man that was one of my favorite books ever. such an eclectic mix of ideas in that one! clockworkpunk if it were actually taken into the future.
I thought of that, too. And I teach college. :)
Experiment and let us know David!
In America's diverse and inclusive environment, good luck trying to convince Muslims to sing or listen to those songs voluntarily, when the quran teaches that music is an abomination and that you should rather have molten lead poured into your ears rather than make music, or something like that. (I might be paraphrasing something their prophet Muhammed said, but the TLDR is: Music is haram for Muslims.)
The only AI channel without clickbait titles and actual relevant and meaningful AI news and updates. Thank you AI Explained
:), thanks David
Nice to have atleast one reliable AI channel.
:)
The goat AI channel
Tried the latest AI generator for music Udio. Please, don't use it and don't help its learning process. It's scary good and the musicians are the people who should avoid it. Just play your damn real instrument and practice! At the moment it gives too high a premium for mere numbers. There’s only one real evil in the world: mediocrity. Soon you will regret it Peace and love! p.s It generated almost identical song and style of playing, as the greatest Satriani!
👹@@diminalbantov
The little robots are super cute. There is potential for a league here, and each bot gets it’s own personality and skill set
I would lose it if they had signature goal scoring celebrations
A.I. football GGO:)
ua-cam.com/video/Ub1Z02dVKXM/v-deo.html
they even learnt to take a dive
what are they called?
I can't believe how the AI-generated Classical Music was so immediately calming.
it's "shoegaze" a 90s/00s style of heavily distorted rock named for the appearance of the guitarists always gazing at their shoes, because they're often looking down at the next distortion pedal to press
Dimmu borgir already covered the Gregorian chants and blast beats part though!
one of my favourite genres. best coupled with shrooms
One of the last genres of rock music not corporatised, and that came from a counter culture, which gave it a unique style. Slowdive for example, and they are back together releasing new music.
@@blakecasimir Slowdive is great!
Fascinating
AI
I am an engineer, and from a technical standpoint AI seems fascinating and the results are more than impressive these days. My main beef is with the crazy focus on arts and now music they have put in. It's a low hanging fruit for roughly three reasons: The architecture of CNNs fit ideally to image data, by design, and also to musical data, I assume, depending on clever choices of representation. Further there is almost an infinite number of training data, ready with categories and tags to be mined off the web. Lastly, due to reasons one and two and the increasing power of GPUs etc, the generative output becomes stunning and the fact that all this art was created to evoke and play with human emotions in the first place, makes the experience of witnessing the results truly a "lost for words" experience at times. This leads to access to financing from venture capital to research grants on state levels etc. On larger scale, this is extremely frustrating. For one, all these resources could be allocated and used to improve human well being on more urgent levels, like healthcare, complex geopolitical issues, hunger, true equality etc. All of this is done already of course, but paling compared to other efforts. Further, I LIKE the fact that good music is scarce to some extent, and I WANT to be thinking about the exceptional human mind who created this piece, how they thought about it and just be in AWE of the human achievement here. Recently I heard a song "Answers" by Nobuo Uematsu. I had never played the game it belongs to but I felt everything that it was about, the pain, despair, beauty and the way it somehow highlights our own journey through this often a bit mysterious thing we call "life" but no one really knows why we are here. I was again lost for words how he could create something so amazing. I want this connection to the artist GUARANTEED, I don;t want to live in a world where I have to wonder everytime I hear something awesome or see an amazing image, if it was just created in-silico to exactly hit the dopamine center of my brain. I think the arts should be left to humans only. Some people talk about "it's just another tool, like photoshop etc were when they were invented". It's not, it's entirely different. With a tool like Cinema 4D it need the same amount of skill, albeit it different skill, to create something impressive, as it did with watercolor. It was months and years of practice, it was just a different tool. AI really allows everyone to create stunning stuff. Well, having the AI create it for you. If you had a slave-artist chained to your room, doing anything you asked, would you call it a tool?
I know exactly what you mean, a lot of real music will be questioned a year from now, which is sad
But aren't humans just highly advanced biological intelligence? We are much token intepreters as the Ai, are we not? Can we not also marvel at the way AI interpret the input data and produce something brilliant?
@@abdvs325 In my opinion no, but I agree that this is debatable.
thanks for the comment I totally feel
The soccer robots is way underrated…. We legit trained robots to run around and play soccer with simulation? What that means for the future of robotics over the next 5 years is actually jaw dropping
That's what I was thinking. AI robotics has potential to seriously disrupt the job market even faster than LLMs.
It would be really cool to see a mini robotic soccer championship similar to battle bots where different teams work on training and building their own AI robot soccer teams. Would be a fun way to see the technology develop
@@some_doofus This already exists ua-cam.com/video/Ub1Z02dVKXM/v-deo.html
@@some_doofus In before soldier robots from actual armies have their own deadly competition
Bruh you see them faking and stutter stepping and shit? Absolutely insane
Just for reference, my company uses GPT 4 for the medical field and the update has made noticeable (but not massive) improvements in reasoning with large context / massive prompts which is good.
OpenAI still needs to release a new model to get back on top.
Great news Oscar, do keep us updated with the next release's impact
4:40 "Hmm, Human Music. I like it!"
MY MAN!
When I saw the bots playing soccer (a hobby I kinda take seriously, competing and all) I literally reacted like Fry from Futurama when he saw baseball from year 3000. "What!? Robots playing soccer!? Hey, it's finally robots playing soccer!"
Judging my how well those little robots moved and actually played its only matter of time before we have the Equivalent of the Terminator running around. Scary!
It's so crazy how I knew this was coming but it's still completely blowing me away
i didn't think anything was going to beat suno v3 so soon...
After creating hundreds of songs in suno, I realize they are really fun but hard to say they are good music. The rhythm and lyrics are highly repetitive, cords are lack of creativity. But def better than suno v2 for sure.
@@biiigdaaaddy "The rhythm and lyrics are highly repetitive, cords are lack of creativity." But also like, turn on the radio and it's the same!
now let's see how much time until Udio is beaten
@@JohnVance you are right. I don’t like to listen to radio, and that could be one of the reasons. But it’s just me 😌
by the way let me just say I love your thumbnails, such a nice break from closeups of people yelling or giant neon letters to get my attention, in addition to the great content ♥
As a musician that recently got diagnosed with Miltiple Sclerosis and is slowly losing his ability to make music due to disability, this type of AI gives me a lot of hope that I'll be able to keep on making music until I die.
Am sorry to hear about your diagnosis Bruno but very glad for what this technology will unlock for you.
Look into parasites
@@alexgordon951parasites? I was going to recommend cannabis, many ms people confirm benefit from it..
Sorry. A friend has that.
I tried Udio yesterday afternoon and it was down for me as well. Kind of disappointed, but their song about the site being down made it worth it.
Praise the agi and praise the non shocking titles!!🎉
I expect corporations behind the music "industry" to be much more organized in their legal crusade against these tools than the collective of random individual painters.
No. They will clone the artists and embrace it.
Microsoft, Google and Amazon won’t let them win. It’s all Fair Use. All generative AI is based on fair use. The IT giants won't allow even the big shots from the music industry to set a precedent where someone loses a lawsuit based on Fair Use.
@@MrSchweppes What do you mean all generative AI is based on fair use? I get that there hasn't been precedent but seems like everyone trains on copyrighted material.
Either way, if there's any incentive to not use AI in an industry, anyone that does can just be blacklisted without needing a lawsuit
@@dweezo2175
@@dweezo2175 If you study Fair Use, particularly Transformative Use, you'll find out that it is perfectly legal. I understand that it is hard to accept that fact, but nevertheless, it's true. Without it, we wouldn't have any progress whatsoever in any field.
3:03 go home Udio, you're drunk
🎯 Key Takeaways for quick navigation:
00:00 *🎵 Introduction to AI developments*
- Overview of recent AI developments, including the release of Udio, updates on GP4 Turbo, and a new paper from Google.
00:41 *🎶 Udio's capabilities and musician reactions*
- Udio's ability to generate music, comedy, and other content.
- Mixed reactions from musicians, ranging from excitement to concern about the impact on the music industry.
05:17 *🤖 Mysterious release of GP4 Turbo*
- OpenAI's release of GP4 Turbo without detailed benchmarks or explanations.
- Speculation on improvements and comparisons to previous versions.
09:41 *🔍 Google's paper on Transformer models with infinite context*
- Discussion of Google's paper introducing Transformer models with infinite context capabilities.
- Potential implications for long-context understanding and model adaptation.
12:17 *⚽ Google's deep learning achievement in football simulation*
- Description of Google's achievement in training football-playing agents through deep reinforcement learning.
- Comparison of the trained agents' performance to a pre-scripted baseline.
Made with HARPA AI
In the bot soccer duel at the end where the bot who loses the ball takes a blatant dive to make a last ditch effort to win a foul by influencing the ref was heartwarming to see. Seems like a lot can be learned in training
Needed an extra roll on the grass for realism
@@aiexplained-official HAHAHA, reminded me of Reyes playing for Arsenal
Thank you for dropping these! Watched amazing. ❤
Love your content! Keep it scientific and down to earth like you always did. You're the best AI news channel on YT!
Thanks Moonscape!
As a musician, I always considered authentic music generation as the final frontier of AI. Now that we have "arrived", and the pace at which this occured, I can only conclude that we are in the midsts of a soft-hard singularity takeoff, or we are at the doorstep of a hard takeoff. The only limit at this point is hardware. We could potentially be one hardware recursion away from ASI. Perhaps Stargate is the precurser to this, should it be required to facilitate AGI
Great video and I’ve been playing w Udio since yesterday. Looking forward to when they life the 30 second limit.
you can already extend songs you've created to add intros, sections before/after, and outros. It's a bit tedious at the moment, but does allow some flexibility for creating a longer song
@@DougJohnston1 Even the guide recommends it to be a 1:30+ intro/main/outro at most. Not sure which of the two is worse when the objective is to make a solid 3:00 song with consistent refrain/chorus. Suno with a limited context and so flying off the rails past a certain point, or this one, where you're stuck to 1:30-ish as far as consistency goes. Tried to extend a song to 3:00+ before reading what the guide said. I was disappointed. Both fail in this area, so I judging purely by quality, Udio is definitely in the lead for me. I quickly found a banger tune I wanted to extend.
Your channel is professional and insightful. You keep doing this, mate. You're great. 😊
I'm working as an animator in a small studio with a sub-shoestring budget. Our project would actually be impossible for people in our economic situation without AI like Udio and Stable Diffusion.
Musician here - if I have an AI that I can say something like "generate me a Mongolian chant at 120bpm" or "8 bar tabla rhythm using notes D, F and A", that would be an immensely useful tool.
Are you worried about AI replacing you?
@@jigsaw2253not if we can actually use it like the comment above. I think in some cases a human can get beter more tailored results
Give it a year or 2. This is the worst it will ever sound.
Listener here. If I learn music I thought had been crafted with love and effort was initially generated by a model, I'll feel extremely betrayed and drop you on the spot.
@@guy_th18 Then you don't understand making music. Why would I spend hours looking for the right sample when I can generate exactly what I want to slot right into the piece I'm composing? If I can't find what I'm after, that's preventing me from being creative unless I buy the instrument, learn how to play it and then record my own sample, which would take literally years. It's not different from using a synthesizer.
Also, you have no right to feel betrayed, I owe you absolutely nothing as a musician.
what wonderful work you are doing! Thank you! :)
one of my fav ytbers currently, consistently puts out enjoyable and informative content. thanks 😃
:))
Great video as always. I find OpenAI’s lacklustre announcement very peculiar, I will be testing the “new” model on our use cases to see if there’s any tangible improvements.
Thank you so much for the link to universal 1.
It's great
0:43 “Dune, the Broadway musical”
It reminds of me of one of those “hit songs” from a musical _within_ an actual musical, i.e., a “fictitious” song (if that’s a thing), which, when you think about it, it kind of is. (It also reminds me of bits of the female chorus in “Prince Ali” from _Aladdin_ but why the AI would emulate _that,_ well, who knows?) The verse works but the tune falls apart at the chorus, for me at least.
_Adding:_ Mike Sharkey over on the “This Day in AI Podcast” sent one of these Udio clips (“Adrenaline Rush”) to several record labels in Australia just to see what the response would be, _not_ indicating the clip was AI-generated, and got interest from several back. (He hasn’t figured out how to respond yet.)
Delivers as usual. Great content. But there is the typical I already read it in full detail missing... 9/10
Haha, will be back I am sure
Another amazing video! Thanks Philip!
Thanks Elijah!
We cannot even image what is being developed that hasn't been revealed, yet. 🤯
Great video again!
Thanks for helping us all stay informed on what’s going on in the ai world! I appreciate the work you do for us!
Thanks Ashton!
13:18 I could watch this all day!
As a professional music producer and musician, I don't believe that Udio represents a ChatGPT-level breakthrough for music. The sound quality is quite poor for both AIs, and there are numerous artifacts in the sound. Suno AI produces simpler track structures and has a better understanding of music theory than Udio. While Udio creates more complex structures, they often lack coherence, they tend to go off in all directions. However, guitars and vocals are better with Udio. The output quality varies by style, and sometimes it can be even worse than what Suno AI offers. They're on par with each other, each having its own strengths and weaknesses. But on the whole, for both Suno and Udio, the sound quality and creativity are quite poor today.
Of course Udio and Suno made some improvements compared to what we had a few months ago and it will be improved. But I think a kind of autonomous agent like GPT-5 or GPT-6 using a music software like Logic or FL Studio and capable of listening to what it writes, is the best way to make Al music. Of course, it will be a little bit slower than Udio / Suno, but the quality will be 100x superior. And you'll be able to make different versions of your track for music licensing, because it's very important for movies or video games.
AI music will be primarily competing within the royalty-free music industry and royalty-free music has been around for years and hasn't stopped movies, video games, and advertisements from securing synchronization deals with artists for copyrighted music. When the music meets their standards, industry professionals are always ready to invest in the work of artists they value. The introduction of AI music is not going to change that. So don't be afraid if you're a musician.
The current path of AI music (generating full audio songs), as seen with Udio or Suno, might be suitable for creating royalty-free tracks but that's it. But it's not necessarily pushing the boundaries of quality (I don't think we can get rid of the artefacts with their method even if it will be improved). What you're seeing everywhere on Twitter represents the best outputs achievable (after 300 attempts).
The only really cool feature in Udio compared to Suno AI is that you can choose to extend a piece of music by selecting sections, such as an intro, break, outro.
The only problem right now is when someone is uploading AI music on streaming platforms. AI-generated music shouldn't flood the platforms (with shitty music right now, but better music in the future); otherwise, human creations will get lost in the radar of releases. AI should benefit humans, not disadvantage human artists.
The only ethical approach I see, is to divide the music industry into two sides: streaming platforms for human artists and streaming platforms for AI music.
Also, just to mention, I'm a huge admirer of your channel. I've been following since the beginning! :)
The music is indistinguishable to me. I don't hear any artifacts. You may be overestimating how much people care or even notice sound quality
This is the worst it will ever be. This is mind blowingly good.
Nice thoughts. Should post it in a blog or sm mb
@@rasuru_dev frfr
Cope, some tracks are indistinguishable already
Now that is trippy, I knew it was coming but that fast is insane. Those robots playing soccer are fantastic, I would love to see teams of robots playing soccer against each other! Thanks again, you made my Friday.
Thanks Mad
🎯 Key Takeaways for quick navigation:
00:00 *🎵 Udio's Impact on Music Generation*
- Musicians react to Udio's capabilities, expressing a mix of excitement and concern.
- Will.I.Am praises Udio as revolutionary technology for music creation.
- Comments from musicians highlight both admiration for Udio's advancement and apprehension about its implications for the industry.
05:01 *🧠 Evaluation of OpenAI's GPT-4 Turbo Update*
- OpenAI releases the GPT-4 Turbo model, claiming significant improvements without providing detailed benchmarks.
- Independent evaluation suggests modest enhancements in reasoning abilities, particularly in handling complex questions.
- Questions arise about the effectiveness of training on more advanced data and the limitations of current AI paradigms.
08:03 *🤖 Latest Developments in AI Models*
- Overview of recent releases from the open weights community, including Mix Trial 8 and Coherent Command R+.
- Introduction of Assembly AI's Universal One model, praised for its accuracy in transcription.
- Discussion on Google's paper introducing Transformer models capable of processing infinite contexts, hinting at potential advancements in AI capabilities.
12:17 *🚀 Origin of Udio and Industry Dynamics*
- Insight into the origins of Udio as a project by former Google DeepMind staff.
- Frustration within the AI community regarding the availability and transparency of advanced AI models.
- Recognition of Google's contributions to AI, despite internal challenges and competition from rivals like OpenAI.
Great video. Udio hallucinating like that is kinda scary.
Imagine embodied hallucinations ... eek
Thank you for your well searched videos!
And well researched!
8:50 Google Gemini 1.5 Pro's Audio capabilities were just released less than 48 hours ago. Try comparing that model's transcription abilities with Other Speech-To-Text Models.
As a bonus, Gemini 1.5 pro Can do more than 4 languages.
I would be shocked if it was anywhere near GPT3.5's audio. Google has always been so horrendous at this stuff that it would be nice for it to change.
... Every D&D book :D
Also, I would love to have my own tiny robot 5-a-side football team :D
well thats certainly interesting, i can foresee udio being used for radio stations in games and whatnot
Music AI is crazy fun to mess around with
4:15 he missed the e. Shoe gaze. It is music of looking at shoes.
And all this besides the amazing papers I regularly read on hugging face paper page.
Now all we need, in addition to infinite context size, is variable "pondering" length/duration/cycles as a parameter.
I’ve tried many things like asking it to take longer, rerun through its responses 10x times, provide deep insights into why it provided the response it did, please take 5 mins to consider and reconsider scenarios/responses, but can’t seem to get it to take more time to ponder… I think the answer lies in asking it to use downtime dips to utilize additional processing power, but I can’t figure or break it, need the language used to guide it in order to avert it.
@@evdm7482it doesnt work like that (yet). When you enter a prompt, the model predicts which tokens are most likely to come next given its system prompt (which is something like: "You are a helpful assistant, answer the users prompt") and the prompt you enter. Thats it. It doesnt have the ability to alter its own architecture or alter any code thats being run based on your prompt. So adding in "pondering" means the developers changing the code
@@evdm7482 Aside from the recent attempt I've seen to use pseudo-code for reasoning, I think the answer lies somewhere between the attention layers and the FF layers. Humans usually use mental simulations to solve a problem. CoT and others mimic this procedure, but are limited to the space of language. Maybe multi-modality is the answer. So not only using CoT in text, but also in visuals or 3D space. Like "Create a visual step by step guide for this problem" or something.
What impressed me about the ball-playing robots at the end was when one of them stumbled and recovered smoothly; it gives a truly organic vibe.
Simulations can be scaled up 10,000x in the next couple years, as they have been already with IsaacGym. I expect the organicness to get noticeably better from here.
@@aiexplained-officialThe first time I saw something like that was a multi-legged robot that looked like a scaled-up bug. It was driven by a neural network copied from a cockroach. It scampered over irregular litter-covered ground, and the "organic" moment was how it coped with shifting pieces when the demonstrator pulled some of the boards out from under it.
That classical music was very good.
I was wondering when we would finally get good AI music. Thank you for your work, and please do make another video about that unlimited context window paper when you're ready because that sounds like a major breakthrough for sure!
Thanks Allister! It could be big, for sure, if it can scale, and that would have been proven if it is indeed in Gemini.
@@aiexplained-official What I'm thinking is that it could make training data somewhat irrelevant, as you could litteraly dump in its context window any data corresponding to your query yourself.
Like you make a general model decent at everything but no more than that, and you create a conversation with it in which you dump a few million tokens of, let's say coding problems, then you make your query and it would be the same as having a model specifically trained for that. On another conversation (or the same, after if its infinite, why bother), you give it a whole library worth of D&D rulebooks and play a solo campaign with it.
Of course, that's assuming it is actually infinite regardless of hardware limitations and that it doesn't slow the process too much. Though I'm sure it's not that simple.
Well, i can't unhear Dune as a big band
Hey Dave 😄🙂
Dune as a broadway showtune…
My ears are still bleeding five minutes later.
8:36 As someone who has been very interested in the transcription of large libraries of audio and video, I actually really appreciated Assembly AI's 'Universal 1' sponsorship of this video.
Their announcement had escaped my radar until I watched this video.
A win win for sure. I only endorse things I actually genuinely think are great, which limits options 99%.
The classical music is arguably the furthest away for the genuine article :) We’re at superhuman country though 😂
Udio is amazing, can specify styles, add lyrics, extend tracks, really good!
It is
I used to wonder where AI explained got all of the great info...... Then I realized it came from AI Insiders - well worth it!
:))
2:45 probably an unpopular opinion, but I prefer Suno V3 over Udio
Its definetly a close match, i wonder who comes out on top in the end
I prefer Udio, without any doubt. Much better instrument separation and stereo image. Vocals are way more convincing and are more diverse. The "songwriting" is much better and sounds more logical. I could go on...
@@bobrandom5545 I agree with you on that. I think Udio is better overall, and the lyrics are more natural - just the feel that the style of song in this case was better (subjectively of course)
udio has better audio quality, Ive made some really amazing songs with suno-ai that ive not been able to do on udio. But on Udio Ive recently finished several parody songs, one im particularly proud of is an Uwuwfied version of Nirvana's Teenage Spirit.
As a poet, Udio lets me explore entire worlds of expression that previously I could only dream of.
The Ilya part got me xD
April, next year: "The last 12 hours have been a rollercoaster for AI development."
Borderline a livestream by 2028
4:53 That’s legitimately an amazing idea and I’ll totally do that.
Udio is insane.
I created this Bob Dylan type song about AI getting crazy.
Just brilliant.
Just used Assembly AI to transcribe your Gemini 1.5 video. Very nice
It is pretty amazing. Underrated ampunt of progress happening in speech to text.
Didn’t think I needed a Dune musical, but now I do:p!
I'm trying Udio and the results are really impressive.
Thank you for the update. Regarding coding, I haven't noticed any improvement with the new GPT-4 Turbo; I've completely switched to Claude 3 Opus :) additionally I tested Gemini Pro 1.5 is accurate in retriving data and huge context window is very convinient. It's great that OpenAI does not have a monopoly anymore :)
Claude 3 is my daily driver now too
From my real world use the gpt4 performance for summarizing academic discourse is strikingly better than in the previous version.
Do others have the same observation?
Strikingly is a big word, wow
'The site is down' is f@#!ing catchy.
Hands down the best AI news outlet on the internet, so happy to have found you early on! Cheers and thanks for being awesome
:))
Once again I slept through!
Thanks a lot, amazing content!
Thank you Andi!
I assume that OpenAI felt that they needed to match Google's announcement beats but this time around didn't really have much to share. So they announced a normal GPT4 update.
I mean the timings are strange
I love that you don't use clickbaity titles!
Me too
I can't stop thinking about what GPT-5+ level intelligence looks like with "infinite" context length. The possibilities...
That's basically ASI.
4:40 Hmm, human music. I like it!
-- Jerry Smith
Amazing classic music, for me that has always been nice, and nothing else
A one-off video like every 2 months for certain AI concepts like context windows, where you do a survey of the SOTA methods for improving along the axis of the video's topic. I know I would deeply appreciate such an enlightening video on any of these topics: context length, agents, fine-tuning, base tuning, multi-modality, simulation, and/or code generation specifically focusing reasoning over needles in a needlestack as the major component of lack in reliability for long form tasks...
Just my experience with opus hitting the message limit at every possible window they would give me over the past few days, building out "timelord" an environment for AI's to test vlm's and llms abilities to perform long form tool-based tasks to deduce if they have sufficient awareness of self-improvement to augment themselves with the skills needed to bootstrap themselves effectively "into" their environment to effectively self-improve. In more concrete terms, they will have access to developing on active separate simulated versions of their environment with the task of maximizing their general faculties across tasks of its choosing.
I think if they are capable of significantly bootstrapping effective rapid skill production into their own prompting mechanisms, well FOOM.
Haha some claim we don't have a route towards AGI, it's funny at face value I can agree, because then again maybe ASI comes first.
I can't even believe I'm real sometimes, so how do you all feel?
It's just one of those days because we're currently still in the era of different accelerations yielding semantics.
Wait until we get into seconds and we all see every coming ontological nuance denoting sentience.
Then again for humans deciseconds are relative enough to comprehend, but by then we will be measuring time in tokens.
Remember 2015 - 2019:
1. The companies were just coming on board with acknowledging we definitely weren't in an AI winter anymore
2. RL was just getting sim2real
3. Neural Architecture Search was still a matter of compute
4. What about GANs, who came up with those again, all I know is it was over a beer with friends.
5. Unsupervised learning was the cake, supervised is the icing, and RL is the cherry, damn Yann Lecun got the Turing award for more than just his contributions, mans was a predictive powerhouse.
6. Remember Elastic Weight Consolidation (EWC), sure we have MOE but imagine if we could subset experts for continual learning based on their implicit biases towards newer information. I'm thinking how Sholto Douglas called to action someone on interpreting specialization in Mixtral, and honestly maybe they missed time specification as a means of historically representing information across experts. Imagine a continual expert designed specifically as a module for integrating new information with other experts.
7. Remember the good times.
I know what you mean. I need to hire a researcher to help me add deeper dive mini-documentaries to what I already do. Put it in bio
6:41 note that he's comparing with previous versions of gpt-4-turbo and not with gpt-4-0613 which is still more accurate as per science-based benchmarks (albeit not preferred by most people)
I worked for years as a professional musician and I'm absolutely blown away by Udio. I will say however when it comes to production control and reliability is key (no pun intended). Don't get me wrong being able to generate random genre based tracks is amazing in itself, but I'd like much more control to the point where I'm able to ask "give me a 124 bpm classic rock track with a 4-5-1-6 chord progression in the key of C#". That way I could design tracks for my vocal range, style, or even the instrument I'm playing.
I honestly think this could eventually be awesome for musicians. Nowadays if you want to do a live show of your own music you either have to put an incredible amount of effort into producing all the tracks yourself or pay professional musicians to either pre-record the tracks or even pay an enitre band to do the rehearse and do a full gig with you. This could allow musicians to design an entire show built around their specific vision and talents with no limitations on funding, scope or conflicting creative ideas.
It reminds me of what one of the artists that got access to Sora said (paraphrasing): the potential of this feels like it unshackles creativity from the established constraints. I think I might dip my toe back into music if this lives up to the potential I think it has!
It will get there. Udio for music is getting close to the first explosion of AI art to artists. But its still justttt not there yet. But like how AI art models eventually solved the "drawing hands" problem, so will AI music models.
Really interesting framing, thank you
I believe they've said they plan more musician-focused features for Udio like more control, stems, etc.
Love the channel, can't believe it took youtube this long to recommend it.
Glad you are here Kano
I was really impressed with Udio - Only copyright law can stand in the way of AI domination of most, non live, music now.
I recommend specifying Hip-Hop if you just want to make something for laughs as you get a lot of cool lyrics 🙂 I tried just specifying a club that I'm a member of that has a web presence and that worked brilliantly.
I wonder if AI music got so good, that a live band need only one front person, with an infinite orchestra behind?
3:03 I mean, comedy was it's goal. It technically nailed it 😂
"Sir" Demis has very little reason to stay at Google. He doesn't need them to open doors any longer.
he should get out and start an open AI/anthropic style lab. Get Karpathy in on it and bring a few other top researchers with him. Things will move a lot faster in a smaller more nimble lab like Deep Mind used to be, as Phillip sort of alluded to here.
Udio is astounding.
Banger, ty
I must say I am not sure why so many people are surprised by cross-domain AI capabilities. So many elements of human mental endeavour can be reduced to the concept of "language", even our sciences are symbolic representations of physics which can be reduced to "language". Just like words in a sentence revolve around context, so to does music, it's not a big jump (conceptually) from chatting to composing to protein folding.
The Gregorian chat with shoegaze. Literally just tried that... we are both Batushka fan I see 8)
My brother is such a lover of broadway this is going to blow his mind
It looked like they were just stumbling around, but there was actually some pretty good shots.
Trained in simulation, transferred zero shot. Incredible
As a singer myself, i tend to get into situations where i need a beat for a certain song i want to sing but because of the high cost that the producer asks for a license, i can not bring out my music. Now with Udio i am finally able to make instrumentals that i can use for my own albums on spotify without having to worry about royalty payments or paid beats.
Very nice
That Dune Musical is the greatest thing ever.
3:04 That’s how I hear most song lyrics to be honest.
Have you been able to get udio to generate lyrics with different rhyming schemes?
I've been using Suno alot and I for one am excited at more competing AI tools for music creation, I really want the ability to isolate specific segments of the generated song and regenerate it or open it up in other music creations methods to give me even more creative control over the whole process. So far the ai music plus ai arrt have been a lot of fun to create silly little additions to our DnD campaigns that we woulddn't have otherwise. I am concerned what will happen the more and more commercial this stuff gets though.