Gemini Full Breakdown + AlphaCode 2 Bombshell
Вставка
- Опубліковано 9 тра 2024
- Gemini is here! All 60 pages of the technical report read, plus the AlphaCode 2 bombshell paper explained and analysed. Is that paper even more consequential than Gemini? Plus the launch of AI Insiders, Gemini demos, Hassabis hints and much, much more!
AI Insiders: / membership
Release Page: deepmind.google/technologies/...
Gemini Technical Report: storage.googleapis.com/deepmi...
MMLU SmartGPT Video: • SmartGPT: Major Benchm...
Availability: blog.google/technology/ai/goo...
AlphaCode 2 Paper: storage.googleapis.com/deepmi...
Google Clips: • Math & physics with AI...
Google Credits: www.nytimes.com/2023/12/06/te...
Hassabis Wired: www.wired.com/story/google-de...
The Verge Hassabis: www.theverge.com/2023/12/6/23...
MMLU: arxiv.org/pdf/2009.03300.pdf
Jim Fan Tweet: / 1731473285668618581 Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
low key best part about this channel is the "videos when merited" tagline. I KNOW I am gonna learn some interesting stuff, not just filler to feed the youtube algorithm.
Thanks tamro, tough sometimes but worth holding that line
Someone actually doing a fair and detailed objective cover and not just reading Google's flashy blogposts. Subscribed.
thanks fabo!
I love your due diligence. Always on time.
I'd like to bring attention to AI Explained's dued illigence. I mean, this guys stuff is super sick. And just like the Beastie Boys, AI Explained certainly has a license to ill. As such, a nod his way is due.
I agree, the moment I heard about Gemini, I was eager for his video to come out
The fact that you get your results and seminal insights on such a shoestring budget is beyond belief.
Trying to fix the shoestring bit!
@@aiexplained-officialgodspeed !!!
Do you have a patron?
@@frank4425 I do! AI Insiders, link in the description, bonus videos, Discord, podcast, the whole shebang
Reading the Gemini news in the morning and then constantly waiting for the ai explained video to be released 😊
Thanks tom
It is hard to believe that you are driving this all alone. I will be joining the patreon without a doubt, the value your UA-cam channel brings alone is outstanding. Besides Tom Scott you are the only other channel I have watched every video of, and keep notifications on for! Thank you once again for the quality video!
Very exciting to see if we get any reaction from OpenAI! 👀
Thanks Face, yeah pretty exhausting!
😮😮😮😂
I love your voice :) @@aiexplained-official
@@aiexplained-officialHave you considered hiring an editor or any position that could save you some work?
You did all of those things AND made a video for us. Thank you. Exceptional, frankly. x
Thanks rich
@@aiexplained-officialplot twist. You're the true AGI.
It never ceases to amaze me how quickly you digest new information and get it formatted for a video. There needs to be studies on your brain 😂
Haha thanks Brian!
@@aiexplained-officialthanks Brain!
probably with the help of AI
Nope
Lol
Talking about multimodality, I think drawing is needed too. Generating images is a whole different process than sketching and drawing.
If it could paint lines, it would be able to annotate images. Like "circle the chairs" or "sketch a bridge over this river".
It would then be able to help with engineering problems where you give it an image or video of where something is needed, and it will be able to show you it's solution by making a sketch or an engineering drawing. Often you can't efficiently describe a solution using text, and text is the input for image generator AIs, you need to sketch your ideas to find out if they might work.
I lov that you highlighted the "cosmic rays" causing delays! Their data center must be too high altitude!
Ultimate test for AGI will be if the model can read newest paper, digest it, test it and prepare a video describing it in accessible manner in just few hours after the paper was released :)
Haha, rip this channel
that's already ASI level 😉
Very interesting. That positive transfer between modalities suggests a genuinely deep sort of multimodaliity -- with semantic units, concepts, derived from text being associated at a low level with objects and events in images, audio, and video.
Yes, I noticed that with GPT-4. It was trained multi-modally, so even when vision was not available in the product, it understood concepts that made sense visually. I described a Chinese character, and then another one was "just like that, but without the hat". It knew what the character *looked* like, not just as an abstract semantic token, and made the connection between taking off a hat and removing a disjoint portion of the glyph that's at the top.
This is also encouraging for robotics, where similar results have been seen. I think there’s a video on this channel, you can ask the AI.
It truly seems like general intelligence is just … easier than lots of specific intelligences. Which is weird.
Assuming everything goes right, even if we never progress beyond gpt-4 quality, 2023 has given lots of stuff to research for various fields over decades.
Love the depth with which you research and analyze everything happening in the AI space. Awesome video as always!
Thanms Jaoquin !
Thanks for the incredibly fast and complete reporting as usual.
I know right, he's so cool :3
Thanks for covering this release! Plenty more to come!
Couldn't wait for your video on this, thanks for spoiling us!
At the risk of being repetitive ... thank you! That was a marvelous update, conveying both the excitement and the necessary skepticism about this latest big advance in AI. Sidenote: moments ago I asked Bard if it is now making use of Gemini. It responded, in part, "I can't confirm or deny whether I am currently utilizing Gemini for generating answers. Google is not disclosing information about which models are used for specific tasks at this time." Amusing, eh?
Interesting! And thanks as always Clay!
I've asked it now and its confirmed its using gemini
@@mesh8349Yes, in the UK about a day or two after the announcement I looked in my Bard updates and it said it had been updated with Gemini. However, it’s quite possibly not the case in the EU, or wasn’t at the time.
Hey Philip. I was sad at first that I couldn’t afford your monetized content, because you are the only channel on UA-cam that I have notifications on. I watch your content as soon as you post it and I learn and discover a lot from it. The awareness you create is priceless and I hope you make a lot of money for this value you provide. I also hope I won’t miss your valuable information because of payment issues.
Thanks so much my man and having that bell on is enough support already. Maybe one day in a few months or years you can get subbed for a couple of months and binge everything!
- Gemini Ultra's anticipated release next year as a competitor to GPT-4 (0:51)
- Nano model designed for mobile applications, with 1.8 and 3.25 billion parameters (5:24)
- Gemini's training includes data from web documents, books, and code, as well as image, audio, and video data (5:52)
- Gemini's multimodal training improves text understanding by learning from images (6:20)
- Gemini Ultra's image, video, speech recognition, and translation benchmarks outperform GPT-4 (6:59)
- Pro and Nano models limited to text and code responses until Ultra's release (7:03)
- Gemini Ultra's machine translation performance exceeds that of Palm 2 large (10:09)
- Alpha Code 2, based on Gemini Pro, achieves significant success in coding contests (11:26)
- Future Gemini models might integrate with robotics for physical interactions (17:21)
Thank you
Phillip you ability to integrate news and knowledge w/out including hype or noise is really good. Definitely subscribing to Patreon
Oh wow thank you!!
We really appreciate you doing this!
Thanks! Keep up the awesome work, it's by far the best channel on AI with a great mix of deep technical insights simply explained! Just great!
Thanks so much! Would love to see you pver on AI Insiders too but regardless very much appreciated!
Another good video. Glad someone is looking at the facts and finer detail. Yep spotted that 32-shot vs 5-shot straight away. I also spotted all the videos Google pushed out are using bespoke UI, so yeah I bet GPT-4 could do a lot of that too with custom interfaces.
Not long ago i was so excited to run an AI model that contained almost the totality of human knowledge on the internet locally in my laptop now we are getting official support for a state of the art AI model on smartphones,this is amazing I can't wait to try it
Wonderful, thank you very much for sharing your time and work Phillip, peace
Excellent, Thanks :) Content much appreciated. Very informative, even though I've just finished 2 other analytic vid's on the Gemini release.
Nice
Thank you for the great content. You’re the only UA-camr I have notifications on for
thanks owen, let me know what you think!
@@aiexplained-officialGreat video. The custom UIs that Gemini will be able to create are insane. I’m looking forward to trying that out when it’s available
This is by far my favourite channel on UA-cam. I feel like all of my time watching other videos is just an interregnum between your videos being released. Thanks for all your hard work and keen insights. 🙏🏻
Thanks Michael! Super kind
I agree. Its super refreshing to see someone so passionate about a topic be able to communicate such important news in a pretty unbiased and informational way. Especially in an age of overblown headlines, clickbait, and misinformation.
The worst thing about this channel is how much it highlights the ignorance of some other AI channels.
Thank you for the update. I can't wait to directly compare the output of Gemini and GPT4 - for me, the real life output and usefulness is the most important. Traditionally, the new Google products are not available; yet we need to wait.
Thank you so much for your commitment to explaining these topics! The rate and degree of advances in this domain are overwhelming for a non-expert, like me. I very much depend upon people like yourself to help me understand. As soon as I'm able, I'll be sure to support your work monetarily.
I think that, if this technology realizes even a fraction of its speculated potential, then the work that you are doing here will be impossible to overvalue.
A lot of people have already stated the same thing, but I also want to congratulate you for creating a great channel out of nothing and earning the credibility to sit down with insiders and experts. All in a very short time frame too! Keep up the good work.
Thanks so much George
What I find most impressive is how GPT-4V, trained Q2 2022 gets comparable scores for a "natively multimodal model" in Q4 2023 trained on Google's data.
you can't forget that ilya sutskever is an actual wizard. now that he left OAI i'm sure that edge they have on everyone else is bound to disappear soon
@@homemdasneves When did he leave? Isn't he still there, just not on the board?
He left? @@homemdasneves
Gemini is trained in Q1 2023.
Finally some quality content regarding Gemini!
Was refreshing youtube for that!
Thanks! Exciting times. 🙏🏼
CoT@32? Seriously? 5-shot was bad enough. There needs to be more pushback against this behaviour and I’m glad you called it out Philip.
Been waiting for this... Thanks!
Incredible! To do so much work and this quickly! THANK YOU!!
Thank you!
We've all said that many times but let me repeat myself, you're the best in the field of AI News! Thanks for the video!
Wow thanks Rick
And here we go again my man! Just started watching.
Thx i should have waited for your summary. Got the same conclusion. Nice to see
Best ai Channel in my opinion. Thank you
thank you!
Great content as always. I wish i had the money to subscribe to your patreon, you really deserve the support and I'm sure the content you create there will be worth every penny.
Thanks daniel! Your support here means the world regardless.
Great, thorough coverage. Thanks very much for sharing this!
Thanks Mark!
You are the most thorough and credible AI channel on UA-cam for me. I always appreciate your persistent hard work!
thanks ethan, too kind!
And for this reason I have subscribed to Insiders. Thanks for your hard work and insights!
Your videos are one of the most valuable educational materials!
Thanks Murat!
9:50 We all knew that cat was never going to make that jump.
Fantastic. One of the if not the most valuable AI UA-cam channels for me. Pricepoint slightly steep but joined. Looking forward to all the greatness on the Patreon ;). Also: appreciated you so much for calling out the "counter" Hype argumetns while at the same time shedding light on whats REALLY unique about Gemini. (didn't know about the pronounciation part for instance which just yesterday I ran into a wall in with GPT Voice.). And coding obv. Plus the big bummer: Not Europe and UK... LOL :D
Thanks so much Blocky, and for making the leap. I will endeavour to really make it a good decision in 2024.
Thanks for keeping some of it free. I like your professionalism.
Thanks so much, always
Thank you. Another big AI channel I follow, was just lapping it up. You managed to gracefully point out what needed to be pointed out. Bigger picture, Gemini was barely a few percentage points better than GPT 4 and you know that they cherry picked all the examples. Remind me of the Ernie Bot demo.
excellent coverage as always man
Thanks paul!
I am way more excited and impressed by this than I ever was by the gpt4 release
Another impressive contribution to our understanding of developments in AI
Thank you! Great appearance on FLI
@@aiexplained-official well, thank you!
My trusty ai news channel never fails to thoroughly inform me. Thank you sir. Cheers from Denmark ❤
Thanks so much Benjamin!
Exciting look at what's to come. This drills home the need for better benchmarking of AI though.
Thank you for your extensive research. Salute.
Thanks as always Calvin
Another great video, thanks Philip
Thanks Martin!
So grateful you could put this together so fast - Also kudos for calling out the marketing fluff, i.e. the 2 degrees of (im)precision - Embarrassing for the data scientists.
Always on time!
Thank you so much! So fricken helpful
Thanks so much phil
Thank you for keeping posting the free content for those who can't afford the top patreon tear!
Of course! Super grateful for everyone who watches here
I appreciate how you break down the academic-speak for the layman.
it's 4:48AM in India, and I am watching this. Thank you so much for your dedication for all the hard work you do. Seriously BIG THANKS
Thanks Karan!
As a dev i tried it out immediatly using javascript. I would rather bash my keyboard with my eyes closed cuz apparently that seems to yeild better results that Bard's Gemini Pro. I'll stick to GPT-4
Pro is a 3.5 competitor. Not 4.
18:58 I can't wait to be able to afford spring this channel. It is by far the best channel on AI. Comprehensive, informative, and just plain above the rest of AI content out there. Kudos to you 👍
Thanks Derrick, very kind of you! And I will work on new content!
This is the best analysis of the AI on UA-cam
Thanks Dr
Your videos on "let's verify step-by-step" sparked interest in my protein design lab.
Thank you for help in staying relevant and informed.
I will wait for a follow-up.
Also I am very interested on CoT idea "to let the transformers think longer than they are". AlphaFold achieved peak performance partially due to feeding it's embedings back into the transformer block several times.
Also this thought of "iterative convergence to the answer" is somewhat similiar to diffusion and flow-matching. May be good idea was reinvented several times. If this isn't interesting enough I would be GREATLY thankful for any redirections for additional info.
That's very sad that there are not more AI info youtubers as good as you
Thanks for this!
I literally wait for your videos to have the summary of what's going on in the world of AI, so pls keep us updated, thank you
Thanks so much zawarkhan!
Thank you! My trust rating for the G company is 1/10 so I'm glad you're calling out the hype.
One of the most interesting modality demos for me was the soccer coaching, based on video or the sequence of frames. I think this new area of world understanding needs more coverage.
That was fast!! Thanks!!
This part from the paper was pretty crazy:
Training at unprecedented scale invariably surfaces new and interesting systems failure modes -
and in this instance one of the problems that we needed to address was that of “Silent Data Corruption
(SDC)” (Dixit et al., 2021; Hochschild et al., 2021; Vishwanathan et al., 2015). Although these are
extremely rare, the scale of Gemini means that we can expect SDC events to impact training every
week or two. Rapidly detecting and removing faulty hardware required several new techniques
that exploit deterministic replay to isolate incorrect computations, combined with proactive SDC
scanners on idle machines and hot standbys. Our fully deterministic infrastructure allowed us to
quickly identify root causes (including hardware failures) during the development leading up to the
Ultra model, and this was a crucial ingredient towards stable training.
OOOOOOF Course this guys learned how to speak Chinese in China! The most impressive thing for me was when they showed how Gemini basically did a systematic review all on its own during "a lunch break", something that takes teams of researchers more than a year to do, collaboratively! So even if you have to hand hold it and account for hallucinations, that's what, day? a week? Why wasn't this around when I was dying while getting my 2 Master's degrees?
Haha epic comment
Been waiting for the notification lets go
I have never been a big fan of Patreon, but you may be just the person to convince me.
Fingers crossed
Do you do anything else? Absolutely incredible!
Perfect timing
You mentioned the ability to answer questions about Chinese tones, I'm hoping that Gemini will be able to answer questions about Japanese pitch accent! (similar, but quite different to tones)
That would be nuts. It's way too hard to get reliable information on that.
how about all the different filipino dialects 😮
Yeah that's hard
Thank you for your great coverage and insight about this latest AI breakthrough from Google.
Thanks indy
Wow, I am more excited about Alpha Code 2, while most attention is on the Gemini Ultra. I have already been using GPT-4 in my daily coding, and I'd say a coder combined with GPT-4 can finish jobs quicker and solve harder problems than working alone. It seems they are exploring a finer collaboration style between coders and LLM.
I am really looking forward to it and imagining where this could lead.
But if one person efficiency drastically increase then won't the company only need few highly efficient people. This will result in massive unemployment
@@wisdomking8305 It will and is already happening. World went from a huge lack of software developers to a surplus in just a year or so. This will rapidly spread to other industries and in 1-2 years it should be clearly visible in unemployment rates. Many people don't realize that jobs will be eaten by increased efficiency long before AI fully replace humans.
@@johnnoren7244 While I find what you say highly plausible, I am unaware of any present state of a surplus of programmers. And your claims of increased efficiency eating jobs stand in contrast to economic history, where even massively increased efficiency often creates more jobs than it destroys even in the exact industries where the change in marginal output is most pronounced. Coding seems, according to my naïve estimation, somewhere one would expect increased efficiency to result in continued increasing demand since it has so many applications.
@@johnnoren7244Explains why I’ve been on the bench for such a long time then! 😢
In the meantime, I’m trying to learn AI!
Amazing video as always
Superstar, thank you Philip. 🙏👍
thanks alert!
Great video as always!
If possible, please make a video on custom instructions for GPT-4!
Incredible! Thank you so so much! Have you joined AI Insiders? That's where something more slightly niche like CI would go, even started talking about it in the future of prompting vid, after your comment earlier!
@@aiexplained-official Yes, Philip!
it’s my first time watching your videos, and it’s hella interesting and informative, since you have your knowledge in this field what is your perspective on the future tech jobs, cause whenever I go I see people either excited or frightened by this advancement. And for me personally I am excited but a bit worried.
Thank you so much again
Hey Riham, thanks so much. I go more into this on my Patreon, AI Insiders, but the short version is I think there will be continued explosion of tech jobs over the next 3-5 years. After that things might get more convoluted, with lower ranked jobs going. Thanks for watching anyway!
I didn't expect it to be AGI, but it's a hell of a step in that direction. Pretty much what I expected.
Plot twist: Ai Explained is bunch of models combined together to generate this awesome content in blink of an eye. Thanks!
Haha thanks, take that as a compliment
Thank you for making good videos.
Now Google can bathe in the warmth of the sunlight for a few weeks until OpenAI bring out their breakthrough model. I have to say, I am very impressed by he specialized models like AlphaCode2.
What I like best about your channel is that thanks to you I can confidently ignore all the click-baity AI youtubers who apparently live in a constant state of MINDBLOWING ALL CAPS SHOCK. I just know you'll offer the best perspective. Thanks again, I'll try AI insider for sure for as long as I can afford it. Keep it up!
Thanks so much my man
Great video as always
Thanks, you are the best.
Thank you!
Been eagerly awaiting your video
Thanks Ect!
Crazy. Will see if i can integrate this for my language learning app once they have their api available.
Just signed up for full Patreon 🙌
Thanks so much Shaun, look forward to your contributions there! Community growing beyond my hopes
You're amazing.
Very interesting to see that we are now heading towards a point where the best solutions are capped in availability by compute cost and that those solutions are starting to edge into the very upper percentiles of human capability for coding. Surely, when the competition is so close and intense, the application of those models to develop even better models must be irresistible. I wonder where that will take us.
Great stuff Philip, I was looking forward to your video. 🎉
Gemini is very fast! I'm quite impressed with it as well. It told me it has 150,000 token context window.
"The size of my context window is 150,000 tokens. This means that I can process up to 150,000 tokens of text at a time."
Hallucinating already!
@@aiexplained-official Haha, it could be lmao. It also told me it is Gemini Pro as well, and I'm in the UK. It does seem quite good though.
It was happy to help with large C++ and Python ML problems, and managed to correctly guess the movie from a single still I snapped. "From the tall building and blue sky I guess it is Towering Inferno, 1974", and then it went on to tell me it likes playing guess the movie games with me.
That is another thing, it uses inclusive language a lot more out of the box than GPT 4 Turbo. I feel dirty, like I'm cheating 😢😢😂
There is no point reading or watching sci-fi movies anymore 🤷♂️❤
I dont think you mentioned it but @9.55 is even added humour using a pun to its description of "what happens next" in the cat video. That seems like some serious nuanced understanding of human language
jesus bro thats a literally fantastic job
Phillip when I tell you I would give anything to help you write or edit or publish or promote these videos, man...
I am calling it now - you WILL be the Editor in-Chief of a Quanta-level publication solely focused on AI within the next two years, your content is just too high quality to keep to a UA-cam channel and a Patreon
Aw thanks papanokiss. Do you have a background in that kind of area?
@@aiexplained-official Yes! I work as the AI implementation specialist within one of the largest digital equity nonprofits on the globe, so communicating and publishing the "so what?" of the actual development advancements of AI for actual end users (individuals, businesses, gov agencies) is half of my entire job! I'd love to talk more if you're interested in exploring that more
Hey, yeah email aiexplained@outlook.com.