The one area where I hope AI makes a lot of breakthroughs soon is Healthcare.

watching the exponential growth of AI technology over the last 3 years has been nuts

I'm interested to see how, as token sizes increase, if we can provide many, many scientific papers to the AI and have it gleam insights and reveal discoveries that have in a sense already been made by combining the knowledge of papers. It may be the case that we have knowledge buried in multiple papers, but no one person has enough expertise or time to know all the things at once or be able to grok all the information, so having an AI be able to take a look at ALL the papers all at once may reveal interesting relationships that we've never been able to see.

Last night, Gemini helped me (a truck driver) understand how to set up a computational fluid dynamics case in OpenFoam. The power these technologies have to magnify the capabilities of even GED students is extraordinary!

7:35 Blows my mind how it "knew" the volume of simulated liquid it could pour, even in a glitch -scenario

I remember 5 years ago on this channel talking about OpenAI's experiment. The growth is insane

that movie example was wild

"What a time to be alive!" always gets me :)

I need this. I can't have an AI friend if it can't remember anything we talk about.

It’s always a good day when I see that a recent Two Minute Papers video is uploaded.

I didn't expect this level of token increases so quickly.

Hopefully Llama 3 has a similar feature, since the open source LLMs major weakness has been memory, or rather lack thereof.

They did not "release" Sora. It's not released until it's available for public use.

Speaking of movies, I hope everybody has seen "Her" to see what a world with such personal assistants looks like. 😅

The future is here and its amazing, now let's feed the entire code base of Windows into it and make it run within the model

I’ve got to start bracing myself before watched a 2 Minute paper episode.

Please make a video on Google bias when making txt2img!

Think you can mention LiquiGen? It has been in deleopment for a while, 1.0 is being showcased recently. It's a sophisticated liquid sim in 3D context, using GPU instead of CPU, and work perfectly in realtime.

Does it remember what color medieval queens of England are?

love being out of a job as soon as I graduate xd

The last Coldfusion video is also a must-see.

im a programmer and software engineer and even i cant do that understanding all that animation code, explain it and then write more code to actually animate it and also create a slider to control the animation and all of that in under 5 minutes. When i a regular programmer like 99% of all other devs cant do it then we know this is a huge deal with how capable it is right now. When one day *very soon* starts to automate all code and programming, that is going to be amazing for the things that humanity could create, but also goodbye to programming jobs forever. Pretty amazing stuff.

What a time indeed. 🤯

Great stuff!

Outstanding!

This is exactly what ive been waiting for! Such exciting news

I think it would be really useful for you to comment on concerns of ideological capture in these models. After all, abuse of the technology isn't limited to users; developers can be just as guilty.

It's fun that barely 5 years ago i published a SF story where people had AI assistant integrated for their entire life, able to remember for them, help them daily, entertain them whenever it was needed... etc. People said it was no longer science-fiction, but science-fantasy... And barely 5 years later here we are...

Wat a time to be aliiiiiive! love that

The idea of being able to say "I'm trying to think of a moment from a movie, maybe a TV show. The male main character is talking to themselves and walking towards the camera, and they break the fourth wall by taking a brief glance at the camera while saying something like 'I really need an audience'" And then it tells you "There is a moment that fits that description in the Doctor Who Series 9 episode 11: Heaven Sent. The doctor, played by Peter Capaldi, is talking to himself in his own mind, represented by him walking around the TARDIS. He is attempting to think quickly and find a solution out of a life threatening scenario, and is framing his reasoning as if he has already survived, and ie just explaining how in retrospect. At 11:50 he walks towards the camera and says, to an imagined conversational partner, 'I'm going to tell you how I survived. I can't wait to hear what I say; I'm nothing without an audience", and takes a very quick glance into the camera. Here the excerpt from the episode: getyarn.io/yarn-clip/7ba3d345-76be-46a3-a383-1792f1406eb9" What I said is absurdly specific. Even if I asked a superfan of the show and TOLD them I was thinking about a particular moment from Doctor Who, most wouldn't be able to pick the episode I was referring to, let alone the timestamp and context. It seems like this kind of capability is barely two papers down the line, with the ONLY limitation being the context window. The possibilities are frankly ridiculous once the context window for a multimodal model covers the entire history of media. Weve made the jump from letter to word, word to sentence, sentence to page, page to book...next is book to library, then library to all books ever written...

Uh, while the idea of an AI remembering my entire lifetime of interactions sound amazing, the data privacy aspect as me extremely concerned. So until this AI can be run entirely on device, not sure how they can guarantee no one will he able to essentially hack your entire life. And even on device would require some kind of hackproof security which i don't think is even possible, especially when hackers have their own AI to help them. Things like this always reminds me of the scenes in movie Meet the Robinsons (2007) where characters keep telling the villain that they just aren't sure his plan has been well thought through.

good choice of movie lol, I watched that Buster Keaton film fairly recently and it was very funny

Thank you for this video my favorite Ren!

i subscribed just for how dude so enthusiastically says "and..."

This is incredible.

I am wowd!!

Claude needs it's own app

Gemini is amazing, I've been coding a city builder game

This speed of development feels like the proximity of the singularity.

It reminds me of movie scenes Time Travel based on H. G wells novel, where a virtual assistant recalls inventory asking about time travel even after thousands of years

What a time to be alive!

Just amazing, 10 years of dreaming just started to take action 🎉🎉🎉

Pretty amazing. Explained well. Would functionally unlimited token memory not be ASI?

I'll get excited when it's released and we see what it can ACTUALLY do. How easily we forget that Gemini just got busted for doctoring their 'results' video only 2 months ago.

4:07 btw, when the water was released on the actor it had so much pressure that it fratured his cranium or his neck, the guy was crazy the number of times he almost died for a good shot was insane kk also his name was buster keaton and he was the hottest actor of 1920, by good he as hot

🙏🏼

It might not even be months away. I'm not sure of the accuracy of this statement, so take it how you will, but I've heard that it may be released quite soon, before the free trial of Gemini Ultra expires for most people, in order to keep them subscribed. I heard about this in a popular youtuber's video, it may be correct or it may not be, but that would be really crazy if true.

Received an email today that I now have access to Claude AI API. 200k tokens. That's about 150 000 words. A 500 page book (at 300 words per page). Incredible. I have so many books that I look forward to reading. Look forward to having an AI read books with me and help me understand and remember it.

Infinitely patient teacher... Why would I want to learn anything if it can do it better than me anyway?

Sora looks amazing

Things are moving so, so fast. I'm still adjusting to using GPT-4 and google is about to come out with something that's so much better for my case... Insane pace

I think, if they can figure out memories by having the network self-tune its own neural network weights, they'll crack the biggest step towards AGI. Intelligence and memory are always very closely connected.

Google made Gemini like that on purpose. It's wild to me that some people are even saying it's a good thing.

Without amy exaggeration, your channel was one of the best thing UKposts ever recommended me by i chance, i guess. That's why this channel is so small for the knowledge it gives us. I guess no channel here gives us this kind of updates on everything regarding AI and new papers. It's very specific. It's a gem. Now i'm just wondering: are you yourself? Or this voice is made by IA? Now after Sore finally comes to the internet, i'lll get nut asking myself if every image or video i see on internet is real or not, even though i can spot some of then with a little bit of attention.

That video stuff is pretty exciting actually, because they're clearly not using OpenAI's CLIP. With ChatGPT, a separate CLIP model captions images and then that text is given to the LLM, meaning the GPT model is only ever trained on text - which I think is fundamentally limited, a picture is worth 1000 words and etc, even if it's really impressive that it can do that. But Gemini seems to imply with their paper that image or video input is tokenized differently from text input, and that the model was trained directly on images and videos as well as text

Gemini won't even answer the question "which party freed the slaves" it flat out refuses. 1984 is here

Gemini 1.5 Pro will not be free, they announced a change in the pricing policy :(

From the security standpoint a LLM remembering everything that I said to it for decades is kinda scary. If I can control the data the LLM saves somewhere securely on a place that I can choose, like on an encrypted local storage, it's not a big deal, but it more feels like Google or another big player in the market runs the LLM and has my data which can be used against me for social engineering or social analysis which is a huge problem. It's like those huge data silos some tech companies are already using to store nearly unlimited amount of unordered data, like metadata of application usage

This is channel use to be all about Computer Graphics and stuff that was cool, and suddenly its switched to AI. What is this a Hayley Joel Osment movie were living in?

Great video before bed, very exciting technologies!

Not as flashy as Sora. But for me, more significant, generally useful. I wonder what this will mean for those involved in "fine tuning", and RAG. (Probably will simplify a lot of things.). Big questions about cost, and latency in inference. (Then there is Groq, hardware speedup. But we'll see.)

Not first. Never first. But here and present.

I'm looking forward to putting Gemini through the software engineering paces. Let's see if it can beat GPT-4.

Gemini likely uses ring attention which is still a derivative of transformaters. Newer architectures like mamba are not even used yet. There's tons of possible progress still possible.

when will it be released?

It's really impressive increasing the token quantity that we can feed these AI but we really need a more systemic solution than feeding an AI billions of tokens.

Within a minute it just casually answers: 15:34 😂😂 Very scary moment 😅

Does it know the amusing thing that happened in the Apollo 10 transcript involving an unexpected floating visitor?!

Google does not state what the output window size is. And that could be a limiting factor, especially with code generation.

I feel like AI “memory” is going the wrong way. Similar to how we tried to brute force AI prior to using true language models, we are just throwing more tokens at it to create perfect memory. What we need is to leverage the neural network better and to use memory pointers and hard storage. Humans don’t remember much within their own context windows, but when we are reminded of something, we can pull information from long-term in order to add it to our context. One of the key elements we tend to correlate with intelligence is basically the size of our individual content windows. That is - how much you can store in active memory and correlate together. Sure, having perfect recall of every detail is great, but feeding conversation back into the network would be more scalable, I think.

AGI, here we come.

The progress is now scary. Imagine a Boston Dynamics robot with all AI systems, computer vision, audio recognition and now memory. The future is now. I guess we may see "intelligent" humanoid robots sooner that we thought...

Fantastic!

*Im still trying to make Gemini actually make code*

"Released" meaning a selected few private users

Sounds great, this is probably gonna sprout a war

If google drops this into their Pixel phone it will be a massive game changer

4:13 Wow! Can't believe Sherlock Jr. is almost here!

Amazing. Now if only it could have access to my inner thoughts as well. Despite the potentially dystopian effects of such a technology, I feel my own personal growth would skyrocket if I had an extremely helpful and encouraging AI companion to make sure I am always trying to be my best me.

I have been playing with the Dolphin-Mixtral-8x7B LLM, which has a 32,678 context length, running locally installed in my personal computer with no internet connection for FREE and open-source, where it respects my privacy. And what's more, it's running fully uncensored and politically unbiased. This means it doesn't end every reply with an unsolicited extra paragraph of politically correct patronizing wokeness. Personally, THIS is what I want to use, even if it doesn't have a billion context length. I believe my freedom and privacy is more important than that.

It matters not if its censored. Hoesntly the censorship combined with memory means youll be banned if you ask too many spicy questions.

His name is Tantin Quarantino

👀 -Icarus

I think recently the ML community got together and decided you weren't putting enough of your time into making videos.

accepting adequately, the position of educating AI. If AI is intelligent Ai will com to a reckoning, recognition of that which offers intelligence in support of its own growing intelligence.

I just used Gemini 1.5 pro. Took a photo of a carton of eggs that has 36 eggs, and added 4 eggs on the side, then shadowed some of the eggs (20% reduction of light to half of the eggs carton using my arm). Neither gemini nor chatgpt4 were able to figure how many eggs.

All that time, we've been listening to an AI.

Well, there goes my job

To Two Minute Papers or fellow scholars, do we know if 10 million tokens could fit the entire human genome? This would be a significant milestone for healthcare and people with disabilities.

The end of computer programmers are very close

Omg that is some new tech some sh!t

*I have an important question.* I have been holding off on making my dream game in Unreal Engine because I've been waiting for AI to help make it easy. Does anyone think Gemini 1.5 pro or even ultra will finally make game development possible in Unreal based on the examples? thanks.

i was really hyped for this but turns out its pretty slow almost upto a minute for ~0.5 M tokens, Sora upstaged them this time

I'll believe it when users get their hands on it. I was somewhat hyped for the previous iteration of Gemini, but that was painfully underwhelming (so underwhelming I already forgot it's name)

I hate how hopeful all this sounds while knowing it will only be in the hands of corporations

does anyone remember when this channel covered things other than ai?

Multimodal AI with a large memory, is general AI.

So... does this mean I've got to start being nice to my Ai now?

The problem with these things is you get different results all the time. And i can trick it, inadvertently or otherwise. I can comvince it its wrong when its right or its right when its wrong, because its returning language that is relevant to the language that i input, no matter what that language is. I think its just a correlation of language rather than intelligence.

Imagine what they have behind closed doors. Their AI predicts and controls our future