Multimodal prompting with a 44-minute movie | Gemini 1.5 Pro Demo

[@HappyBirthdayGreetings]

I like that this was done in a honest way

[@CK-kd5pn]

Can you elaborate on what you mean by done in an honest way?

[@ShouryanNikam]

@@CK-kd5pn They faked the last demo

[@leeleo9615]

they released a demo video a while back showcasing Gemini Ultra. The video was carefully edited in a way and sped up to give you an illusion that the model can give really high quality respond consistently and quickly, which is not the truth. To be precise the demo is half accurate, and Google sugarcoated it.@@CK-kd5pn

[@LudwikTrammer]

@@CK-kd5pn They pointed out the sped-up waiting times and showed how long it actually took to generate a response. During the original Gemini launch, they released a famously misleading demo video - it portrayed real-time interactions with the model, which appeared to be based on a live video stream. In reality, the responses were pre-generated from carefully curated static images. It looks like they've learned their lesson, though.

[@TamVuongvan-hj3hw]

@@CK-kd5pn 0:37

[@laStar972chuck]

1. This is amazing ! 2. They clearly learned their lesson about last "edited" demo. Always good to see perpetual improvement in an organisation

[@KinqNick]

absolutely. this time they have a clear advantage against their competitors (with their token length). Last time with gemini ultra it was "just another llm who cares". I hope that it is as impressive as it looks so far.

[@laStar972chuck]

@@KinqNick I truly doubt they can maintain this cost structure for long because that must be expansive to run They'll prolly educate the customer into being patient (long response time) or some sort of job queuing/wait But now they need to catch-up and it's all to users benefits. Gonna be a wild ride while it lasts !

[@KinqNick]

@@laStar972chuck ya could be. Even in the demo some responses were above 60 sec, what I would accept but how would it change, if thousands if not millions use the service.

[@HassanHashemiFarahani]

This is how you create a demo. Simply amazing 👌

[@raaajit1]

This is insane , totally insane 😮usage : you don't have to watch your whole CCTV footage , Instead, you can just inquire what needs to be identified.

[@poxer1]

Great example. Did not think of that use case. Take a document which is 100 pages like a law, insert it and ask it to find a specific part. Or to tell you if the documents of a person allow that person to be accepted in university. Or huge medical file of a patient which needs analysis

[@siloporcen]

@user-vx9tn3iz4n you already know the answer to that question lil bro.

[@apache937]

Yes. and it can be runnning in near real time scanning the footage of all the cameras you have for suspicious activities

[@Whiztlex]

​@@apache937With government having access to all the cameras😈 Big brother is about to become all seeing

[@Danuxsy]

Exactly, the more data the better, EYES WILL BE EVERYWHERE.

[@strikewave1918]

With OpenAI's video generation and Google's video comprehension it seems like almost every aspect of digital media is now processable and synthesizable via AI. Absolutley incredible.

[@paulinaolender6034]

❤

[@RichsrSanch-d3y]

@@paulinaolender6034

[@Hobby_Technology]

The most impressive thing is that when you asked for the time code it responded with simply with "15:34" and not "You have shown me an image of what appears to be a crude drawing of a stick figure below a water tower, with a spout coming from the tower showering water onto the stick figure character ...[3 more paragraphs later]... This is similar to a moment in the film Sherlock Jr. by Buster Keaton in which a character from the movie gets covered in water from a water tower, this happens at 15 minutes and 34 seconds." Sometimes I get really sick of how long-winded ChatGPT is.

[@CK-kd5pn]

Can't you just ask for a brief response?

[@apache937]

Yes but if the context window is long it ignores that @@CK-kd5pn

[@noahwalsh8426]

Gemini advanced has so many guardrails you spend half the time training the model smh

[@TuxraGamer]

I really appreciated seeing that too, it really feels like something that would be compatible with the Unix philosophy (the computer must do *explicitly* what it was asked for).

[@MendigoLouco]

@@CK-kd5pn "Of course, I will try to keep it brief. The time code for the scene depicted in the image you provided is 15:34. Do you need anything else?"

[@julienhovan5725]

I would definitely like to see an example of asking it a sketch or description of an event that didn’t happen in the Movie, as its one thing to find it, its another to recognize that something doesn’t exist.

[@srinivasjakkamsetti2635]

..... 😅😊

[@ИгорьМясников-т5п]

❤

[@Tracesoftradition]

That’s amazing! Can’t wait to see what else Gemini 1.5 Pro is capable of doing.

[@joelhulsey8387]

This is crazy impressive. Wow.

[@zaiologyy]

So neeat! The possibilities are truly endless

[@zergar5671]

This will be amazing for going through 24 hour surveillance videos for police investigators to find the exact crime window.

[@utomo8]

I hope we can have it on consumer CCTV also to better surveillance and alerts. when somebody enter some areas. it can greet or warns by time and known people also

[@mistycloud4455]

that is dystopian

[@DeveloperJS314]

Can't wait to try out 1.5 pro! I am looking forward for 1.5 ultra!

[@Romathefirst]

is the pro the free one ???

[@Ricolaaaaaaaaaaaaaaaaa]

@@Romathefirst yes

[@FartuunMohamed-ke5ki]

1:41

[@Moh1mmed]

Wow, this is really impressive

[@bob3160]

Fantastic. Can't wait to put it to work. Thanks

[@shrikarmadhu2776]

damn!! really impressive!! cant wait to try it out ....

[@erichwschneider]

I have to close my mouth, amazing!!!❤️

[@crosbja360]

That seems very impressive! AI keeps getting better every few months. Good job Google!

[@finbeats]

Every day!

[@CodingAqyanoos]

Wow very interesting. Now you can search in videos. That is amazing.

[@josephhuelbig]

There is an indexer setup you can use in combination with quantized images to find missing people, evidence of crimes, etc. You can draw up quantized images from off sequence outside the square. Needs data mining of quantized images, database, application to Google Maps and index

[@sorryaboutyourcats]

Great to see massive improvements already! When Bard was originally released, it felt rushed out and was confusingly poor considering how long Google has been in the generative AI space for. Looking forward to trying it out. 😸👍

[@Someone7R7]

Openai is in serious danger 😂

[@Tracesoftradition]

I don’t it is. They’re just going to step it up a notch.😊

[@finbeats]

@@Tracesoftraditionsora vs this. Just another race but this time to agi. Exciting stuff

[@ilyass-alami]

Gemini pro Evolved and became stronger ❤❤

[@SanaagSomaliland]

Minority Report moment is coming sooner than I thought.

[@SanaagSomaliland]

If this is what the private entreprises are now capable of, imagine what the Gov't has been capable of secretly.

[@3ndriu999]

I don't understand why I've always seen the demos perform better than the official versions...

[@finbeats]

Your Prompt issue

[@DevolperOperation]

Stunning.

[@dishcleaner2]

This is bewildering.

[@gabrielfernandez3782]

Amazing, simply amazing.

[@raphetsonatelier]

Simply incredible 😮 I can't believe it can parse that much information in so little time

[@mohamedkarim-p7j]

Thank for sharing👍

[@chlee4256]

As impressive as this one is, let's hope Google doesn't get called out like the Gemini demo.

[@anav587]

Doesn't even matter, I doubt they're lying about the 10m window

[@nawabifaissal9625]

@@anav587 yeah same i doubt they would lie about that especially since there's not really a big marketing project to introduce a new model or else, it's "just" an improvement

[@pebre79]

Also their Google Assistant making phone calls demo :/

[@AkameAGKxv]

Cool😮❤️‍🔥

[@muru603]

Freaking amazing!

[@_ahrorjon]

It’s absolutely insane 😍😍😍😍

[@Илья-е4э8ю]

I remember 2 years ago, I wanted to find something like these models, But I didn't find them, it was like a closed door with no doorknobs. Thank you for making my dreams come true, This is a dream come true, thank you.

[@_abdul]

Yup, We're There.

[@ayukay4538]

Sorry google, i didn't have 2 minutes to spare watching this demo video, so i asked an AI to watch it in my stead, and to produce this comment as a result.

[@mansoor8228]

😂😂

[@sumansaha295]

Now this is impressive

[@ilyass-alami]

We know that Gemini is still constantly developing, because it does not currently compete, and Google is a giant company It should be the queen of artificial intelligence, and provided to users for free

[@amjadhussainamjad786]

Good luck Gemini

[@darkphoenix2]

Ok, I don't want people like artists to lose their jobs to AI, but this kind of thing is really cool. Don't know how useful it will be, but it's cool.

[@wrcz]

thanks, just bought some Google stock

[@vikramjeetsingh4627]

Now this was cool!!

[@abhiram978-]

this is soooo good

[@ShpanMan]

Awesome work. Competition will bring the AI gods faster.

[@florianfeldhaus6478]

Mindblowing

[@HiddenExp]

This means we could generate long and consistent storytelling, damn! And longer AI NPCs memory!

[@web3global]

Just WOW! 🤩

[@gptty]

Good job

[@mistycloud4455]

The speed of ai progress is insane

[@kjb8449]

Super!

[@jeanchindeko5477]

Gemini Pro 1.0 was released in December and 2 months after we already have a huge upgrade with version 1.5. What will this world look like in December 2024?

[@wanarchives]

When the real dev makes their own video demo instead of marketing team

[@peztopher7297]

I'd think the simpler the drawing, the better. You have to decide what the important features are. Detail would be noise. Years back, I realized this was how cartoons (images, not video) work. What is the minimum representation, for example, of a mouse?

[@MichealHawrylyshen]

Where was Base 211 located? Wasn't Feugo Patagonia or the Falklands...

[@muhammadasiffarooqi7672]

Mind blowing

[@alexissvetrev]

I. Amazed with all these, how come no one is building home server robots. All the elements are there...

[@compareFirepower]

This is nice.

[@bensoos]

Are these voices artificial?

[@FazriIhsan]

When will the Gemini app be released in Indonesia

[@dimii27]

Am I the only one who finds it funnny that this is filmed on a mac?

[@ZweSim]

Amazing

[@johnflux1]

Very very nicely done. Absolutely amazing, and so much better video than the last one.

[@eeuno-v2r]

the stock price reflect everything

[@aigriffin42604]

Why does Gemini require tokens?

[@r5LgxTbQ]

just speechless

[@danielmartinmonge4054]

This is 1.5 but PRO is It? Not ultra. Just giving a crazy amount of context to the more basic model right?

[@eyescreamcake]

How much did that one minute of compute cost?

[@jd5787]

Great, i will know how to use gemini next time I need to get info on a movie and associated time code. Now: how do I leverage gemini for my business? Pre-sales, sales, finance operations for instance... How about being able to get proper analysis of legal documents, suggestions and clauses drafting (which chatgpt can do pretty well) before sending to a lawyer for vetting/approval? How about using Gemini in Google meet to get call report/summary that can be inserted in a CRM? How about being able to get relevant mockups, designs and images that can be used and exported in usable format (eos, ai etc). Right now, it looks like Gemini is a powerful model with limited use cases.

[@deejay2023i]

This app not available in France

[@hellodavidryan]

Dear future overlords, be merciful. Signed, squishy humans.

[@eliporter3980]

Man this stuff is accelerating fast.

[@wellingtonmnf_ti]

I can't find the correct words for explain what I saw right now...

[@poxer1]

Ok, so inteligence and knowledge is now owned by companies like google. Time to learn carpentry

[@mansoor8228]

Nice

[@Yewbzee]

It baffles me that so many “AI experts” are still predicting AGI is many years away. I can only assume they are doing this in bad faith in an attempt to fend off heavy handed regulation.

[@christoschri9193]

Why was Gemini created? you can just improve Google Assistant .... you can just merge the two apps into one ...

[@pablonolberto6557]

Openai already feels cornered, how is it possible that Google will present Gemini Pro 1.5 some after Openai takes its ace on generating videos.

[@eclectice]

Now I can ask Gemini Pro 1.5 to identify any potential shoplifters.

[@MohammadFarman-dr9hg]

Samsung and Google system fuchsia os 🎉😮😊❤

[@eukaryote-prime]

Its a good thing that liquid was blue

[@RUMPshit]

Is google just using RAG calling it 1m token context?.... I dont think the LLM actually reads 1m tokens...

[@hooooman.]

Whats that token actually means? Why google raised to a new standard of 1 million token?

[@FirasMohamed96]

Can you Imagine Gemini 10.5 what would be? 😳

[@theeternalnow6506]

I think we'll have an AI overlord before then at this pace.

[@killstreak4767]

This needs a separate app

[@na-gi1xb]

🤫🧏‍♂️

[@user-ji1ko2vs5g]

❤ 0:32

[@TwoTeaTee]

Holy sheets!

[@JazevoAudiosurf]

i do believe this demo

[@maylahsyari]

why google sign in😶

[@PeggyBoggs-lr7xc]

0:50 1:01 10 🎉

[@tony.cortez]

Application for this is, use in investigation.

[@lowlufi]

Nex gen: Jarvis

[@vietlongpham]

Now who could say GPT is better. Google beat them up

[@lomfulanolwazi7241]

amen

[@JOHN.Z999]

An efficient multimodal system with a significantly expanded context has impressed me! This is precisely what I need, Google: efficiency and improved reasoning from these models. ❤

[@rahneshin752]

❤❤❤

[@ApatheticPerson]

I really hope this is not fake just like the last time.

[@Mackcolak-xf5bk]

Here are some key takeaways from the video: 1. Gemini 1.5 Pro is an experimental new model from google that can understand very long context - up to 1 million tokens. 2. They demonstrated this capability by using a 44-minute Buster Keaton film (over 600,000 tokens) as the context. 3. The model was able to find a specific moment in the film when prompted - it located the exact timecode when a pawn ticket was removed from a person's pocket, and extracted accurate text details from the ticket. 4. The model was also able to match a simple drawing to the correct scene and timecode, showing an understanding of multimodal input combining text and images. 5. Long context understanding opens up new possibilities for complex and detailed questions about books, films, etc that weren't possible with previous models. 6. Responses will vary and may not always be perfect, but the examples show promising capabilities in extracting details from extremely long context. 7. Simple drawings can be an effective way to probe the model's understanding without having to describe the full context again. In summary, Gemini 1.5 Pro demonstrates an early but meaningful capability in long-form reasoning and understanding from both textual and multimodal context.

[@CK-kd5pn]

Anthropic?? Don't just copy paste from ai bro read it over

[@apache937]

bot

[@Mackcolak-xf5bk]

@@CK-kd5pn were these takeaways wrong?

[@CK-kd5pn]

@@Mackcolak-xf5bk Just look at the first point...

[@Mackcolak-xf5bk]

@@CK-kd5pn i see now, thank you.

[@omkarchavan5940]

Google using mac os and not chrome os?

[@ianlewin8888]

Next step would be making the AI can tell what's real and not.

[@BalveerSingh4956-t4t]

और सुनो जी आप सत्य मार्ग सबसे बडा धर्म है अलवर राजस्थान भारत से बलवीरसिंह ❤