WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=zQZiHOpkq_s

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/zQZiHOpkq_s.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:12.719 --> 00:00:16.880
trickle in from other locations and then we'll get started pretty shortly.

s2
00:00:16.880 --> 00:00:20.880
Um, ideally there will be some time for lots and lots of live demos.

s3
00:00:20.880 --> 00:00:27.439
Um, so you'll ideally learn some of the things about our new models, how to use them as part of your projects.

s4
00:00:27.439 --> 00:00:37.840
Um, and then there also should be time for questions, but it also seems like we'll have a small enough audience where if you ask questions throughout the duration of the presentation, we can also do that.

s5
00:00:37.840 --> 00:00:42.079
Um, so end goal is to make this as useful as possible for all of y'all.

s6
00:00:42.079 --> 00:00:43.120
Um, my name is Paige.

s7
00:00:43.120 --> 00:00:49.680
I work at DeepMind and excited to show you all what we've been working on too.

s8
00:00:54.160 --> 00:00:57.440
And as mentioned, we'll get started in just another couple of minutes.

s9
00:00:57.440 --> 00:01:01.520
It looks like people are still getting scanned at the door.

s10
00:02:38.720 --> 00:02:41.120
All right, let's get started.

s11
00:02:41.120 --> 00:02:41.519
Cool.

s12
00:02:41.519 --> 00:02:42.720
So, greetings everyone.

s13
00:02:42.720 --> 00:02:44.160
As mentioned, my name is Paige.

s14
00:02:44.160 --> 00:02:49.120
If you have questions throughout the duration of the presentation, please feel free to raise your hand and shout them out.

s15
00:02:49.120 --> 00:02:56.879
Um, ideally, we'll be able to take some uh take some as I kind of carine along doing live demos.

s16
00:02:56.879 --> 00:03:08.879
Um, I also want to give a kind of the requisite caveat at the very beginning that um, DeepMind's entire mission is to build AI responsibly and for the benefit of humanity.

s17
00:03:08.879 --> 00:03:18.159
So trying to cast as much light as we can um in this uh in this world that increasingly has uh increasingly has shade.

s18
00:03:18.159 --> 00:03:29.200
Um this translates to things like alpha folds um much of the work that we do with our open models including medge gemma um and a lot of our robotics and science use cases

s19
00:03:29.200 --> 00:03:33.440
as well as things like AI for math AI for frontier research.

s20
00:03:33.440 --> 00:03:42.159
Um, so in general, everything that I show you today, um, is being used for those projects that are deeply deeply embedded in AI for science.

s21
00:03:42.159 --> 00:03:55.599
Um, and if any of y'all work in the AI for science space, um, please feel free to send questions and ask afterwards about how you can use AI to to kind of transform and accelerate that uh that good work.

s22
00:03:55.599 --> 00:04:00.959
Um, I don't think it's a secret that Google has been a little bit busy over the course of the last few months.

s23
00:04:00.959 --> 00:04:06.000
um feels like we've been releasing new models, new features, new products every single week.

s24
00:04:06.000 --> 00:04:12.959
Um most notably, uh just recently we released a computer use API which we'll be talking about a little bit.

s25
00:04:12.959 --> 00:04:24.639
um something called managed agents which gives you the ability to take um kind of a higher order task uh describe it in a natural language and have a fleet of agents execute on it in a Linux workstation

s26
00:04:24.639 --> 00:04:33.360
like a sandbox environment where you can add skills you can add um kind of dependencies that get pulled in along the way and a whole bunch more.

s27
00:04:33.360 --> 00:04:38.800
Um and then also our speechtoech translation API which we'll take a look at in a second.

s28
00:04:38.800 --> 00:04:45.440
And one of the things that I really really love about Google is that not only are we shipping Frontier models.

s29
00:04:45.440 --> 00:04:50.080
So you might have heard about Gemini 3.5 Flash which got released at IO.

s30
00:04:50.080 --> 00:04:55.199
Um but we've also been focused pretty significantly on our open model releases.

s31
00:04:55.199 --> 00:04:59.280
Um so how many folks in the room have heard of Gemma 4?

s32
00:04:59.280 --> 00:05:01.040
Um quite a few hands.

s33
00:05:01.040 --> 00:05:02.240
That's excellent.

s34
00:05:02.240 --> 00:05:05.520
Uh Gemma 4. Oh yes, absolutely.

s35
00:05:05.520 --> 00:05:06.400
Go for it.

s36
00:05:06.400 --> 00:05:09.919
Um, Gemma 4 is our open model family.

s37
00:05:09.919 --> 00:05:11.520
It's the latest iteration.

s38
00:05:11.520 --> 00:05:14.000
Um, we have many different sizes available.

s39
00:05:14.000 --> 00:05:16.560
So, there's a two billion parameter, a 4 billion.

s40
00:05:16.560 --> 00:05:19.280
We just released a 12 billion parameter.

s41
00:05:19.280 --> 00:05:27.360
Um, and then we also have a 26 billion and a 31 billion um, mixture of experts and dense model um, respectively.

s42
00:05:27.360 --> 00:05:41.680
Um, they're useful for a lot of things and they're also Apache 2 licensed which means you can download them, use them as part of your company, fine-tune them, uh, and kind of, uh, expand on them however you feel like would be most useful.

s43
00:05:41.680 --> 00:05:48.080
Um, and this Gemma universe is actually pretty cool to see, um, in terms of people building.

s44
00:05:48.080 --> 00:05:54.000
I I hate slides, so we're going to see how few slides I can get through today.

s45
00:05:54.000 --> 00:06:04.880
Um but uh this is an example of something that someone from the community has built using Gemma 4 um and Fable 5 before it was taken off the market to rewrite some of the kernels.

s46
00:06:04.880 --> 00:06:08.160
Um you can load the model directly within the browser.

s47
00:06:08.160 --> 00:06:12.800
So this is loading Gemma directly in the browser using it via web assembly.

s48
00:06:12.800 --> 00:06:14.240
It's sandboxed.

s49
00:06:14.240 --> 00:06:18.880
Um and then it can do things that that feel pretty magical, right?

s50
00:06:18.880 --> 00:06:23.759
So this model is uh this model is currently running completely locally.

s51
00:06:23.759 --> 00:06:42.000
Um and if I say something to the effect of create a table with emoji comparing and contrasting all of the uh Harry Potter books um based on which are the funniest

s52
00:06:42.000 --> 00:06:44.319
and most exciting.

s53
00:06:44.319 --> 00:07:06.479
um make sure to give me recommendations on which to read and also um um maybe incorporate Harry Potter Harry Potter in the methods of rationality um or and then

s54
00:07:06.479 --> 00:07:11.759
you get kind of a a response that feels almost instantaneous.

s55
00:07:11.759 --> 00:07:21.039
Um, you have uh kind of the Deathly Hollows, Half Blood Prince, Sword or the Phoenix Goblet of Fire, etc. Um, I'm not sure if I agree with the ranking,

s56
00:07:21.039 --> 00:07:26.639
but that's uh but that's a pretty interesting pretty interesting comparison.

s57
00:07:26.639 --> 00:07:33.360
Um, and then also uh since it is local um to the machine, none of your data is getting sent elsewhere.

s58
00:07:33.360 --> 00:07:35.199
This is not using an API.

s59
00:07:35.199 --> 00:07:41.039
It's just something that's running locally in the browser with transformers.js um and the Gemma 4 model.

s60
00:07:41.039 --> 00:07:52.800
Gemma 4 is also pretty cool in the sense that you can use um Google AI edge gallery in order to analyze um some of the model capabilities that we have on device.

s61
00:07:52.800 --> 00:07:56.560
You can download it and use it with Android and with iOS.

s62
00:07:56.560 --> 00:08:04.319
Um and it includes everything from kind of taking images and describing them to automatically transcribing audio in multiple languages.

s63
00:08:04.319 --> 00:08:08.479
I believe Gemma supports over 140 different languages.

s64
00:08:08.479 --> 00:08:13.280
Um, and then also testing it out with function calling directly on device.

s65
00:08:13.280 --> 00:08:20.720
If you haven't had a chance to take a look at the Google AI edge gallery, um, we have a collection of skills as well.

s66
00:08:20.720 --> 00:08:36.560
Um, so things like building games, doing haikus, um, asking queries about weather, being able to schedule events on your calendar for you, um, that are all available to use just with this AI edge gallery app completely for free and just with locally installed models.

s67
00:08:36.560 --> 00:08:40.959
Um, so if you haven't downloaded it, definitely try it out.

s68
00:08:40.959 --> 00:08:52.640
There's also a way to use um the the kind of accelerator on your local device in order to power the model and do all of the inference work as opposed to just the fall back to the CPU.

s69
00:08:52.640 --> 00:09:07.040
Um, so if you do have something like a Pixel 10 or a higherend um a higherend mobile device, you're already able to use Gemma kind of locally um for for all of that work, which is quite cool.

s70
00:09:07.279 --> 00:09:15.360
Gemma 4 just to to give a recap or to place how the model performance versus size shapes up.

s71
00:09:15.360 --> 00:09:23.600
Um this is the two of the two kind of largest versions of Gemma that I had mentioned before the 31B and the 26B.

s72
00:09:23.600 --> 00:09:32.560
Um they're quite small in comparison to some other models that are on the market but they're performing way above what you might expect.

s73
00:09:32.560 --> 00:09:37.920
um so more than models that are in order of magnitude or larger than their size.

s74
00:09:37.920 --> 00:09:42.240
This is great because it kind of translates into a cost savings perspective.

s75
00:09:42.240 --> 00:09:46.160
You don't have to worry about distributed inference for local models.

s76
00:09:46.160 --> 00:09:49.920
Um you don't have to worry about as large of a GPU footprint.

s77
00:09:49.920 --> 00:09:57.680
Um and it's really intended to run super super well on a single commodity GPU as opposed to needing something a little bit more fully featured.

s78
00:09:57.680 --> 00:10:02.480
You can also run Gemma, some of the variants on even things like Jets and Nanos.

s79
00:10:02.480 --> 00:10:06.320
Um, and for 12b and below, you can run it locally on your laptop.

s80
00:10:06.320 --> 00:10:10.240
Um, 2B can even fit handily on mobile devices.

s81
00:10:10.240 --> 00:10:12.959
We've also released quantized versions.

s82
00:10:12.959 --> 00:10:21.920
So, the the model that we saw at the very beginning that was running so so blazingly fast um was using some of the QAT checkpoints that we have for Gemma.

s83
00:10:21.920 --> 00:10:36.000
Um and the QAT checkpoints um are quite small when you uh when you take a look even less than a single gigabyte in size for the the uh kind of two billion parameter version.

s84
00:10:49.767 --> 00:10:51.767
[music]