Latest / In The Loop / Two Insane AI Models By Google That Broke The Internet

Transcript

Two Insane AI Models By Google That Broke The Internet · In The Loop

• 0:00
  This week we have a new language model announcement, and in fact,
• 0:03
  not even just one, but two announcements, and those two
• 0:07
  have broken the Internet and they're both from Google.
• 0:10
  Now, we know that Google, you could argue, has been not
• 0:13
  winning the AI race. You could probably argue that
• 0:15
  it's not losing, but it's definitely not winning for some
• 0:18
  time. But obviously now everything
• 0:20
  just changed to the point where Sam Altman has come out to his
• 0:24
  entire company and said that he is worried.
• 0:28
  He's worried about slower growth and huge competition.
• 0:31
  And for context, the Gemini app and the Nana Banana applications
• 0:35
  have hit #1 on the App Store, have generated hundreds of
• 0:38
  millions of images and have gone completely viral.
• 0:42
  It's scored higher on almost every single benchmark than most
• 0:45
  language models out there. And it can do things that make
• 0:49
  image generation finally truly useful for everyday users.
• 0:54
  Anyway, in today's episode, I'm going to go through the
• 0:56
  different things you can now do with these Gemini models.
• 0:59
  This is in the loop with Jack Horton, and I hope you enjoy the
• 1:02
  show. So let's start with a little bit
• 1:16
  of background on this because on November 18th, Google launched
• 1:20
  Gemini 3 Pro, the main competitor model to the Open AI
• 1:26
  Chachi BTS of the world. Two days later came Nano Banana
• 1:30
  Pro, which is their image generation model.
• 1:32
  Now officially it's called Gemini 3 Pro image, but
• 1:36
  everyone's been calling it Nano Banana from the previous earlier
• 1:39
  versions that have been teased online.
• 1:41
  And if you've been online recently in the AI community,
• 1:45
  you will know that this model has been hyped by the community
• 1:49
  in a way that we've probably never seen in a model since
• 1:52
  early Chatchi BT days. Now, if you remember the episode
• 1:54
  I think five or six weeks ago where I talked about the new
• 1:58
  Google Pixel phone having AI in it, it was actually overshadowed
• 2:02
  in some ways by the fact that everybody was hoping that Nano
• 2:06
  Banana was going to be released and they didn't release it.
• 2:09
  So that gives you a kind of insight into just how much hype
• 2:12
  there's been around this model release.
• 2:14
  And I mean, they're really building Google here because
• 2:16
  they've got about 650 monthly users on their Gemini
• 2:19
  application. They have 13 million developers
• 2:22
  now building on top of their models, and now both of these
• 2:25
  models they've just released are available inside the main
• 2:27
  applications and as we speak, being rolled out across every
• 2:31
  single product feature you can possibly think of.
• 2:34
  Now let's get into all the cool things you can now do with these
• 2:36
  models. OK, so I think of the two models
• 2:40
  I want to talk about today. So Gemini 3 Pro, which is kind
• 2:42
  of the competitor to Open AI, ChatGPT and Cloud, and then the
• 2:47
  Nano Banana, which is the competitor to the likes of Saura
• 2:51
  and other big image generation models.
• 2:53
  I think I want to start with Nano Banana because it's
• 2:56
  probably provided some of the most exciting new capabilities
• 2:58
  I've seen in a very long time. So the first major capability
• 3:14
  it's developed and built on is the way it deals with text.
• 3:18
  So it's now supports every single language and deals with
• 3:22
  text perfectly. So every other image model
• 3:25
  breaks when it comes to text. You know?
• 3:27
  Ask Dali or other providers to create a product label with
• 3:32
  English, Korean and Arabic. You often get a lot of mistakes
• 3:36
  and AI slop mid journey. Can't handle non Latin scripts
• 3:41
  at all. Chinese characters might come
• 3:44
  out as random strokes, Arabic reads right to left incorrectly.
• 3:48
  There's just a lot of errors, and it's not something that's
• 3:51
  easy to do because there's a lot of detail there.
• 3:53
  Well, none of Nana completely flips that on its head.
• 3:55
  And it enables you to do really cool things.
• 3:57
  So first, it can generate multilingual text from scratch.
• 4:01
  So someone tested, you know, generating product cans with
• 4:04
  text on it across different languages, and it worked
• 4:07
  perfectly. Second, big leap is the way it
• 4:10
  is able to now translate text that's already in images.
• 4:13
  This is genuinely new, you know, take a photo of a friend's
• 4:17
  restaurant menu and it'll be able to translate it to English.
• 4:20
  But going even further than that, it can match handwriting,
• 4:23
  or it can take information and text from one image and generate
• 4:28
  it inside another. So one of the most amazing
• 4:31
  examples so far I've seen is its ability to even get this example
• 4:35
  up on my phone. Here, match your handwriting so
• 4:39
  you can actually write something on a notepad.
• 4:41
  Imagine writing on a notepad, a maths equation or a problem and
• 4:45
  then asking it to replicate that handwriting but complete the
• 4:48
  equation or write a page of notes and it will just do that
• 4:52
  on the fly and it matches your handwriting.
• 4:55
  So yeah, the way it deals with text is incredible.
• 4:57
  Another one as a result is its ability to just generate entire
• 5:00
  infographics and visual ways of representing information using
• 5:06
  text. Again, this is something that
• 5:08
  models just sucked up previously and anyone that's tried to do
• 5:11
  this has probably been disappointed over and over
• 5:13
  again. So you can now just generate
• 5:16
  entire amazing whiteboard and infographic images instantly.
• 5:20
  So just imagine how this impacts marketing teams, for example,
• 5:23
  generating entire marketing campaigns for different regions
• 5:27
  at scale. You can just imagine Facebook or
• 5:30
  meta ads that changes ads for not just a country, but a region
• 5:36
  or a street or so many different layers of subculture.
• 5:41
  Now, I, I personally find that kind of dark and terrifying, but
• 5:44
  you can see where this rolls out to and how this impacts the way
• 5:49
  things like marketing and information sharing go.
• 5:51
  Now, I, for one, find this quite a almost dark and scary concept.
• 5:55
  I always find big advertising and using data to generate new
• 5:59
  images to persuade us to buy things a little bit.
• 6:02
  Yeah, as I say, dark and evil. But you can see how much of an
• 6:05
  impact this language model is going to have in that arena.
• 6:08
  So another major upgrade here, number 2 is, is its ability to
• 6:12
  maintain image identity across many different images.
• 6:16
  So within say 48 hours of launch, you suddenly saw images
• 6:20
  of sort of Elon Musk SAT with a cigar with the Google CEO, with
• 6:25
  Jensen Twangs, Sam Altman, other people like that and it looked
• 6:28
  completely real. Now a lot of the time other
• 6:31
  image providers, they might be able to take lots of single
• 6:34
  images and combine them into a single image, but often it's
• 6:36
  just not that good. With Nano Banana you can blend
• 6:39
  up to 14 images at once whilst maintaining the appearance of up
• 6:43
  to five people. So now suddenly you've got
• 6:46
  people being generated in 3D. People uploaded image selfies
• 6:49
  and turning themselves into figurines.
• 6:52
  Now the people from A16Z use this for their leadership teams
• 6:56
  infographics. So they had five exact headshots
• 6:59
  with their actual appearances, family photos from individual
• 7:01
  shots, and they blended it together with Nana Banana and
• 7:05
  everyone looks perfect. But it's now in a completely
• 7:08
  different context. That looks super professional,
• 7:10
  but exactly like everybody now going on to #3 which really
• 7:14
  blows my mind is how it can ground real time data with its
• 7:18
  image generation. So previous image models might
• 7:22
  only know what they learned during training.
• 7:25
  So you asked Dali to create an infographic showing today's
• 7:27
  weather. It'll invent A plausible looking
• 7:30
  weather icon or image, but it won't make any sense because it
• 7:33
  won't be the actual weather. Or perhaps give it some
• 7:36
  coordinates of an event that's just happened and ask it to
• 7:39
  generate an image of that. What that event would have
• 7:41
  looked like, it wouldn't have be able to do that.
• 7:43
  We'd have no clue. Well, Nana Banana now can use
• 7:47
  Google search for injecting that real time information about
• 7:50
  what's happening in the world that second to help it generate
• 7:53
  that information and image. So you can say generate me an
• 7:56
  infographic or a visualization of an event that happened today.
• 7:59
  This search just happens automatically within Nano
• 8:02
  banana. You don't have to use a
• 8:03
  different model or see. I just think about the output of
• 8:06
  creative in the world that's going to that's going to be
• 8:09
  online as a result. I mean, anyone that's been
• 8:11
  online recently, you can just see so many Nano Banana images
• 8:14
  and videos out there all the time right now.
• 8:18
  And I mean, one of the things that I found amazing here is you
• 8:20
  can things like you can upload GPS coordinates and just ask
• 8:24
  Nano Banana to generate images of what that event might have
• 8:27
  looked like, both historic events and real time events
• 8:30
  happening today. See, it's safe to say that I've
• 8:32
  been blown away by this new model update.
• 8:35
  It's really set the tone and expectation of things going
• 8:38
  forward. Now moving on to Gemini 3 Pro
• 8:54
  and it's upgrading capabilities and importantly how good it's
• 8:58
  got at coding. So this is the fourth big
• 9:01
  upgrade by Google. Here they claim that it's going
• 9:05
  to have multi hour autonomous coding sessions because previous
• 9:08
  coding assistants, you know, like Cloud Code or others might
• 9:12
  run for 30 minutes, 40 minutes, 45 minutes without losing
• 9:15
  context. Cursor can handle maybe an hour
• 9:19
  of continuous work before needing to be restarted or
• 9:21
  readjusted. Or they hit token limits, lose
• 9:25
  track of what they're doing, or just start to make a lot of
• 9:27
  errors without any human intervention.
• 9:29
  Well, Google's what they call anti gravity agents apparently
• 9:33
  can work for over 24 hours continuously.
• 9:36
  So Google's documentation says in our internal evaluations we
• 9:40
  observe Codex Max works on tasks for more than 24 hours.
• 9:45
  Now apparently it uses something called compaction, which is a
• 9:48
  technique that essentially prunes its history whilst
• 9:52
  preserving contacts. So if you imagine a model going
• 9:54
  out there and doing things, it's constantly updating the
• 9:57
  information about that task. Now if you can imagine you
• 10:00
  working for three hours with this agent, it suddenly
• 10:04
  collected 8 hours worth of information, contacts, things
• 10:07
  that you said you do and don't like.
• 10:09
  And after a while it's trying to take too much in every time you
• 10:12
  want it to do a new thing and so it's performance can drop.
• 10:16
  This innovation essentially is auto reducing, improving and
• 10:20
  compressing the amount of context it has so it can be more
• 10:23
  efficient and therefore go on for a longer and longer amount
• 10:26
  of time. So this means that Google claims
• 10:27
  that you could refactor an entire code base that might take
• 10:30
  weeks of time with a single prompt.
• 10:34
  Now whether that's true or not, we'll see when we start to test
• 10:37
  it, and I'm sure the team internally at Mindset will be
• 10:39
  testing this very soon as well. So moving on to the fifth big
• 10:42
  update by Google, you can now build complete games, which
• 10:46
  although might not sound exciting, these games can have
• 10:49
  complete physics and audio instantly.
• 10:52
  So previous models could generate a game quite simply,
• 10:55
  but it doesn't actually always work.
• 10:57
  And you know, you could ask Claude or GBT to build a game
• 11:00
  with physics. You might get code that's quite
• 11:02
  plausible, a very simple prototype that will sometimes
• 11:05
  will work OK, but often has a lot of bugs in it.
• 11:07
  Well, Gemini 3 in their demos built entire games like
• 11:11
  Ridiculous Fishing, the iOS game from 2013, which had working
• 11:15
  mechanics, sound effects and music like the full game, and it
• 11:20
  had accurate physics for catching fish.
• 11:22
  And previous models really struggled with this types of
• 11:25
  activity, whereas Gemini 3 can do it in just a few prompts.
• 11:29
  Again, devil's in the detail here, and we'll be really
• 11:31
  interested to see at what point you get stretched to its limits.
• 11:36
  OK, moving to number six, which is the ability for Gemini to
• 11:41
  read very complex screenshots and take action on them.
• 11:44
  So previous models are really good at describing what might be
• 11:47
  in a screenshot, but they can't interact with it.
• 11:49
  So if you give it access to your web browser, it can take
• 11:53
  screenshots of things maybe. So this is agent mode, but it's
• 11:56
  not very good at figuring out, you know, I need to press this
• 11:59
  button and that button to do this thing.
• 12:02
  So it could tell you, you know, the information on screen, but
• 12:04
  not really good at figuring out where to press.
• 12:07
  And that's why I think agent mode and autonomous agents, they
• 12:10
  get branded online have not been hugely successful or popular.
• 12:14
  Well, Gemini 3 has got 72% on the screenshot protest, which is
• 12:18
  double the previous best score of 36%.
• 12:21
  And this test is only testing a model, whether it can look at a
• 12:24
  screenshot, understand the UI it sees and figure out where to
• 12:28
  click and take the right action quickly.
• 12:29
  And really interesting ways this has been used is, for example,
• 12:32
  giving it access to 3 1/2 hours of footage from a meeting and
• 12:37
  getting it to produce transcripts, speaker names, time
• 12:40
  stamps, key sections, action items, all those types of
• 12:45
  important parts, even where people disagreed and agreed, but
• 12:48
  doing so by watching the actual video and the actual footage as
• 12:51
  well. So the type of innovations is my
• 12:53
  unlock is the ability for agents to navigate software themselves.
• 12:57
  Again, this isn't an area that I've been particularly excited
• 13:00
  by within AI because I think AP is so just, you know,
• 13:03
  transferring information through the back end is much quicker and
• 13:07
  easier. But it wouldn't surprise me if
• 13:09
  there were really some interesting use cases emerging
• 13:12
  because of this over time. You know, the personal shopping
• 13:15
  assistance, that type of idea #8 is the ability for Gemini to use
• 13:20
  multiple valid approaches to figure out a solution to a
• 13:24
  problem. So often a language model can
• 13:27
  struggle when there's no single correct answer to a problem,
• 13:31
  which is most problems in our world because it's a very
• 13:33
  complex world. You know, ask GPT 5.1 for, you
• 13:37
  know, three different approaches to reduce our customer churn.
• 13:40
  You'd often get different variations of the same, same or
• 13:43
  similar approach. Ultimately, the model is pattern
• 13:45
  matching to common solutions rather than exploring very
• 13:49
  different strategies. Now again, think about much of
• 13:53
  these innovations not being suddenly made possible because
• 13:56
  you can make a model do these things, but doing so within a
• 13:59
  single prompt or just a couple of prompts.
• 14:01
  So the speed to this raw intelligence and the ease to
• 14:05
  access and utilize this raw intelligence in other models,
• 14:08
  this might take a lot of, you know, twisting its arm, making
• 14:11
  sure it's got the right information and context.
• 14:14
  Whereas what Gemini is claiming to do is make this 10 times
• 14:18
  easier. And a good example is Gemini
• 14:20
  scored 45.1% on the ARC AGI two test, which is designed to have
• 14:26
  multiple potential solutions. In this test, it shows you a few
• 14:29
  different examples of patterns and we'll ask the model to infer
• 14:32
  the rule, for example, with more often than not more than a
• 14:36
  single valid interpretation. PPT scored 17.6, Gemini 2.5
• 14:41
  scored 4.9%. And again, as I said, Gemini
• 14:45
  just scored, the Gemini 3 model just scored 45.1%.
• 14:49
  That's that's a massive leap. And I think Matt Schumer said
• 14:52
  this perfectly, which is he said previous models often had a
• 14:56
  certain spikiness. Their quality varied wildly
• 14:59
  depending on different tasks. You could get brilliant at 1
• 15:02
  task, but just OK results on another, whereas Gemini 3 feels
• 15:06
  just more consistent across the board and when a model is
• 15:10
  reliably good at most things instead of occasionally
• 15:14
  brilliant, that's where you build entire systems around it.
• 15:28
  So yeah, let's let's conclude for today's episode I mean, so
• 15:31
  what's actually changed this week?
• 15:33
  Well, Google has just released 2 models that have broken the
• 15:36
  Internet and they've broken thresholds that people have
• 15:41
  really struggled to break for a very long time.
• 15:44
  Nana Banana now can merge 14 images perfectly.
• 15:48
  It can generate text perfectly. Gemini 3 can apparently do a
• 15:52
  single task within a couple of prompts for over 24 hours.
• 15:56
  CSA to say Google is super competitive again.
• 16:00
  And as I said at the beginning, Sam Altman announced to open AI
• 16:04
  that he is worried, and I mean I would be too because Google is a
• 16:10
  huge Goliath as incredibly talented people and is now
• 16:13
  making huge leaps into the future and being a market leader
• 16:18
  provider. Anyway, as always, the most
• 16:19
  important thing I can now recommend is go try it for
• 16:22
  yourself. Go play, go have fun, see if it
• 16:25
  works in your day-to-day work. Anyway, that's it for today.
• 16:29
  I hope you enjoyed the episode and I'll see you next week.