Where Does Consumer AI Stand at the End of 2025?

0:00:06 For most of the year, less than 10% of Chat2BT users even visited another one of the big LLM providers.
0:00:10 When you open Gemini, it has a pop-up that says, we got a nano-banana.
0:00:11 Would you like to do something with it?
0:00:13 A little pane where you have to type something.
0:00:15 I don’t know what to do.
0:00:20 These are hot-up nuances that I think makes people actually take the first.
0:00:25 The models have gotten to the level of quality that you can build a real scalable app on top of them.
0:00:30 And so the hope is 2026 will be a huge year for consumer builders.
0:00:36 As 2025 comes to a close, consumer AI is starting to look very different than it did at the beginning of the year.
0:00:40 A small number of products now dominate everyday usage.
0:00:46 New multimodal models have gone viral, and the big labs have pushed harder than ever into consumer experiences.
0:00:53 To take stock of the year, the A16Z team, Anisha Charya, Olivia Moore, Justine Moore, and Brian Kim
0:00:56 break down what actually worked in 2025 and what didn’t.
0:00:59 They discuss which model launches and interfaces change user behavior,
0:01:03 why small product details matter more than raw model quality,
0:01:07 and whether the consumer AI market is trending toward win or take most.
0:01:12 The conversation also looks ahead to 2026, where there is still room for startups,
0:01:15 how templates and multimodality are reshaping creation,
0:01:19 and why this may finally be the moment when scalable consumer AI apps break out.
0:01:25 Today we’re talking about who won consumer AI in 2025.
0:01:29 This was arguably the year that we saw the big model providers,
0:01:34 OpenAI and Google, most out of everyone, make a major push of their own into consumer,
0:01:39 both in terms of new models they release, but also in terms of new products, features,
0:01:41 and interfaces that target the mainstream user.
0:01:45 You might wonder, why does it matter who is in the lead here?
0:01:51 There are some early signs that the general LLM assistance space might be trending towards winner-take-all,
0:01:53 or at least winner-take-most.
0:01:59 So only 9% of consumers are paying for more than one out of the group of ChatGPT,
0:02:01 Gemini, Claude, and Cursor.
0:02:08 And for most of the year, less than 10% of ChatGPT users even visited another one of the big LLM providers like Gemini.
0:02:17 If we had to call it now, ChatGPT is currently in the lead, by far, at 800 to 900 million weekly active users.
0:02:23 Gemini’s at an estimated 35% of their scale on web and about 40% on mobile.
0:02:25 And everyone else significantly trails this.
0:02:29 So Claude, Grok, Perplexity are all about 8% to 10% of the usage.
0:02:34 But, especially in the last three to six months, things are changing very quickly.
0:02:37 With the launch of new viral models like NanoBanana,
0:02:42 Gemini is now growing desktop users 155% year-over-year,
0:02:46 which is actually accelerating even as they reach more scale, which is pretty crazy to see.
0:02:50 And ChatGPT is only growing 23% year-over-year.
0:02:56 And we’re starting to see players like Anthropic almost specialize within consumer,
0:02:58 owning different verticals like the hyper-technical user.
0:03:03 So today, we’ve brought together the A16Z consumer team to recap what we saw this year
0:03:08 from the big model companies in consumer, and also to predict what might be ahead of us in 2026.
0:03:10 Cool. Well, thank you, Olivia.
0:03:11 It’s been a super fun year.
0:03:14 If we kind of wind the timeline back to last January,
0:03:17 maybe we should start with what we saw, launches, products,
0:03:18 what worked, what didn’t.
0:03:20 So, Justine, tell us what you saw this year.
0:03:22 OpenAI, Google, what are you paying attention to?
0:03:23 What have you changed your mind on?
0:03:27 Yeah, those two in particular had a ton of consumer launches, like Olivia mentioned.
0:03:31 From a model perspective, I would argue their most viral models this year,
0:03:34 at least among consumers, were in image and video.
0:03:39 So, for OpenAI, it was the Chack GPT 4.0 image, the Ghibli moment,
0:03:41 which is crazy that that was this year.
0:03:42 It seems like this is this year.
0:03:44 It feels like it was years ago.
0:03:47 And then Sora, obviously, Sora 2.
0:03:51 And then for Google, it’s VO, VO3, and VO3.1,
0:03:54 and then Nano Banana and Nano Banana Pro in image models,
0:03:57 which went insanely viral, probably comparable to,
0:04:00 if not beyond, the Ghibli moment for OpenAI.
0:04:03 I think in terms of the product layer,
0:04:09 what we saw was OpenAI tended to keep more things in the ChatGPT interface.
0:04:14 So, like, Pulse, group chats, shopping, research, tasks,
0:04:17 all of these features launched inside ChatGPT as the core.
0:04:21 The exception there is obviously Sora as a standalone video app,
0:04:24 whereas Google tended to launch more things as standalone products.
0:04:27 So, they did ship a lot through, like, Google AI Studio
0:04:30 and Google Labs and Gemini
0:04:34 and the plethora of Google Surfaces there are to launch a product.
0:04:36 But they would also ship things as standalone websites
0:04:38 that you could go to and visit,
0:04:40 which basically allowed for a more custom interface
0:04:42 for each type of product,
0:04:46 not just the kind of chat entry, chat exit, or image video exit.
0:04:48 Well, so, Justine, I have a question for you on that.
0:04:51 So, it felt like 18 months ago, we were talking about mid-journey,
0:04:55 and most of the multimodal models were defined by aesthetics and realism.
0:04:56 Is that still true?
0:04:57 What changed this year?
0:05:00 Yeah, I think there was definitely different styles still.
0:05:01 And I think mid-journey,
0:05:04 when you talk to people really deep in image and video,
0:05:07 it still kind of stands apart for this, like, aesthetic sensibility
0:05:10 that a lot of the models don’t have if you don’t know how to prompt for it.
0:05:12 But I would say, this year in particular,
0:05:14 we made a lot more strides on realism
0:05:18 and also on reasoning within both image and video.
0:05:22 Like, all of the little details that make an image or a video actually seem real.
0:05:25 For example, if you have a person walking and talking,
0:05:27 the people in the cars in the background,
0:05:30 if they’re on a street, should be moving in the correct direction,
0:05:33 like they shouldn’t be morphing and looking strange.
0:05:37 In image, we were able to have multiple input images and text
0:05:40 and sort of reason across all of those uploads
0:05:43 to create, like, a cohesive design or something like that,
0:05:45 which was not something we saw happening last year, for sure.
0:05:49 Yeah, I remember when we were excited about having
0:05:52 a letter show up correctly in images.
0:05:55 And now we have insane infographics.
0:05:55 Yes.
0:05:58 We can just put up an amazing YouTube video and say,
0:05:59 give me an image that explains this.
0:05:59 Yeah.
0:06:01 That’s incredibly different.
0:06:03 Nano Banana Pro can even generate, like, market maps.
0:06:05 Like, I would tell it, generate a market map in the space.
0:06:06 You generated a market map. It’s incredible.
0:06:10 And it either has or will go do the web research within the image model,
0:06:12 which is crazy, to get the correct list of companies
0:06:14 and then pull their correct logos, which is insane.
0:06:15 I know.
0:06:18 There’s one benchmark left that the reasoning image models have not cracked.
0:06:21 I tested GPT Image 1.5 yesterday.
0:06:25 They sometimes struggle with both reasoning and multi-step reasoning.
0:06:29 So what I’ve been testing is you upload a picture of a monopoly board
0:06:31 and you say, remove the names of all the properties
0:06:35 and replace them with names of AI labs and startups.
0:06:39 And GPT Image 1.5 is actually the closest,
0:06:44 but it’s very hard for them to do all of those steps.
0:06:46 Remove it, come up with the new names,
0:06:47 put all of the new names in the correct places,
0:06:51 make sure there’s not overlaps or one thing you mentioned three times
0:06:53 and another big player you never mentioned.
0:06:56 So there’s still some room to go on the image evals.
0:06:57 And it’s interesting that,
0:06:59 especially from the image model from ChatGPT,
0:07:02 where you can actually see perseverance of like,
0:07:06 it carries a character over into multiple image generation,
0:07:07 the same style.
0:07:07 Yeah.
0:07:09 And I thought that was like,
0:07:10 oh, like this is actually very interesting.
0:07:11 We’re storyboarding.
0:07:11 Totally.
0:07:13 makes you want to generate more.
0:07:14 Yeah.
0:07:15 You know, for me,
0:07:17 it felt like the most underhyped aspect of Nano Banana
0:07:19 was the integration with Search,
0:07:21 because it feels like there’s realism,
0:07:24 which is physics and sort of other things
0:07:25 that feel like we’re on Candy Valley.
0:07:27 There is reasoning,
0:07:29 which is apply modifications
0:07:31 that are adherent to what the user asked for.
0:07:33 But then there’s also sort of accuracy.
0:07:36 And for me, a good example of this is product photography.
0:07:38 If you say, hey, generate a photo of this album cover
0:07:42 or a historically accurate photo of this moment in time,
0:07:43 you have to actually have the search integration.
0:07:45 And that was sort of non-intuitive,
0:07:46 but it’s actually very useful.
0:07:47 Totally.
0:07:47 Yeah.
0:07:49 It’s kind of like the VO3 moment
0:07:52 when I don’t think it was intuitive to people
0:07:53 that video would be cracked necessarily
0:07:56 by bringing audio together with video in the same place.
0:07:57 And that ended up being the thing
0:07:59 that made AI video go viral.
0:08:00 Yeah.
0:08:01 Like since VO3,
0:08:03 and now Sora maybe dominates,
0:08:04 but like since VO3,
0:08:07 my social feeds have been like full of
0:08:08 really realistic looking.
0:08:09 I counted.
0:08:11 About one-fifth of my feeds are AI generated.
0:08:12 Amazing.
0:08:12 Wow.
0:08:13 What do you guys do?
0:08:15 There’s so many launches this year
0:08:16 and many of them went well,
0:08:17 like VO and Nano.
0:08:18 What do you think is under-hyped
0:08:19 or products that you think
0:08:20 didn’t get enough attention?
0:08:21 Brian?
0:08:23 It’s a good question.
0:08:26 I think under-hyped pulse of the world
0:08:27 is probably still under-hyped.
0:08:29 And we’re talking about
0:08:31 OpenAI, Google,
0:08:33 which to me falls under productivity category.
0:08:35 So if you actually think about,
0:08:36 if you go to App Store today,
0:08:39 top five out of top 10 productivity apps
0:08:40 are all Google.
0:08:41 It’s insane.
0:08:42 And ChatGP is number one.
0:08:44 So we’re talking about a productivity category
0:08:46 where it helps you do things.
0:08:48 And I feel like a lot of people
0:08:50 are trying this from a different angle.
0:08:52 Like how do I actually ingest your data
0:08:53 or your schedule,
0:08:54 your email,
0:08:55 to make it more helpful
0:08:56 and give more proactive
0:08:58 and notification to you?
0:09:00 I think a lot of people are working on it.
0:09:01 Given the frequency
0:09:03 of people using ChatGPT,
0:09:04 which I think is what,
0:09:06 25 times a week,
0:09:07 pretty good,
0:09:08 pretty good,
0:09:08 three to four times a day,
0:09:11 it feels like it’s a really good position
0:09:14 to actually give you proactive nudges
0:09:15 and summary
0:09:17 and help your life in general.
0:09:20 So I feel like the Everything app
0:09:21 was always this myth
0:09:22 in the Western world.
0:09:24 I think OpenAI is trying to move
0:09:24 in that direction
0:09:26 where it’s ingesting enough,
0:09:27 people are going there enough
0:09:29 to start giving really useful,
0:09:30 proactive nudges.
0:09:32 and I think that’s a space
0:09:33 that I’m excited about.
0:09:34 It’s interesting.
0:09:35 But are you a DAU?
0:09:37 I am not a DAU.
0:09:37 A Pulse?
0:09:39 Not a Pulse.
0:09:41 Similarly, I tried Pulse for a while
0:09:43 and have kind of largely turned off of it.
0:09:44 But I would agree with you
0:09:46 that I feel like Pulse
0:09:47 and a couple other examples
0:09:49 that OpenAI launched this year
0:09:50 are kind of new primitives
0:09:52 or ideas that feel underhyped.
0:09:53 But because the execution
0:09:55 is a little off.
0:09:55 I think it’s execution.
0:09:57 The usage is off.
0:09:58 Another example that I would give,
0:09:59 which is similarly like
0:10:01 personal contacts,
0:10:02 would be their connectors.
0:10:03 So now you can,
0:10:05 and you can do this on cloud as well.
0:10:06 You can connect your calendar,
0:10:07 your email, your documents.
0:10:08 And so hypothetically,
0:10:10 you could say to ChatGPT,
0:10:11 you know,
0:10:12 read all of my memos
0:10:13 over the past six months
0:10:14 and like summarize
0:10:15 what’s most interesting,
0:10:16 least interesting.
0:10:17 I think when that works,
0:10:18 it’s really exciting.
0:10:19 I have found it to be
0:10:21 a little bit unreliable so far,
0:10:22 but I think as the models get better,
0:10:24 they have a real chance
0:10:25 to kind of own
0:10:26 the prosumer workspace
0:10:27 if they get that right.
0:10:28 Prosumer is a perfect category
0:10:29 because we talk about it sometimes,
0:10:31 but 99% of people
0:10:33 don’t run their lives on calendar.
0:10:33 Yeah.
0:10:34 We do.
0:10:34 Right.
0:10:35 So that’s what I’m thinking about,
0:10:38 the actual average frequency
0:10:39 of using ChatGPT.
0:10:40 And look,
0:10:42 if it’s 24 times a week,
0:10:44 that’s a pretty good place to start.
0:10:44 Yeah.
0:10:45 Olivia,
0:10:46 I feel like you are
0:10:47 the ultimate power user.
0:10:49 What are you still using?
0:10:50 What’s your stack?
0:10:51 It’s a great question.
0:10:53 From all of the larger model companies,
0:10:54 actually,
0:10:55 I would have to say the thing
0:10:56 that I’m still using the most
0:10:57 and was maybe the most impressed
0:10:58 by this year
0:11:00 was the Perplexity Comet browser.
0:11:01 And I don’t,
0:11:04 and was not using Perplexity
0:11:06 as my core general LLM assistant.
0:11:08 I use ChatGPT and Claude much more,
0:11:11 but I think they really executed on it
0:11:12 in a first class way
0:11:14 in terms of both the agentic model
0:11:15 within the browser,
0:11:17 but also perhaps more importantly,
0:11:18 all of the workflows
0:11:19 that you can set up
0:11:21 that allow you to basically run
0:11:23 the same task over and over,
0:11:24 either at a preset time
0:11:25 or when you trigger it
0:11:26 on a certain web page.
0:11:26 So that to me
0:11:27 was a really exciting launch.
0:11:28 And if you look at the data,
0:11:30 like the spike at launch
0:11:31 and the sustained traffic
0:11:32 for Comet
0:11:33 was actually much higher
0:11:35 than for ChatGPT’s
0:11:36 own browser launch,
0:11:36 Atlas,
0:11:38 which is kind of crazy
0:11:39 given how much more distribution
0:11:40 ChatGPT has
0:11:41 than Perplexity.
0:11:42 But I think
0:11:43 they also launched
0:11:45 an email assistant this year,
0:11:45 Perplexity did,
0:11:46 and they made
0:11:47 a couple acquisitions
0:11:48 of really strong
0:11:49 agentic startups.
0:11:51 And so what I would love
0:11:52 to see from them next year
0:11:53 is like more of these
0:11:55 dedicated prosumer interfaces.
0:11:56 I feel like that would be
0:11:57 an awesome direction
0:11:57 for them to kind of
0:11:58 double down in.
0:11:59 They do feel like
0:12:00 the startup that has
0:12:01 the biggest breadth
0:12:02 of ambition,
0:12:02 you know,
0:12:03 alongside the labs
0:12:04 and sort of big tech,
0:12:05 like it’s very,
0:12:05 very impressive
0:12:06 just the number of things
0:12:07 they’ve shipped this year.
0:12:08 Yes, definitely.
0:12:10 What, you know,
0:12:11 one thing I wanted to ask
0:12:11 you, Justine,
0:12:12 was sort of Gemini
0:12:13 feels like it’s having
0:12:14 a real moment
0:12:15 because of all the
0:12:16 image and video models.
0:12:17 Do you think it can
0:12:18 overtake ChatGPT?
0:12:19 Is there truly
0:12:20 that much demand
0:12:21 for these types of models?
0:12:23 I think, yeah.
0:12:23 So what we,
0:12:24 what I’ve seen basically
0:12:27 is there is always
0:12:29 nearly infinite demand
0:12:30 for like the best
0:12:31 in class image
0:12:32 or video model
0:12:34 because then you have
0:12:35 a mix of tons
0:12:36 of different people
0:12:36 seeing it
0:12:37 and wanting to use it.
0:12:37 You have,
0:12:39 like if you’re using
0:12:39 it professionally,
0:12:40 if you’re marketing
0:12:41 or in entertainment
0:12:42 or storyboarding
0:12:42 or whatever,
0:12:44 you always want
0:12:45 to be using
0:12:45 what’s at the forefront
0:12:46 of the field
0:12:47 and so you’re totally
0:12:47 fine to go somewhere
0:12:49 other than ChatGPT
0:12:49 and Sora
0:12:51 to get access to VO.
0:12:52 Even if you’re
0:12:53 an everyday consumer,
0:12:55 so many new viral trends
0:12:55 are created
0:12:57 around new capabilities
0:12:58 of the best in class
0:12:59 image and video models
0:13:01 and so that ends up
0:13:02 driving users
0:13:03 into different products
0:13:05 that they may have
0:13:05 never tried before.
0:13:06 Like you might be
0:13:08 downloading the Gemini app
0:13:10 or accidentally ending up
0:13:11 on Google AI Studio,
0:13:12 which I know they’re trying
0:13:12 to make be more
0:13:13 for developers
0:13:15 to use Nano Banana Pro,
0:13:16 which a lot of users
0:13:17 I think experienced
0:13:19 in the past couple of months.
0:13:19 Yeah.
0:13:21 The interesting thing
0:13:22 about Gemini to me
0:13:23 is like hypothetically
0:13:24 they benefit from
0:13:25 the massive Google
0:13:26 distribution advantage.
0:13:27 Like if you look
0:13:27 at Android,
0:13:29 Gemini is at like
0:13:31 50% of ChatGPT’s
0:13:32 scale on mobile,
0:13:33 whereas on iOS
0:13:34 it’s like 17%.
0:13:35 So like clearly
0:13:36 something is working there.
0:13:37 They launched a little
0:13:37 Gemini widget
0:13:39 within Chrome recently
0:13:40 that encourages you
0:13:41 to use it.
0:13:41 They’re launching it
0:13:42 within Google Docs
0:13:43 and Gmail and other things.
0:13:44 Yeah.
0:13:46 But I think that most
0:13:47 the average person
0:13:48 is still just using
0:13:49 one AI product
0:13:51 and ChatGPT is like
0:13:52 the Kleenex of AI.
0:13:53 Like it is the brand
0:13:53 that has become
0:13:54 the noun.
0:13:55 Exactly.
0:13:55 Yes, yes, yes.
0:13:56 And so I think
0:13:57 that Gemini still has
0:13:58 a pretty big hurdle
0:13:59 to overcome
0:14:01 just in terms of that.
0:14:01 Yeah.
0:14:03 But if they keep
0:14:03 doing what they’re
0:14:04 doing on these
0:14:06 amazing viral
0:14:07 consumer creative
0:14:08 tool launches
0:14:09 and model launches
0:14:10 like they could
0:14:11 get there next year.
0:14:12 I’m thinking about this.
0:14:13 It’s really interesting
0:14:15 when you look at
0:14:16 Gemini,
0:14:17 which is everywhere.
0:14:18 Yeah.
0:14:18 Yeah.
0:14:19 But yet nowhere
0:14:20 to some extent, right?
0:14:21 You don’t like,
0:14:22 you know,
0:14:22 when you look at
0:14:23 the actual usage,
0:14:24 people still think
0:14:24 of the Kleenex.
0:14:25 Yep.
0:14:26 And they go to
0:14:27 ChatGPT.
0:14:28 But the interesting
0:14:29 thing also is on
0:14:30 the product sensibility.
0:14:32 So this morning
0:14:33 I had like two panes open,
0:14:35 OpenAI’s image model
0:14:37 and Google’s Gemini
0:14:38 and basically use
0:14:40 an image functionality.
0:14:41 When you open Gemini,
0:14:43 it’s a blank screen.
0:14:45 It has a pop-up
0:14:45 that says,
0:14:46 we got a nano banana.
0:14:47 Would you like
0:14:48 to do something with it?
0:14:50 And it’s a little pane
0:14:51 where you have to
0:14:51 type something.
0:14:52 Yeah.
0:14:53 I don’t know what to do.
0:14:53 Yeah.
0:14:55 ChatGPT,
0:14:55 you go in
0:14:56 and it has a very
0:14:58 TikTok-like style
0:14:58 of like,
0:14:59 here’s a trending
0:15:00 themes that you
0:15:01 might want to generate
0:15:02 and you click on
0:15:03 I want a sketch pen
0:15:04 or whatever
0:15:05 and then just like
0:15:06 use one other picture
0:15:06 and it creates
0:15:07 something amazing.
0:15:08 And then it says,
0:15:08 would you like
0:15:09 a holiday card?
0:15:09 Would you like
0:15:10 a blah, blah, blah, blah, blah.
0:15:12 These are product nuances
0:15:13 that I think
0:15:14 makes people actually
0:15:15 take the first step
0:15:17 to generate it.
0:15:18 And then once you have it,
0:15:19 you have character consistency.
0:15:19 Yeah.
0:15:20 So you keep going.
0:15:21 Right.
0:15:22 So that’s interesting
0:15:22 in that I think
0:15:23 OpenAI
0:15:24 and ChatGPT
0:15:25 has proven
0:15:25 that there is
0:15:27 deeper product sensibility.
0:15:28 Yeah.
0:15:29 But then I,
0:15:30 this is a funny thing,
0:15:31 maybe a little
0:15:32 non-cosher thing to say,
0:15:33 but you know,
0:15:35 I worked at Snap.
0:15:36 So when you look
0:15:37 at Meta versus Snap,
0:15:38 famously,
0:15:39 Evan Spiegel
0:15:40 was chief product
0:15:40 officer of Meta.
0:15:41 Yeah.
0:15:41 Yeah.
0:15:44 I wonder if there’s
0:15:44 a world where
0:15:46 the ChatGPT team
0:15:47 that innovates
0:15:48 on the product front
0:15:49 again and again,
0:15:50 Google with distribution,
0:15:51 looks at them like,
0:15:52 that’s cool.
0:15:53 Let’s just,
0:15:54 let’s just integrate it
0:15:55 and keep going
0:15:56 and actually play that game.
0:15:57 The interesting thing there
0:15:59 is that images pane
0:16:00 just launched yesterday
0:16:01 when we’re filming this.
0:16:02 In ChatGPT.
0:16:03 In ChatGPT.
0:16:03 Yeah.
0:16:04 Brand new.
0:16:04 Yeah.
0:16:04 And it took them,
0:16:06 like they had image models
0:16:07 for years
0:16:08 and it took them that long
0:16:09 to come up with a separate
0:16:11 relatively basic interface
0:16:12 for generating images.
0:16:14 I would almost argue
0:16:16 the application layer companies
0:16:17 like the CRIAs,
0:16:18 the Hedras,
0:16:19 the Higgs fields of the world
0:16:22 popularized that template format
0:16:23 and did it first
0:16:23 and did it better.
0:16:24 I agree.
0:16:26 And they are ChatGPT’s
0:16:27 product people
0:16:28 and then maybe
0:16:28 the ChatGPT product people.
0:16:29 So it’s a supply chain
0:16:30 of product ideas.
0:16:30 Exactly.
0:16:32 Always.
0:16:33 Well, maybe going
0:16:34 in a slightly different direction,
0:16:34 BK,
0:16:35 I’m very curious for your take
0:16:37 on OpenAI’s social features
0:16:38 because it does feel like
0:16:39 that’s something
0:16:40 that you really have to get
0:16:41 product execution right on
0:16:42 but also network design.
0:16:43 You know,
0:16:43 there’s some efforts
0:16:44 around Sora too.
0:16:45 We should talk about that.
0:16:46 There’s also group chats
0:16:47 within ChatGPT.
0:16:49 You’re our sort of social guy
0:16:50 or have been historically
0:16:51 bullish, bearish.
0:16:51 Where’s your head at?
0:16:53 Bearish for now.
0:16:54 Okay.
0:16:56 And the reason to me
0:16:57 is twofold.
0:16:58 Historically,
0:16:59 we look at sort of,
0:17:00 it’s funny,
0:17:01 I look at products
0:17:03 based on what I call
0:17:04 inception theory.
0:17:05 You go like three
0:17:05 to four layers down
0:17:06 to figure out
0:17:07 what the one liner is
0:17:08 which is like
0:17:09 I want my dad to love me.
0:17:10 And so,
0:17:10 you know,
0:17:11 when they think
0:17:12 about products,
0:17:12 Is that for you
0:17:13 or for the world?
0:17:13 That’s for me
0:17:15 as well as for a lot of people.
0:17:15 Okay.
0:17:16 Yes, yes.
0:17:17 And so,
0:17:18 I look at some of the,
0:17:19 you know,
0:17:20 products like,
0:17:21 like ChatGPT.
0:17:22 Ultimately,
0:17:23 when you peel the onion
0:17:23 five times,
0:17:24 I think essentially
0:17:25 is help me be better.
0:17:26 Like,
0:17:27 help me get that information.
0:17:28 Help me be more productive.
0:17:29 Help me be more efficient.
0:17:31 And then when I think
0:17:32 about social features,
0:17:33 meta,
0:17:33 Instagram,
0:17:34 what have you,
0:17:35 or even TikTok,
0:17:37 the two layers
0:17:37 of information
0:17:38 or the,
0:17:39 you know,
0:17:40 emotion that it’s trying
0:17:40 to address to me
0:17:41 is for TikTok,
0:17:43 entertain me.
0:17:44 I want my clown
0:17:45 to entertain me.
0:17:45 Yeah.
0:17:46 And then the other layer
0:17:47 is I’m lonely,
0:17:48 I want to be seen,
0:17:49 I want to connect
0:17:49 with people.
0:17:51 And to me,
0:17:52 these are pretty
0:17:54 two different parallels
0:17:55 in the product direction.
0:17:57 And OpenAI’s product
0:17:58 is incredible.
0:17:59 It’s magic.
0:18:00 It’s amazing.
0:18:01 But it’s ultimately
0:18:02 a see me
0:18:03 or help me category,
0:18:04 which essentially
0:18:05 is why
0:18:06 it’s the number one
0:18:07 in productivity category.
0:18:07 Yeah.
0:18:09 Now we’re trying
0:18:10 to take this
0:18:11 and shove it
0:18:11 in people’s life
0:18:12 and say,
0:18:12 guys,
0:18:13 connect,
0:18:14 connect better
0:18:15 and like actually
0:18:17 feel like you’re being seen.
0:18:18 And even the group chat function,
0:18:19 which I love,
0:18:20 it’ll be so good
0:18:21 to plan a trip
0:18:22 and like actually
0:18:23 have that common pain.
0:18:25 But I think
0:18:26 it still stops
0:18:28 at probably
0:18:28 end count
0:18:30 of two to three people
0:18:31 planning something
0:18:32 in a help me way
0:18:32 versus,
0:18:33 oh,
0:18:34 I feel like
0:18:35 I understand
0:18:36 a niche so much better
0:18:38 because I’ve sort of
0:18:38 done that.
0:18:40 So largely over time,
0:18:42 I think that’s the reason
0:18:43 of that division.
0:18:44 But that is not to say
0:18:45 you can build
0:18:46 a separate product
0:18:47 that completely
0:18:48 sort of addresses that.
0:18:49 I think Sora,
0:18:51 so we talked about group chat,
0:18:52 Sora 2 was the other big
0:18:53 I think social push
0:18:54 this year
0:18:55 from all the consumer AI giants.
0:18:56 Which was basically
0:18:58 like a TikTok feed
0:18:59 but all AI generated
0:19:00 video and you can make
0:19:01 cameos of your friends.
0:19:03 The cameos was a very good bet.
0:19:03 Yeah.
0:19:04 It was a strong bet.
0:19:04 Yeah.
0:19:06 But I think what we’ve seen
0:19:08 is like in the retention data
0:19:09 and how we’re seeing it used
0:19:11 is it was massively successful
0:19:12 as a creator tool.
0:19:13 Like now,
0:19:15 my feed is probably
0:19:16 two-thirds AI slop
0:19:16 if not more.
0:19:19 And over 50% of it
0:19:20 is now Sora.
0:19:20 Whereas before
0:19:21 it was like all VO
0:19:22 and some cling.
0:19:24 but it has not been
0:19:25 as successful
0:19:26 as like a social app.
0:19:27 Consumption.
0:19:28 Yeah.
0:19:28 People are like
0:19:30 a small number of creators
0:19:31 are creating a ton of content
0:19:33 and then bringing it out
0:19:34 to like TikTok,
0:19:34 Instagram,
0:19:35 X,
0:19:35 Reddit
0:19:36 where it’s going
0:19:37 massively viral.
0:19:38 But it doesn’t seem
0:19:39 like there’s a
0:19:40 as much
0:19:42 consumption happening
0:19:42 in the app.
0:19:43 Yeah.
0:19:44 As much remixing,
0:19:44 as much commenting,
0:19:46 especially as there was initially.
0:19:47 You know,
0:19:47 in a funny way,
0:19:48 the way I think about it
0:19:50 is like Sora’s competition
0:19:51 or analogy
0:19:53 isn’t actually TikTok.
0:19:54 It’s actually Capcom.
0:19:57 It’s like a funny way.
0:19:57 It’s almost like
0:19:58 a creative tool.
0:19:58 Yes.
0:19:59 Interesting.
0:19:59 Yes.
0:20:00 Olivia,
0:20:00 what’s your thing?
0:20:00 Well,
0:20:01 I was going to say
0:20:02 like I think it goes back
0:20:03 to your earlier point
0:20:04 which is like
0:20:06 the kind of motion
0:20:07 that drives social apps
0:20:09 is both these like
0:20:10 positive and negative
0:20:11 feelings of like
0:20:12 oh,
0:20:13 I’m publishing this thing
0:20:14 of myself
0:20:15 that’s kind of sensitive
0:20:16 or that I want people
0:20:17 to think it’s this
0:20:17 or that
0:20:18 or this other thing
0:20:19 and so that’s kind of
0:20:21 what drives participation
0:20:21 on the app.
0:20:22 Yeah.
0:20:22 The status game.
0:20:23 Yeah.
0:20:24 A little bit of the status game.
0:20:25 Exactly the status game
0:20:27 and when it’s AI generated content
0:20:28 and people know
0:20:29 it’s not real
0:20:31 like a real representation
0:20:32 of you as a human being
0:20:33 the status game
0:20:34 is lost a little bit.
0:20:35 Absolutely lost.
0:20:35 Yeah.
0:20:36 I think the status game
0:20:37 comes then with
0:20:39 can you prompt
0:20:40 something very cool?
0:20:40 Yeah.
0:20:41 But that’s a different type
0:20:42 of product
0:20:43 and that’s why
0:20:44 I think it goes viral
0:20:45 on like Twitter
0:20:45 and all these other
0:20:46 existing platforms.
0:20:47 I mean,
0:20:48 my sort of counterpoint
0:20:49 or bull case
0:20:49 for Sora 2
0:20:50 is I actually think
0:20:51 the status game
0:20:51 was about humor
0:20:53 more than anything else
0:20:54 and humor is the intersection
0:20:56 of knowing how to prompt
0:20:56 and sort of being
0:20:57 culturally aware.
0:20:58 Yeah.
0:20:59 So I think that if they
0:21:00 iterated on that
0:21:01 that’s like a direction
0:21:02 that nobody has captured before.
0:21:02 Yeah.
0:21:03 Yes,
0:21:04 but if you can export
0:21:05 those videos
0:21:06 isn’t it true
0:21:08 that like TikTok
0:21:09 with Sora videos on it
0:21:11 is strictly better than Sora?
0:21:12 We talked about it so much
0:21:13 where like
0:21:13 the ultimate
0:21:15 social product
0:21:16 is where consumption
0:21:17 and creation
0:21:18 both live together
0:21:19 and that the output
0:21:20 of it is
0:21:21 not native
0:21:22 to other platforms
0:21:23 like TikTok
0:21:24 like YouTube Shorts.
0:21:27 So what do folks
0:21:27 think of the challengers?
0:21:28 You know,
0:21:29 we’re talking about
0:21:29 Sora 2
0:21:30 who I mean
0:21:31 Meta
0:21:32 it’s crazy
0:21:32 to talk about
0:21:33 Meta as a challenger
0:21:34 I guess in this context
0:21:34 they are
0:21:35 but I think
0:21:36 Claude,
0:21:36 Perplexity,
0:21:37 Grok
0:21:38 are the more obvious
0:21:39 names for challengers.
0:21:39 Olivia,
0:21:40 what’s your take?
0:21:41 I love Claude.
0:21:41 I talk to Claude
0:21:42 all the time.
0:21:43 Claude is somewhat
0:21:44 replaced ChachiBT
0:21:45 for me as my
0:21:46 general LLM.
0:21:47 I think Claude
0:21:48 is opinionated
0:21:49 in an interesting way.
0:21:51 I also love Claude
0:21:52 because I’m willing
0:21:53 to invest time
0:21:53 into building out
0:21:54 AI workflows.
0:21:55 I think Claude
0:21:56 actually launched
0:21:56 a lot of really
0:21:57 powerful things this year
0:21:58 around like artifacts
0:21:59 and skills
0:22:00 where you can
0:22:01 essentially set up
0:22:03 tasks or workflows
0:22:05 to run over time.
0:22:06 I do think the reason
0:22:07 it hasn’t hit
0:22:08 the mainstream yet
0:22:08 is even the way
0:22:09 they built those things
0:22:10 is geared towards
0:22:12 a technical user
0:22:13 or an engineer.
0:22:14 It’s
0:22:15 I think they tried
0:22:16 to make skills
0:22:17 as easy as they could
0:22:18 to create
0:22:19 and it still
0:22:20 was not anywhere
0:22:21 near easy enough
0:22:22 for the mainstream consumer.
0:22:23 Another example
0:22:24 would be
0:22:25 they were actually
0:22:25 the first
0:22:27 of the big players
0:22:28 to kind of launch
0:22:29 file creation
0:22:30 slide deck creation
0:22:31 editing
0:22:33 and they branded it
0:22:33 as like
0:22:34 file generation
0:22:35 and analysis
0:22:36 or something
0:22:36 and it was like
0:22:37 a toggle feature
0:22:39 within a setting bar
0:22:40 of a setting bar
0:22:41 or something
0:22:41 so like
0:22:42 very few people
0:22:42 used it
0:22:43 and yet to me
0:22:44 it’s still
0:22:45 the best product
0:22:46 across all of them
0:22:47 at doing that
0:22:48 kind of complex work.
0:22:50 So I love Claude
0:22:50 but I think
0:22:51 if they want
0:22:52 to be a true
0:22:53 mainstream consumer
0:22:54 product
0:22:55 they need
0:22:56 to
0:22:57 dumb it down
0:22:58 even more
0:22:59 in terms
0:22:59 of accessibility.
0:23:00 There was
0:23:00 that survey
0:23:00 you found
0:23:01 recently
0:23:01 of U.S.
0:23:02 teens.
0:23:03 Yeah there’s
0:23:04 I think it was
0:23:05 three times
0:23:05 more U.S.
0:23:06 teens have
0:23:06 ever used
0:23:07 character AI
0:23:07 than have
0:23:08 used Claude.
0:23:08 Yeah.
0:23:09 So I think
0:23:09 that shows
0:23:10 that like
0:23:10 Claude is
0:23:11 a pretty broad
0:23:12 thing.
0:23:13 Claude is beloved
0:23:13 amongst
0:23:14 tech people
0:23:15 but outside
0:23:16 of tech people
0:23:16 I think
0:23:17 they are
0:23:18 maybe struggling
0:23:18 to pick up
0:23:19 relevance.
0:23:20 It is
0:23:20 interesting
0:23:20 though
0:23:20 like
0:23:21 if you
0:23:21 look at
0:23:21 the sort
0:23:21 of
0:23:22 aesthetics
0:23:22 the product
0:23:23 design
0:23:24 the craft
0:23:24 like three
0:23:25 things that
0:23:26 Anthropic did
0:23:27 were MCP
0:23:29 skills and
0:23:30 command line
0:23:30 interface
0:23:31 Claude code
0:23:32 like those
0:23:32 are three
0:23:33 surprising bets
0:23:35 especially Claude code
0:23:35 I would have
0:23:36 said command line
0:23:37 interface really
0:23:37 like is this
0:23:38 the way that
0:23:38 people want
0:23:38 to interact
0:23:39 I thought
0:23:39 you were
0:23:39 going to talk
0:23:40 about taking
0:23:41 over airmail
0:23:41 and the thinking
0:23:42 cap
0:23:43 yeah that too
0:23:44 they’re a consumer
0:23:46 so three
0:23:46 things like
0:23:47 where’s the
0:23:47 thinking cap
0:23:49 but it’s
0:23:49 sort of
0:23:50 very high
0:23:50 minded
0:23:51 design
0:23:51 it’s
0:23:52 sort of
0:23:52 like
0:23:52 versus
0:23:52 mass
0:23:53 market
0:23:54 or maybe
0:23:54 that’s
0:23:55 apologetic
0:23:55 on their
0:23:55 behalf
0:23:56 but I
0:23:56 think
0:23:56 it is
0:23:57 that
0:23:57 it’s
0:23:58 opinionated
0:23:58 and it’s
0:23:58 great
0:24:00 I do need
0:24:00 to hear
0:24:01 Justine’s
0:24:01 take on
0:24:01 both
0:24:01 Meta
0:24:02 and
0:24:02 Grok
0:24:02 as I
0:24:03 feel like
0:24:03 they both
0:24:03 had
0:24:04 fascinating
0:24:04 years
0:24:05 in different
0:24:05 ways
0:24:06 so
0:24:07 Meta
0:24:08 hired
0:24:09 all those
0:24:09 researchers
0:24:10 I think
0:24:10 their
0:24:10 strongest
0:24:10 models
0:24:11 are
0:24:11 actually
0:24:12 not
0:24:12 consumer
0:24:12 facing
0:24:13 models
0:24:13 it’s
0:24:13 their
0:24:14 SAM
0:24:15 3
0:24:15 series
0:24:15 so
0:24:16 like
0:24:16 segment
0:24:16 anything
0:24:17 for
0:24:17 video
0:24:18 for
0:24:18 image
0:24:19 and
0:24:19 for
0:24:19 audio
0:24:20 and
0:24:20 basically
0:24:20 like
0:24:21 for
0:24:21 video
0:24:22 for example
0:24:22 you can
0:24:23 upload
0:24:23 a video
0:24:23 and you
0:24:24 can describe
0:24:24 a natural
0:24:25 language
0:24:25 like
0:24:26 find
0:24:27 the kid
0:24:27 in the
0:24:27 red
0:24:28 t-shirt
0:24:29 and it
0:24:29 will find
0:24:29 and track
0:24:30 that person
0:24:31 across
0:24:32 every
0:24:32 the entire
0:24:33 video
0:24:33 even if
0:24:33 they’re
0:24:34 coming in
0:24:34 and out
0:24:34 of the
0:24:34 frame
0:24:35 it will
0:24:35 let you
0:24:35 apply
0:24:36 effects
0:24:36 like
0:24:36 blurring
0:24:37 them out
0:24:37 or removing
0:24:38 them
0:24:38 or whatever
0:24:39 and
0:24:39 you
0:24:39 can
0:24:39 imagine
0:24:39 a
0:24:40 similar
0:24:40 thing
0:24:40 with
0:24:41 audio
0:24:42 with
0:24:42 different
0:24:43 stems
0:24:43 and
0:24:43 then
0:24:44 with
0:24:44 image
0:24:44 with
0:24:45 different
0:24:45 objects
0:24:45 in an
0:24:46 image
0:24:47 I
0:24:47 think
0:24:47 we’re
0:24:47 going
0:24:47 to
0:24:47 see
0:24:48 next
0:24:48 year
0:24:48 hopefully
0:24:49 some
0:24:49 incredible
0:24:50 consumer
0:24:50 products
0:24:50 built
0:24:51 on top
0:24:51 of
0:24:51 those
0:24:52 models
0:24:53 but
0:24:53 today
0:24:54 they’re
0:24:54 more
0:24:54 of a
0:24:54 playground
0:24:55 for
0:24:55 developers
0:24:56 than
0:24:56 they
0:24:56 are
0:24:56 a
0:24:57 consumer
0:24:57 which
0:24:57 is
0:24:57 surprising
0:24:58 given
0:24:58 just
0:24:59 like
0:24:59 the
0:24:59 DNA
0:24:59 of
0:24:59 the
0:25:00 company
0:25:00 yeah
0:25:01 so
0:25:01 the
0:25:01 one
0:25:02 good
0:25:02 consumer
0:25:02 feature
0:25:03 I
0:25:03 think
0:25:03 they’ve
0:25:03 launched
0:25:03 this
0:25:04 year
0:25:04 with
0:25:04 AI
0:25:05 is
0:25:05 the
0:25:05 Instagram
0:25:06 AI
0:25:06 translations
0:25:08 where
0:25:10 when you’re uploading a reel now
0:25:12 you can opt in to enable translations
0:25:13 and it will
0:25:14 clone your voice
0:25:16 translate it into five different languages
0:25:19 apply the translation with your voice
0:25:20 so you know
0:25:21 and then re-dub with
0:25:22 the lip sync
0:25:23 wow
0:25:25 and so it basically makes it seem like you’re a native speaker
0:25:26 in whatever language
0:25:29 so I would love to see more of that stuff come to
0:25:31 to the meta products
0:25:33 Grok I think has had
0:25:35 so Grok had a crazy year
0:25:36 with like the companions
0:25:37 yes
0:25:38 with all of the LLM progress
0:25:39 and the coding progress
0:25:42 I think their image and video progress
0:25:44 is probably the steepest slope
0:25:46 I’ve seen of any of the companies
0:25:47 like
0:25:50 it was probably like six months ago
0:25:52 they didn’t even like have image and video models
0:25:53 and
0:25:55 they’re shipping so fast
0:25:56 to launch new features
0:25:58 like it was initially just image to video
0:26:00 they added text to video
0:26:01 they added audio
0:26:02 then they added lip sync with speech
0:26:04 then they added 15 second videos
0:26:05 like
0:26:07 they’re just not slowing down the speed of progress
0:26:08 and
0:26:10 Elon has made a bunch of statements about like
0:26:14 wanting more interactive video game type content out of Grok
0:26:17 and wanting movies out of Grok by the end of next year
0:26:17 so
0:26:20 let’s hope it continues to go at that pace
0:26:22 do you feel like it’s a pinster movement
0:26:23 where like on one hand
0:26:24 there’s like a very
0:26:27 infrastructural model layer of like
0:26:27 let’s get to the
0:26:30 let’s top the LL Marina charts
0:26:31 and then the other one is like
0:26:33 let’s go Annie
0:26:35 I think
0:26:37 it’s like a little bit of like a bifurcated move
0:26:38 right
0:26:39 like the entertainment and the like
0:26:40 absolutely
0:26:41 but entertainment in a way that like
0:26:43 we’re talking about you know
0:26:46 anthropic and chat to these general population
0:26:49 but you just said character AI is way more popular
0:26:49 yes
0:26:51 so then like how do we think about that
0:26:52 and I think you know
0:26:53 it’s a very interesting
0:26:55 strategy in my mind
0:26:57 and Grok like in the image and video app
0:26:59 since pretty early on
0:27:00 they’ve had templates
0:27:01 of popular things
0:27:02 like
0:27:03 you’re standing somewhere
0:27:05 and suddenly like a thing drops
0:27:06 a rope drops from the ceiling
0:27:07 and you grab onto it
0:27:08 and it like swings you out of the scene
0:27:09 like
0:27:10 some really good ones
0:27:11 that go viral regularly
0:27:13 on TikTok and other places
0:27:14 yeah
0:27:15 really really interesting
0:27:17 well so maybe switching gears
0:27:18 from 25 to 26
0:27:20 what are some of all
0:27:21 your predictions for next year
0:27:22 what do you think we’ll see
0:27:23 hardware
0:27:24 models
0:27:24 commerce
0:27:26 we haven’t spoken about yet
0:27:28 so what do we think will play out
0:27:29 I think
0:27:30 I know this is
0:27:31 we’re talking about consumer
0:27:32 but one of the things
0:27:33 that’s been really
0:27:35 maybe underrated for me
0:27:35 about ChatGPT
0:27:37 that we might see more of next year
0:27:39 is they’ve really made a push
0:27:40 into the enterprise
0:27:41 both with the traditional
0:27:42 enterprise licenses
0:27:43 and then working with
0:27:44 specific companies
0:27:46 to even like train models for them
0:27:47 and I think when we think about
0:27:48 the fact that
0:27:50 most consumers only use
0:27:52 one general LLM product
0:27:54 ChatGPT enterprise usage
0:27:55 they publish a big study
0:27:57 but it’s up somewhat like
0:27:58 8 or 9x year over year
0:27:58 yeah
0:28:00 and so if we’re entering a world now
0:28:01 where people
0:28:02 have to
0:28:03 use ChatGPT
0:28:04 for their company
0:28:05 or as part of their work
0:28:06 yeah
0:28:07 that could really translate
0:28:08 into consumer usage
0:28:08 yeah
0:28:09 or
0:28:11 maybe they become
0:28:12 the workspace
0:28:13 with the connectors
0:28:14 and some of the other things
0:28:15 that they’re investing in
0:28:16 and someone else
0:28:18 owns the consumer
0:28:19 consumer use cases
0:28:19 yeah
0:28:20 I think
0:28:21 to that end
0:28:22 we have to talk about
0:28:23 their push into apps
0:28:23 and I think
0:28:24 whether or not
0:28:25 that works
0:28:25 is going to be
0:28:27 kind of the defining question
0:28:27 for them next year
0:28:28 yeah
0:28:29 and I think that the
0:28:30 we’ve all discussed
0:28:30 the importance
0:28:31 of the apps SDK
0:28:32 and the apps directory
0:28:33 as they’re calling it
0:28:34 and it’s going to be
0:28:35 a huge new channel
0:28:35 for a consumer
0:28:36 I think what’s less discussed
0:28:37 is it’s hyper relevant
0:28:38 to enterprise
0:28:39 so I think
0:28:40 where ChatGPT shines
0:28:41 is where it’s able
0:28:42 to operate
0:28:43 across a number
0:28:43 of tools
0:28:44 for one workflow
0:28:45 and if you think
0:28:46 about the number
0:28:46 of things you do
0:28:47 in your sort of
0:28:48 business day to day
0:28:49 that operates
0:28:50 across many tools
0:28:51 it’s most of those things
0:28:52 yeah
0:28:52 so I think
0:28:52 that will have
0:28:53 very interesting
0:28:54 implications
0:28:55 for the SaaS ecosystem
0:28:56 and it’s a part
0:28:57 of the app store
0:28:57 we’re not talking
0:28:58 about as much
0:28:58 yeah
0:28:59 yeah
0:29:00 maybe less of a prediction
0:29:01 but I’m thinking
0:29:03 through 2025
0:29:04 and we talked
0:29:05 about all the big
0:29:06 moves from big labs
0:29:08 and from the startup
0:29:08 point
0:29:09 I think one of the
0:29:11 biggest trend
0:29:11 we’ve seen
0:29:12 is app generation
0:29:14 and I think
0:29:15 there is a real world
0:29:17 where we see
0:29:18 the big labs
0:29:19 with the distribution
0:29:20 and the frequency
0:29:20 of usage
0:29:22 of people coming in
0:29:23 to start saying
0:29:24 look like
0:29:24 maybe there is
0:29:25 a common
0:29:27 type of product
0:29:27 and apps
0:29:28 that we could
0:29:28 actually help you
0:29:29 generate
0:29:30 within the confines
0:29:32 of the big lab
0:29:32 products
0:29:32 yeah
0:29:33 I think that’s
0:29:34 like one of the
0:29:34 interesting thing
0:29:35 which you know
0:29:35 again going back
0:29:36 to the supply chain
0:29:37 of ideas
0:29:37 and research
0:29:38 maybe that’s
0:29:38 one thing
0:29:39 and again
0:29:40 nothing groundbreaking
0:29:41 but as we know
0:29:43 the Ghibli
0:29:44 broke the internet
0:29:45 my cousin
0:29:47 who knows
0:29:47 nothing
0:29:48 about tech
0:29:50 sent me
0:29:50 a Ghibli
0:29:51 photo
0:29:52 well let’s not
0:29:53 send this to your
0:29:53 cousin then
0:29:54 yeah
0:29:56 and I think
0:29:57 that goes
0:29:58 to show
0:29:59 that templates
0:30:00 matter
0:30:00 yeah
0:30:01 that style
0:30:01 matters
0:30:02 yeah
0:30:03 and I think
0:30:03 about video
0:30:04 and like
0:30:05 it’s pretty
0:30:06 freaking good
0:30:06 yeah
0:30:08 and it’s possible
0:30:08 that we’re
0:30:09 already at a point
0:30:10 that it’s not
0:30:11 necessarily
0:30:12 just about
0:30:12 the capability
0:30:13 of models
0:30:14 of the big labs
0:30:15 but the
0:30:16 stylistic things
0:30:16 the template
0:30:17 think of
0:30:18 TikTok
0:30:19 the large
0:30:20 capability
0:30:21 largely still
0:30:21 the same
0:30:22 music
0:30:22 trend
0:30:23 dance
0:30:23 go
0:30:25 except the
0:30:26 trend and format
0:30:26 keeps on changing
0:30:27 keeps it extremely
0:30:28 fresh
0:30:29 so I feel like
0:30:29 there’s a real
0:30:30 world where
0:30:32 the repurposor
0:30:32 team or what have you
0:30:33 can start thinking
0:30:34 about ways to
0:30:35 actually really
0:30:36 build in
0:30:37 video first
0:30:38 products
0:30:38 into these
0:30:39 lab models
0:30:40 and I think
0:30:40 the cost will go
0:30:41 down enough
0:30:41 for people to
0:30:42 try it out
0:30:42 and I’m
0:30:43 excited to
0:30:43 see that
0:30:44 yeah I think
0:30:45 what I’m
0:30:45 most excited
0:30:46 about is
0:30:47 sort of along
0:30:48 those lines
0:30:49 basically everything
0:30:50 becoming multimodal
0:30:51 like I call it
0:30:52 like anything in
0:30:53 to anything out
0:30:54 which is
0:30:55 basically initially
0:30:56 especially with
0:30:57 these image and
0:30:57 video models
0:30:59 it was
0:31:00 you put in
0:31:00 a text prompt
0:31:02 and you get
0:31:03 an image out
0:31:04 or a video out
0:31:04 you couldn’t
0:31:05 really do much
0:31:05 with it
0:31:06 and now
0:31:08 we started
0:31:08 to see this
0:31:09 with the
0:31:09 image edit
0:31:10 models
0:31:10 with like
0:31:11 nano banana
0:31:11 and with
0:31:12 flux
0:31:14 and with
0:31:14 the new
0:31:14 open AI
0:31:15 model
0:31:16 where you
0:31:17 can put an
0:31:17 image in
0:31:18 now and get
0:31:19 another image
0:31:19 out
0:31:20 you can put
0:31:20 an image in
0:31:21 with a text
0:31:21 pair in a
0:31:22 direction
0:31:23 or put an
0:31:23 image with
0:31:24 a template
0:31:25 another reference
0:31:26 image and get
0:31:26 another image
0:31:26 out
0:31:28 what happens
0:31:28 when you can
0:31:29 put a video
0:31:30 in and get
0:31:31 images out
0:31:32 that are
0:31:33 related to
0:31:33 or the
0:31:34 next iteration
0:31:35 of the video
0:31:35 or you can
0:31:36 put a video
0:31:37 in and a text
0:31:38 prompt about
0:31:38 what you want
0:31:39 to edit
0:31:39 and get the
0:31:40 edited video
0:31:40 out
0:31:42 from my
0:31:43 conversations
0:31:43 with the
0:31:44 labs
0:31:44 a lot of
0:31:44 them are
0:31:45 trying to
0:31:46 basically
0:31:47 combine all
0:31:48 these largely
0:31:48 separate efforts
0:31:49 they’ve had
0:31:50 across like
0:31:51 text reasoning
0:31:52 and intelligence
0:31:52 the LLM
0:31:53 space and
0:31:54 image and
0:31:55 video into
0:31:56 like what
0:31:56 if we can
0:31:57 put merge
0:31:58 those all
0:31:58 into like
0:31:58 a mega
0:31:59 model that
0:32:00 can take
0:32:00 a lot
0:32:01 different
0:32:01 forms of
0:32:02 content
0:32:03 and produce
0:32:04 much more
0:32:04 I think
0:32:05 it’s also
0:32:05 going to
0:32:05 have huge
0:32:06 implications
0:32:06 for like
0:32:07 design
0:32:08 because if
0:32:08 you think
0:32:09 about it
0:32:09 a lot of
0:32:10 design is
0:32:10 combining
0:32:11 images
0:32:12 with text
0:32:13 with video
0:32:14 with different
0:32:14 elements
0:32:15 in kind
0:32:15 of interesting
0:32:16 ways
0:32:16 yeah
0:32:17 I guess
0:32:18 if I think
0:32:18 about like
0:32:18 a macro
0:32:19 level
0:32:19 prediction
0:32:21 I think
0:32:21 it’s actually
0:32:22 going to be
0:32:22 more of the
0:32:22 same
0:32:23 in that
0:32:24 when we
0:32:24 talk about
0:32:25 what all
0:32:25 of the
0:32:26 labs have
0:32:26 launched
0:32:27 in consumer
0:32:28 they’ve done
0:32:28 a great
0:32:29 job with
0:32:29 models
0:32:30 and they’ve
0:32:30 done a great
0:32:31 job with
0:32:31 incremental
0:32:32 things that
0:32:33 improve the
0:32:33 core experience
0:32:34 of using
0:32:34 like a
0:32:35 chat
0:32:35 gbt or
0:32:36 gemini
0:32:37 in my
0:32:37 opinion
0:32:38 we’ve gone
0:32:38 through dozens
0:32:39 of things that
0:32:40 they’ve launched
0:32:40 or tried as
0:32:41 new consumer
0:32:42 products or
0:32:42 new consumer
0:32:43 interfaces like
0:32:44 group chat
0:32:45 like pulse
0:32:46 like atlas
0:32:47 like sora
0:32:48 google has had
0:32:49 a long tail
0:32:49 like stitch
0:32:50 gems
0:32:51 opal
0:32:51 doppel
0:32:52 tons
0:32:53 none of
0:32:53 those are
0:32:54 really working
0:32:55 and I think
0:32:55 it’s because
0:32:56 it’s not the
0:32:57 core competency
0:32:57 of these
0:32:58 companies
0:32:58 anymore
0:32:59 to build
0:33:00 opinionated
0:33:01 standalone
0:33:01 consumer
0:33:02 ui
0:33:02 out of
0:33:03 all of
0:33:03 those
0:33:04 I think
0:33:04 the product
0:33:04 that’s
0:33:05 working the
0:33:05 most is
0:33:06 like notebook
0:33:06 lm
0:33:07 and that’s
0:33:07 one of
0:33:08 like maybe
0:33:09 20 things
0:33:09 that google
0:33:09 has tried
0:33:10 or experimented
0:33:10 with
0:33:11 so I think
0:33:11 it’s actually
0:33:12 very positive
0:33:13 for startups
0:33:13 in that
0:33:14 consumer
0:33:15 startups
0:33:15 and that
0:33:15 the models
0:33:16 will keep
0:33:16 getting better
0:33:17 which the
0:33:17 startups
0:33:18 can use
0:33:19 and they’ll
0:33:19 keep
0:33:19 you know
0:33:20 they’ll make
0:33:20 chat gbt
0:33:21 better and
0:33:21 better but
0:33:21 I don’t
0:33:22 necessarily
0:33:22 think that
0:33:23 chat gbt
0:33:25 like verticalizes
0:33:26 into all of
0:33:26 these other
0:33:27 amazing use
0:33:27 cases or
0:33:28 products and
0:33:28 there’s still
0:33:29 room for
0:33:29 startups to
0:33:30 be building
0:33:30 there I
0:33:30 have a
0:33:31 yes and
0:33:31 to that
0:33:32 okay
0:33:32 where
0:33:33 absolutely
0:33:34 but however
0:33:35 when the
0:33:36 input and
0:33:37 output is
0:33:37 text
0:33:38 yep
0:33:39 where
0:33:40 chat gpt
0:33:40 and gemini
0:33:40 of the world
0:33:41 shine the
0:33:42 most no
0:33:43 matter how
0:33:43 deeper you
0:33:44 go no
0:33:44 matter how
0:33:45 specific you
0:33:45 think your
0:33:46 text output
0:33:46 is going
0:33:47 to be
0:33:47 essentially
0:33:48 given the
0:33:49 frequency of
0:33:50 usage of the
0:33:50 main
0:33:51 big lab
0:33:51 product
0:33:52 yeah
0:33:52 I think
0:33:52 it’s going
0:33:52 to be
0:33:53 really hard
0:33:54 to stitch
0:33:54 that and
0:33:55 get that
0:33:55 away from
0:33:56 that usage
0:33:57 if your
0:33:57 product is
0:33:59 mainly text
0:33:59 and text
0:34:00 out
0:34:00 yeah
0:34:01 so I do
0:34:02 think you
0:34:02 have to be
0:34:03 creative around
0:34:04 what is the
0:34:05 angle that
0:34:06 you can like
0:34:07 go steal people
0:34:07 away from
0:34:08 you know I love
0:34:09 that you use
0:34:09 the word
0:34:10 opinionated
0:34:10 because I
0:34:11 think that
0:34:11 for labs
0:34:12 certainly for
0:34:13 big tech and
0:34:14 perhaps increasingly
0:34:14 for labs
0:34:15 the priorities
0:34:16 get set in
0:34:16 their promo
0:34:16 committee
0:34:17 always and
0:34:18 if you’re a
0:34:18 PM and
0:34:19 it’s always
0:34:19 the sort of
0:34:20 mid-career
0:34:20 PMs and
0:34:21 I’ve been one
0:34:21 of these
0:34:22 and like the
0:34:22 incentives are
0:34:23 always to get
0:34:24 promoted and
0:34:24 the way to get
0:34:25 promoted is to
0:34:25 build something
0:34:26 safe that
0:34:27 extends a core
0:34:28 metric and a
0:34:29 core feature
0:34:30 so building
0:34:31 opinionated products
0:34:32 is a very risky
0:34:32 way to manage
0:34:33 your career
0:34:33 you know because
0:34:34 they’re probably
0:34:34 not going to
0:34:35 work they’re
0:34:35 probably going to
0:34:36 have a bunch of
0:34:37 implications for
0:34:37 legal and
0:34:38 compliance and
0:34:39 the CEO might
0:34:39 yell at you
0:34:40 so I just think
0:34:41 that they are so
0:34:42 structured to do
0:34:43 incremental things
0:34:43 the more founders
0:34:44 do opinionated
0:34:45 things the more
0:34:46 advantaged they are
0:34:47 I think honestly
0:34:48 the big thing
0:34:48 we haven’t
0:34:49 discussed here
0:34:49 too is compute
0:34:50 which is the
0:34:51 labs have this
0:34:52 inherent tension
0:34:53 between there’s a
0:34:54 limited amount of
0:34:55 compute and they
0:34:56 either spend it on
0:34:57 like training models
0:34:58 or they spend it on
0:34:59 inference and even
0:34:59 with inference
0:35:01 there’s this split
0:35:01 between like the
0:35:02 entertainment Ghibli
0:35:03 use cases and the
0:35:04 like coding
0:35:05 intelligence use
0:35:06 cases I think
0:35:07 XAI is probably the
0:35:08 only model company
0:35:09 that is not
0:35:09 bottlenecked on
0:35:10 compute from my
0:35:11 understanding whereas
0:35:12 the others have
0:35:13 to make really
0:35:14 like serious and
0:35:15 significant calls
0:35:16 of like if we
0:35:17 let if we release
0:35:18 Nano Banana and go
0:35:19 super viral like it
0:35:20 may slow down the
0:35:21 next like big LLM
0:35:22 we’re trying to push
0:35:23 forward whereas
0:35:24 startups who focus on
0:35:25 the app layer don’t
0:35:26 have that problem
0:35:27 because there’s no
0:35:28 tension there
0:35:29 absolutely yeah
0:35:30 we’ve talked about
0:35:31 this before I also
0:35:32 think that there are
0:35:33 categories in which
0:35:34 being multi-model
0:35:35 is just allows you
0:35:36 to deliver a better
0:35:37 proposition to the
0:35:38 customer and the
0:35:39 labs and big tech
0:35:39 are always going to
0:35:40 be sort of
0:35:41 definitionally first
0:35:41 party model only
0:35:42 yeah so I think
0:35:43 as all the models
0:35:44 get better perhaps
0:35:45 80% of what you
0:35:47 need can be received
0:35:47 from a single
0:35:48 model but for the
0:35:49 power users and so
0:35:50 much of AI is a
0:35:51 power user story
0:35:52 yeah you know you
0:35:53 always said that like
0:35:54 well power users are
0:35:55 just power users and
0:35:55 I think that’s true in
0:35:57 a pre-AI world but
0:35:58 now the kind of depth
0:35:59 of value and the
0:35:59 depth of monetization
0:36:01 is so much higher that
0:36:02 maybe all of AI is
0:36:03 actually a power user
0:36:04 story you know and
0:36:05 everyone else is just
0:36:06 traffic yes yeah
0:36:07 which is why we’re
0:36:08 also seeing like
0:36:09 consumer products for
0:36:10 the first time ever
0:36:11 have more than 100%
0:36:12 revenue retention
0:36:13 yes yes and that’s
0:36:14 separating the good
0:36:15 from the great from
0:36:16 the exceptional in
0:36:18 consumer AI and to
0:36:18 be clear how that
0:36:20 happens is they charge
0:36:21 for usage often in
0:36:22 addition to a
0:36:22 subscription so you
0:36:23 can use beyond
0:36:24 whatever your quota
0:36:25 is for the month
0:36:26 given your subscription
0:36:26 and pay more
0:36:27 it’s either upgrade of
0:36:28 the tier yeah or
0:36:29 actually buying tokens
0:36:32 or more usage yeah
0:36:33 it’s that’s what
0:36:34 differentiates it like
0:36:35 you know if you told
0:36:38 me pre-AI we see a
0:36:38 consumer company with
0:36:40 100 plus retention and
0:36:41 money I’m like that
0:36:41 that doesn’t make any
0:36:42 sense doesn’t that
0:36:45 compute yeah yeah no
0:36:46 pun intended exactly
0:36:48 exactly well guys okay
0:36:49 maybe let’s talk about
0:36:50 start with specific
0:36:51 recommendations like
0:36:52 after this pod what are
0:36:53 the products people
0:36:54 should download or the
0:36:54 features or the
0:36:55 models what should
0:36:56 folks be using today
0:36:59 I guess on the
0:37:00 multimodal point I
0:37:01 think one really
0:37:03 under hyped product
0:37:04 that people should
0:37:05 check out not because
0:37:05 they’ll use it every
0:37:06 day but because it
0:37:07 shows sort of what is
0:37:08 possible when you
0:37:09 combine like an agent
0:37:11 with image with text is
0:37:13 Pomelli so this is like
0:37:14 the Google Labs product
0:37:15 where you put in the
0:37:16 URL of your business
0:37:18 and it has an agent go
0:37:20 to the website pull all
0:37:20 of the product and
0:37:22 brand photos summarize
0:37:23 what it thinks your
0:37:25 brand’s aesthetic is
0:37:26 what it stands for what
0:37:27 kind of customers it’s
0:37:28 targeting and then it
0:37:30 will generate three
0:37:31 different ad campaigns
0:37:32 for you and it will
0:37:33 generate not only the
0:37:34 text but it will
0:37:35 generate like the
0:37:36 Instagram posts it will
0:37:37 generate the flyer it
0:37:38 will generate like the
0:37:40 photo of your product in
0:37:42 this you know whatever
0:37:43 wherever it thinks it
0:37:44 should be based on your
0:37:47 customer and very cool
0:37:49 product would be hard to
0:37:49 become a giant
0:37:50 standalone product within
0:37:52 Google I think but
0:37:53 show sort of the future
0:37:54 of what happens if we
0:37:56 combine agents with
0:37:58 generation models that
0:37:59 have sort of really
0:38:00 deep understanding of
0:38:02 context that an image
0:38:03 model or video model
0:38:04 normally wouldn’t have
0:38:06 startup products though
0:38:06 do you have a favorite
0:38:07 startup startup
0:38:09 creative tool yes in
0:38:10 creative tool yes I
0:38:11 think I mean we’re
0:38:12 investors in Korea so
0:38:14 this is bias but I think
0:38:15 they they’ve they’ve
0:38:16 really done an
0:38:17 exceptional job of
0:38:19 being the best place to
0:38:21 use every model or
0:38:22 every quality model across
0:38:24 every modality and also
0:38:25 building more of the
0:38:26 interface on top of
0:38:27 these models like I
0:38:28 now prefer to use
0:38:30 nano banana pro on
0:38:31 Korea because Korea
0:38:33 allows you to save
0:38:34 elements which are
0:38:35 essentially characters or
0:38:37 styles or objects that
0:38:38 you can like at tag to
0:38:40 reprompt versus having
0:38:41 to drag the same image
0:38:42 reference into nano
0:38:43 banana over and over
0:38:44 again it’s a good one
0:38:47 I suppose it falls under
0:38:49 startup category against
0:38:50 shilling companies but you
0:38:52 know the one that I use
0:38:53 the most is actually a
0:38:55 11 labs reader and the
0:38:57 reason is we’ve seen an
0:38:58 explosion in podcasts and
0:38:59 there’s I think a reason
0:39:00 for that right people are
0:39:02 a lot more on the go the
0:39:04 reading capability of us
0:39:05 reading I think it’s going
0:39:08 down over time and so you
0:39:09 know let’s not fight the
0:39:10 reality let’s embrace it and
0:39:12 okay so then like let’s
0:39:13 actually find a written
0:39:15 material translate into
0:39:17 listening and and do that
0:39:19 and I used to be a power
0:39:20 user of tools like
0:39:22 pocket you know I didn’t
0:39:23 have time to read
0:39:23 everything that I wanted
0:39:25 to read and it’s a saving
0:39:26 behavior right you’re
0:39:27 you’re going around and
0:39:28 saving all the things you
0:39:29 eventually want to consume
0:39:30 but I think what I do now
0:39:32 is similar where I go get
0:39:34 all the things I want to
0:39:35 read and I just put it
0:39:36 either PDF it or put it
0:39:38 on 11 reader and just
0:39:39 like once in a while when
0:39:40 I’m on a walk and I’m
0:39:42 like three four minute you
0:39:43 know 1.5x speed or 2x
0:39:44 speed and just listen to
0:39:45 one of these and get the
0:39:46 gist of it so I think
0:39:47 that’s been a good way
0:39:50 to use a little bit of
0:39:51 time as a sort of
0:39:53 semi-normal person yeah
0:39:55 well first of all I love
0:39:56 this question because I
0:39:58 am strongly opinionated
0:39:59 that by far the best way
0:40:00 to get up to speed on AI
0:40:01 is just to try a ton of
0:40:03 products and you get
0:40:04 opinionated really quickly
0:40:06 yeah justine and I
0:40:07 actually for the whole
0:40:08 month of December are on
0:40:10 Twitter publishing one new
0:40:11 consumer product a day for
0:40:12 people to check out so
0:40:15 that’s one way I’ll name
0:40:16 three others that I think
0:40:18 are super maybe relevant
0:40:19 or interesting that people
0:40:19 can plug into their
0:40:21 workflows so one would be
0:40:22 gamma for slide deck
0:40:23 generations you can go
0:40:25 text prompt a slide deck
0:40:26 you can go document a
0:40:28 slide deck I use it for
0:40:29 everything also the slides
0:40:31 are flexible sizes so
0:40:31 you’re no longer like
0:40:33 editing every little pixel
0:40:34 in your Google slides to
0:40:35 get it to fit into one
0:40:37 which is great granola for
0:40:38 no taking you might not
0:40:39 have any meetings over the
0:40:40 holidays but in the new
0:40:42 year and it just gets
0:40:43 better and better the more
0:40:44 meetings you have on it
0:40:45 because it has the context
0:40:46 of what you talked about
0:40:48 before and then lastly I’m
0:40:49 still going to plug try the
0:40:50 comment browser if you want
0:40:52 to try kind of an AI native
0:40:53 workspace I think that’s one
0:40:54 of the most accessible ones
0:40:55 to start with
0:40:57 I mean for me I’ve spent my
0:40:59 whole year obsessed with
0:41:01 coding and AI code it’s just
0:41:02 been so tremendously fun I
0:41:03 by the way Brian would take
0:41:04 the other side of your
0:41:06 argument that the big labs or
0:41:07 big tech will win an app
0:41:09 generation I think they just
0:41:10 lack the focus products like
0:41:12 Opal have been you know
0:41:14 released with a whimper and
0:41:16 they’re one model only so
0:41:17 I didn’t think they will win
0:41:18 it I think that we will see
0:41:20 them doing it yes yes I think
0:41:21 that’s true but I think for
0:41:22 the pure consumer side of
0:41:24 course Wabi is really fun and
0:41:26 really capable and I think
0:41:27 they are they’re creating the
0:41:30 right sort of constraints on
0:41:32 app generation so that you can
0:41:33 get a really satisfying
0:41:34 functional result and I think
0:41:36 so far there’s been a lot of
0:41:37 over promising in app
0:41:38 generation which has
0:41:39 discouraged the early users
0:41:41 I also think if you haven’t
0:41:43 tried you know GPT-5.2 in
0:41:45 Codex or in Cursor it’s worth
0:41:46 trying even for non-technical
0:41:48 people it’s just amazing I
0:41:49 think almost being technical
0:41:50 is sort of a constraint because
0:41:52 you have a pre-existing idea
0:41:54 for what these models can do
0:41:55 and they can do a lot more and
0:41:56 I’m hearing increasingly about
0:41:57 people doing knowledge work
0:41:59 and writing essays in Cursor
0:42:00 instead of just writing code
0:42:02 just one thing I’m going to do
0:42:04 at the end year end it’s just
0:42:06 to plug in like a popular trend
0:42:08 I see on TikTok where there
0:42:09 are people who said what is a
0:42:11 most unhinged thing I said
0:42:13 this year okay and it actually
0:42:16 does a review of all the things
0:42:17 that you said but I think
0:42:19 similarly it’ll be a good thing
0:42:21 I’m going to do this at the
0:42:23 year end tell me how to live a
0:42:25 better life next year yeah give
0:42:27 me give me actual unvarnished
0:42:30 opinions and some directions and
0:42:32 I think it’ll be helpful I love
0:42:33 that idea I’m going for a worse
0:42:37 life next year fantastic let’s go
0:42:39 full dgen guys any closing
0:42:44 thoughts that was I mean the
0:42:45 obvious one is we are very
0:42:46 actively investing in consumer
0:42:48 companies and I genuinely I think
0:42:49 a lot of people say this I
0:42:52 genuinely believe that the models
0:42:53 have gotten to the level of
0:42:55 quality that you can build a real
0:42:57 scalable app on top of them
0:42:59 Wabi is a great example of this
0:43:02 and so the hope is 2026 will be a
0:43:04 huge year for consumer builders
0:43:06 not just like consumers as
0:43:10 consumers being consumers of a
0:43:12 product yes well thank you all for
0:43:14 a super fun year in consumer and
0:43:15 AI we’ll be back with more next
0:43:17 year and Merry Christmas guys this
0:43:19 is a wrap yeah happy holidays
0:43:25 thanks for listening to this episode of
0:43:27 the a16z podcast if you like this
0:43:29 episode be sure to like comment
0:43:31 subscribe leave us a rating or
0:43:32 review and share it with your
0:43:34 friends and family for more
0:43:36 episodes go to YouTube Apple
0:43:39 podcast and Spotify follow us on X a
0:43:41 a16z and subscribe to our sub
0:43:44 stack at a16z.substack.com thanks
0:43:46 again for listening and I’ll see you in
0:43:49 the next episode as a reminder the
0:43:50 content here is for informational
0:43:52 purposes only should not be taken as
0:43:54 legal business tax or investment
0:43:56 advice or be used to evaluate any
0:43:58 investment or security and is not
0:43:59 directed at any investors or
0:44:01 potential investors in any a16z fund
0:44:04 please note that a16z and its
0:44:05 affiliates may also maintain
0:44:06 investments in the companies
0:44:08 discussed in this podcast for more
0:44:09 details including a link to our
0:44:12 investments please see a16z.com
0:44:14 forward slash disclosures