The Large Language Model Race with Pete Huang, Founder of The Neuron

0:00:02 Google said, we’ve got a new large language model that’s awesome.
0:00:04 And then open AI went, hold my beer and watch this.
0:00:05 That was crazy.
0:00:06 That was absolutely crazy.
0:00:06 Right.
0:00:12 And look, I think the whole thing around the one person business that can
0:00:16 generate 10 million a year, 20 million a year, a billion dollars, whatever sort
0:00:18 of fancy thing you want to put on it is possible.
0:00:23 Hey, welcome to the next wave podcast.
0:00:23 My name is Matt Wolf.
0:00:25 I’m here with my co-host, Nathan Lanz.
0:00:28 And today we’ve got an amazing guest for you.
0:00:33 Someone that loves to nerd out around AI and keeps his finger on the pulse of AI
0:00:35 just as much as we do today.
0:00:40 We’ve got the founder of one of the most popular AI newsletters, the neuron on the
0:00:42 show. Today we’re talking with Pete Huong.
0:00:44 Thank you so much for joining us, Pete.
0:00:44 How are you doing?
0:00:45 How’s it going, everyone?
0:00:46 Good, it’s good to be here.
0:00:51 Can you give us the quick rundown of what the neuron is, what type of reader
0:00:54 you have, what your goal with each issue is?
0:00:56 So the neuron is a daily AI newsletter.
0:00:59 It’s filtered through the lens of knowledge workers, non-technical
0:01:01 audience, small business owners.
0:01:04 And we give you updates every day, money through Friday.
0:01:05 It’s a very, very quick read.
0:01:09 We do two stories and a bunch of links just showing what’s going on.
0:01:13 And what we try to do is filter through the lens of like really the average
0:01:16 person. There’s a lot of, you know, talk around the technicals of AI.
0:01:21 But if we zoom way back and make sense of how does it actually manifest in
0:01:25 your work, in your business and how you make changes in how you do work?
0:01:26 What does it actually look like?
0:01:27 Yeah, that’s amazing.
0:01:29 I mean, congrats on the success of that newsletter.
0:01:30 That is huge.
0:01:34 I have to ask you one question because it’s the question that people ask me
0:01:37 the most. So I actually want to turn it around on somebody else.
0:01:41 Is there anything like in the AI world that you’ve kind of like heard of?
0:01:45 It’s in the pipeline that maybe you’re aware of or you’ve seen already
0:01:47 that has you really excited right now?
0:01:49 Oh, it’s the agents for sure.
0:01:50 Agents for sure.
0:01:55 Because if anything, because my lens is on what running a business
0:01:59 looks like in the future, and today, right now, a lot of business owners,
0:02:02 enterprises, small business, everyone included, they’re a little bit confused
0:02:06 by AI because they’re like, okay, well, the press and the media,
0:02:09 they’re promising me these transformative things.
0:02:12 I should be able to translate, run a completely different business
0:02:13 because of AI today.
0:02:18 And that’s not actually happening because the reality of AI, at least in
0:02:22 the chat to the world as it is today, is it’s automating what we call
0:02:24 like the bits and pieces of work, right?
0:02:30 It is, I’m a marketing coordinator and I can save 10 to 30 minutes of my
0:02:34 day every day because of this small little thing that I now created a prompt
0:02:36 in Claude or chat to BT to do.
0:02:39 And then now it’s giving me a first draft and I have to spend that time doing
0:02:43 it. That is not as transformative as it is for that one particular person.
0:02:48 For the business owner, it’s not like wholesale transformative, right?
0:02:53 Agents, I think have a much bigger promise along that front because you’re
0:02:58 now going to unlock end to end automation for an entire process, right?
0:03:02 So if you have an agent chain that can go from creative concept to a brief,
0:03:06 to a proposal, to all the iterations, to all the back and forth, like really
0:03:08 all those things and handle all those pieces of work.
0:03:12 Now you’re talking about for a large corporation, for example, a set of AI
0:03:15 agents could take on even 10%, 20%.
0:03:17 That’s huge, right?
0:03:18 That is a huge amount of work.
0:03:20 And those people can now do other things.
0:03:26 And now you’re looking at a real multiples of creativity, not this 10 to 20%
0:03:28 thing, you know, on, on a per person basis.
0:03:32 You’re talking about entire sort of like chunks of work at a time.
0:03:38 I think the whole thing around the one person business that can generate 10
0:03:41 million a year, 20 million a year, a billion dollars, whatever sort of fancy
0:03:44 thing you want to put on it is possible, right?
0:03:45 But you need agents to get there.
0:03:49 I don’t think you sitting in front of chat to BT every day is going to do it.
0:03:52 That is again, giving you 50% leverage on your time.
0:03:54 We’re really looking for a hundred acts, right?
0:03:56 And I think agents represent that hundred acts.
0:04:02 When all your marketing team does is put out fires, they burn out fast.
0:04:06 Sifting through leads, creating content for infinite channels, endlessly
0:04:09 searching for disparate performance KPIs, it all takes a toll.
0:04:13 But with HubSpot, you can stop team burnout in its tracks.
0:04:17 Plus your team can achieve their best results without breaking a sweat.
0:04:22 With HubSpot’s collection of AI tools, Breeze, you can pinpoint the best leads
0:04:27 possible, capture prospects attention with clickworthy content and access
0:04:29 all your company’s data in one place.
0:04:31 No sifting through tabs necessary.
0:04:33 It’s all waiting for your team in HubSpot.
0:04:37 Keep your marketers cool and make your campaign results hotter than ever.
0:04:40 Visit hubspot.com/marketers to learn more.
0:04:46 So when, when you say an agent, let me just kind of clarify.
0:04:50 A, so I know that we’re we’re talking the same language, but B, the
0:04:51 listeners know what we’re talking about.
0:04:55 When you say an agent you’re talking about, you give it, you go into like
0:04:59 a chat window, like a chat GPT, you give it a task that you want it to complete.
0:05:04 But instead of just giving you a single response back, it will keep on working
0:05:09 on that task and iterating and iterating and possibly even using external tools
0:05:13 and APIs until it reaches the end goal that you initially prompted it with.
0:05:14 That’s exactly right.
0:05:18 And so I think to me, the, the easiest way to explain it is there are three
0:05:21 parts, exactly all the things that you just said, Matt, today, right now in
0:05:26 chat, it’s like very directed, which is like, I want you to write a blog post
0:05:26 about this.
0:05:28 And it’s like a very specific thing.
0:05:29 Like it’s very self-contained.
0:05:35 There’s no extra research involved, but an agent has the ability to plan.
0:05:39 It has the ability to reason and has the ability to use other tools.
0:05:42 And so instead of saying, write a blog post, you might give it something a
0:05:46 little bit more abstract and say, come up with them, come up with and
0:05:50 execute the marking strategy for my business and it’ll go through everything
0:05:53 around, okay, like I’m going to research, what does the marketing strategy mean?
0:05:55 What does it mean to have a good marketing strategy?
0:05:57 What does it mean to execute it?
0:05:59 What are the sort of best practices for your line of work?
0:06:02 What are all the pieces and sort of like compile everything.
0:06:06 And as it’s sort of reflecting on its own work, it’s reasoning through the problem.
0:06:09 It’ll then go into execution phase and start to use tools, right?
0:06:12 Which is, okay, now I know that I need to create an Instagram post.
0:06:17 So let me go draft the thing and then go call the Instagram API and
0:06:20 upload the post every week, every two days, whatever it is.
0:06:25 And so now all these things that were normally done by a human going from chat
0:06:30 GPT to clicking over to Instagram.com, whatever it is, is now replaced by code,
0:06:30 right?
0:06:35 It’s, it’s, it’s code and, and these large language models, um, that have just
0:06:37 like developed enough capability, right?
0:06:40 And, and so this is sort of the prospect of GPT five, llama three, et cetera.
0:06:42 Um, to do all these things.
0:06:47 And so this is getting very close, honestly, to what it feels like to have an AI employee.
0:06:47 Right.
0:06:52 Like in some cases, I know previous jobs that I’ve had where the only value
0:06:55 ad that I’ve brought to the table was the fact that I could think because I have
0:06:57 a human brain and I was clicking on software all day.
0:07:01 Like I was just moving bits and pieces of information from one software to another.
0:07:03 Like an agent could totally do that.
0:07:04 You know what I mean?
0:07:07 Like an agent can go and build a Salesforce dashboard for you.
0:07:09 It can go run SQL for you.
0:07:10 It can do all these things.
0:07:14 So there’s like this whole new wave of, of capabilities and startups and
0:07:18 products that are all agentic in this way that, that, that’ll come hopefully in
0:07:19 the next couple of years.
0:07:22 And again, my lens is on just like the future of business, right?
0:07:26 Like I think that’s going to like really fundamentally change how businesses are run.
0:07:26 Yeah.
0:07:31 If it feels like probably GPT five will be the first time where agents kind of
0:07:34 start working because like, you know, I wrote this big Twitter thread back, you
0:07:36 know, like a year ago about baby AGI, right?
0:07:37 It weren’t really viral.
0:07:40 I went on TV and talked about it and it was really promising, but like, you know,
0:07:43 all of those, you know, you’d boot them up and get them working on something.
0:07:45 And it would just end up spinning in circles because it wouldn’t be able to
0:07:46 complete the task.
0:07:48 It would basically get lost, you know?
0:07:52 And so I think as soon as you get where they can actually complete tasks and come
0:07:56 back and show you the results or even go on and do the next logical step, that’s,
0:07:59 that’s going to be such a huge unlock for businesses and entrepreneurs.
0:08:04 And, and the rumor, the rumor is that with GPT five, people have seen demos and it
0:08:05 can not that it’s perfect.
0:08:07 I’m sure it’s going to like take many, many years to get amazing.
0:08:11 But apparently, yeah, you can tell to do some task and it just goes off and does
0:08:13 them and come back and reports to you.
0:08:14 Yeah.
0:08:14 Yeah.
0:08:18 And, and Nathan, that getting lost thing, I think is the unlock, right?
0:08:26 Cause right now, even if it’s 2% off, 3% off, you take point, like 97% times 97%
0:08:30 times 97% times however many steps you want to put to complete.
0:08:34 And like at some point there’s like a guarantee that it will not complete it
0:08:36 correctly if you have enough steps, right?
0:08:38 It’ll, that number will converge to zero.
0:08:45 And so the, the real part is like, how do you get that 97% correct to as high as
0:08:50 possible, like 99.999999999% or even 100% theoretical?
0:08:51 I have no idea, right?
0:08:52 But that’s the biggest gap.
0:08:55 There are a couple of platforms out today, but that’s just how it is, right?
0:08:57 Like it is just not there yet.
0:09:01 And so once you get past three, four steps, it just starts to crash every
0:09:02 single time.
0:09:05 I think everyone needs to be thinking about how they’re going to use agents when
0:09:08 they get good, because it’s going to create so much competition that if you’re
0:09:10 not using them, they’re going to be so far behind.
0:09:13 If you think of like the Peter levels of the world, who just like crank out, like,
0:09:16 Oh, I made a new startup and now it’s making $20,000 a month.
0:09:18 And like, in the next month, he does another one, right?
0:09:20 He just keeps doing this somehow.
0:09:24 Imagine when he has agents where he can literally just like, here’s my idea.
0:09:25 I have a social following.
0:09:26 So I can kind of jumpstart it.
0:09:28 I don’t have to worry about like no one’s going to try it.
0:09:30 At least a thousand people will try anything I do.
0:09:34 The agent goes and makes the landing page an hour after I thought about it on
0:09:38 the beach and now I’m testing the idea.
0:09:38 It didn’t work.
0:09:38 Okay.
0:09:40 Well, tonight we’ll try some more ideas.
0:09:43 It’s just like, that’s going to be amazing.
0:09:43 Totally.
0:09:47 The week that we’re recording this kind of the biggest thing going on in the AI
0:09:50 world is all of these new large language models dropping.
0:09:54 We got Gemini 1.5 now became available publicly.
0:09:56 We’ve got GPT-4.
0:10:01 They just released a new April 9th version that they claim is a major update,
0:10:03 but didn’t really tell anybody what they updated about it.
0:10:04 Just that it’s a major update.
0:10:12 We got Mixerl 8X7B that just got released, which I believe is open source.
0:10:14 The 8X7B was open source.
0:10:16 So I believe this one is as well.
0:10:19 And then we also got the announcement from META that Lama 3 is most likely
0:10:20 coming in May.
0:10:25 So lots of large language model news going on right now.
0:10:29 What is what is your take on large language model news?
0:10:34 Do you see like one of these companies like being a clear winner?
0:10:38 Do you think there’s place for all of these large language models to all coexist?
0:10:40 Like what’s your take on it?
0:10:41 It’s a race.
0:10:44 I mean, it’s all a race to build the best models, right?
0:10:46 And so, of course, you have your GPT-4.
0:10:49 There’s that upgrade that happened this week.
0:10:52 Everyone’s sort of already talked about GPT-5, though, you know what I mean?
0:10:56 It’s like when when both Sam Altman and Satya Nadella are talking
0:10:58 openly about GPT-5, like it’s coming.
0:11:01 Like everyone knows it’s in the works and it’s just a matter of time.
0:11:05 I think rumors, if I have this correct, guys, it’s like summer, I believe.
0:11:07 That’s GPT-5. Is that right?
0:11:07 Yeah.
0:11:09 And so it’s just like all these models are great.
0:11:13 I mean, like Claude 3, I think was is amazing.
0:11:17 And I still think Claude today, chat bot wise, is is the best option.
0:11:19 So that was a big upgrade over GPT-4.
0:11:25 Yeah, we had Gemini 1.5 public preview availability for for all developers.
0:11:30 A lot of advancements there that they claim, you know, Gemini 1.5 beats GPT-4.
0:11:34 You have all these the sort of like second group of the
0:11:37 mixed rolls, et cetera, that are trying to make their mark and and and find a
0:11:40 position in the market, you have Lama 3 with with Meta.
0:11:46 But all of this is sort of looming with the sort of shadow of GPT-5 coming.
0:11:46 You know what I mean?
0:11:48 Like everyone’s comparing it against GPT-4.
0:11:51 But GPT-4 has been out for a year already.
0:11:53 And that was like the old GPT-4, you know what I mean?
0:11:58 And so you factor in like this week’s upgrade with some math capabilities.
0:12:01 And then you just have to ask yourself like, OK, if everyone’s just now
0:12:05 catching up to GPT-4, are they just going to get blown out of the water by GPT-5?
0:12:06 You know what I mean?
0:12:09 Yeah, yeah, I’ve been thinking, too, that like the GPT-5 is probably
0:12:11 way ahead of everything else we’re seeing.
0:12:18 And some evaluations came out for GPT-4, the new upgrade, maybe a few hours ago,
0:12:22 and it’s showing that actually it’s outperforming Clod again in coding.
0:12:25 And it’s kind of weird, too, because like they released it and it’s so vague.
0:12:27 And like you said, yeah, GPT-4 has been out for a year.
0:12:30 They’re still releasing updates to GPT-4.
0:12:32 They’re not calling it 4.1 or 2.
0:12:33 It’s kind of confusing.
0:12:36 And I kind of wonder if they’re doing it on purpose, just to like, you know,
0:12:39 they’re so far ahead, it’s like, oh, yeah, compared to the old model and we’ll
0:12:43 slightly like you, you know, Clod gets ahead, we’re always just barely
0:12:47 better than the best model until we release the real thing we’re actually
0:12:51 focused on, which the thing they’ve been really focused on is GPT-5, not 4.
0:12:57 When you look at it from a product marketing and PR lens, OpenAI knows
0:12:58 that they’re ahead, right?
0:12:59 Like, let’s just be, let’s put it out there, right?
0:13:01 Like they’re they’re pretty much ahead.
0:13:05 And if you see their behavior over the last year and a half, every single
0:13:09 time Google has launched something, they’re always ready.
0:13:12 And in fact, I remember, if you remember last spring, every time Google
0:13:16 made an announcement, Google or OpenAI made an announcement that was
0:13:17 like 10 times better.
0:13:20 And everyone was just like, it created this atmosphere of just like,
0:13:21 what is Google doing?
0:13:22 You know what I mean?
0:13:25 And Google’s trying, they’re like trying to release new things or prepping
0:13:29 their launches, their PMs or PMMs, like going nuts, trying to craft a
0:13:30 campaign and a message that works.
0:13:33 And OpenAI was like, boop, let me just go, like, drop something out.
0:13:37 Oh, here’s another one from the backlog that just like beats you, right?
0:13:39 And, you know, so this one is ready.
0:13:43 And to your point, Nathan, like this thing that dropped on the April 9th
0:13:47 update for GPT-4 wasn’t actually a real launch, you know what I mean?
0:13:50 From a product marketing perspective was not a real, there was no blog post.
0:13:54 There was no specificity into like what actually improved other than like, I
0:13:55 think it’s better than math.
0:13:56 You know what I mean?
0:13:57 Yeah, it got a little bit better.
0:13:57 Yeah.
0:13:59 Well, there was somebody who said like major update or major upgrade.
0:14:03 And there was others like upgrade or update is like, that was that was
0:14:04 different messaging from OpenAI.
0:14:05 Totally, totally.
0:14:09 And so it’s just like, they had to have been sitting on this for so long that
0:14:11 they’re kind of like, okay, we know Google Cloud Nexus coming.
0:14:14 We know they’re going to do something, Google Gemini 1.5.
0:14:14 Yeah.
0:14:15 Let’s just be ready.
0:14:19 And just like even a half baked release or an upgrade of whatever, like, wouldn’t
0:14:21 even need to be specific about it.
0:14:22 We’ll just drop it.
0:14:25 And then people, you know, then we’ll be talking about it on this podcast, right?
0:14:27 Versus Gemini 1.5, right?
0:14:32 So it’s like, it’s just a masterclass, I think, in OpenAI flexing on the
0:14:35 rest of the market and knowing very clearly what position they’re in.
0:14:40 Yeah, Paul Graham always said that like Sam Altman really shined in terms of
0:14:42 strategic thinking and planning.
0:14:44 And so I think a lot of people are not realizing that.
0:14:47 Like he’s like one of the masters at being strategic and planning things out.
0:14:50 And I don’t think a lot of the people think like, oh, they’re just kind of
0:14:52 floundering right now because GPT-5 is not out.
0:14:54 I was like, no, I think this is like part of the plan.
0:14:58 And then when it comes out, it’s going to be so much better than everything else.
0:15:00 And I think Claude and other ones are going to be like, oh, wow.
0:15:02 They’re like way behind.
0:15:03 That’s my personal belief.
0:15:06 Even going back to Pete’s point about just sort of OpenAI always sort of
0:15:10 having something queued up to be ready, wasn’t it the same week that
0:15:15 Gemini 1.5 was announced that OpenAI dropped the Sora video?
0:15:18 I’m pretty sure that was like the same week.
0:15:21 Google said, we’ve got a new large language model that’s awesome.
0:15:23 And then OpenAI went, hold my beer and watch this.
0:15:25 And then put out the Sora videos.
0:15:26 That was crazy.
0:15:27 That was absolutely crazy, right?
0:15:30 And look, to Google’s credit, and maybe this is where we start talking
0:15:35 about Gemini a little bit, but like that release when Google dropped Gemini 1.5,
0:15:36 that was a pretty big deal.
0:15:41 Like there’s a lot going on with Gemini 1.5 that matters a ton.
0:15:44 And it wasn’t just like this, let me catch up to GPT-4 thing.
0:15:46 It was like a pretty landmark statement.
0:15:48 But it made for a crazy news week, right?
0:15:52 Because you had Gemini 1.5 and everyone’s just like, oh my gosh,
0:15:55 now video is a real thing with Sora.
0:15:58 And like before, like those clips weren’t like the video
0:16:00 genders weren’t actually doing a thing.
0:16:02 And then now it’s like a real sort of like thing to watch out for.
0:16:05 I remember that it was absolutely wild.
0:16:11 But I’m curious, Matt, if we step back here and kind of just go back to Google
0:16:14 Cloud next and all the sort of announcements there.
0:16:17 Gemini is definitely, I think, one of the the headliners coming out one.
0:16:20 Can you just help me level set the Gemini conversation?
0:16:21 Just like, what is it?
0:16:24 Like, what is the big deal about it over something like GPT-4?
0:16:29 And then let’s talk about what’s coming with this model and all the other
0:16:30 rest over the next few months.
0:16:34 We’ll be right back.
0:16:37 But first, I want to tell you about another great podcast you’re going to want to listen to.
0:16:41 It’s called Science of Scaling, hosted by Mark Roberge.
0:16:45 And it’s brought to you by the HubSpot Podcast Network, the audio
0:16:47 destination for business professionals.
0:16:52 Each week, host Mark Roberge, founding Chief Revenue Officer at HubSpot,
0:16:56 senior lecturer at Harvard Business School and co-founder of Stage 2 Capital,
0:17:00 sits down with the most successful sales leaders in tech to learn the secrets,
0:17:03 strategies, and tactics to scaling your company’s growth.
0:17:09 He recently did a great episode called How Do You Solve for a Siloed Marketing
0:17:11 in Sales, and I personally learned a lot from it.
0:17:13 You’re going to want to check out the podcast.
0:17:17 Listen to Science of Scaling wherever you get your podcasts.
0:17:24 So one of the biggest things about the Gemini 1.5 model when they announced
0:17:29 it was that it was going to have up to a 10 million token context window.
0:17:34 But I think the release version that was going to be made available was going
0:17:37 to have up to a million, but then the version that we actually have now,
0:17:39 I think it’s like 248,000 or something like that.
0:17:44 So the version that we have available to us right now inside of like the Gemini
0:17:49 platform, I think is like a 258,000 somewhere in that range.
0:17:51 I don’t know the exact number I’ll top my head, but it’s somewhere in that range.
0:17:54 So it’s the largest context window we have available.
0:17:56 You know, supposedly it’s much better at coding.
0:18:00 I actually haven’t tested it myself for coding yet, but the context window,
0:18:02 I think was really, really that big leap.
0:18:07 And in the article that they shared, they did the needle in a haystack test,
0:18:12 right, where they had a, you know, 700,000 word article.
0:18:15 They put a single sentence that was unrelated to the rest of the article
0:18:19 somewhere inside of it and then asked questions about it.
0:18:24 And the needle in a haystack test found the answer like 99.9% of the time.
0:18:27 So it did really well on the needle in a haystack test for like the,
0:18:30 like the retrieval augmented generate generation tests.
0:18:32 And it also has that huge context window.
0:18:34 I’d say those are probably the biggest factors.
0:18:38 Well, I saw a tweet saying that you could upload files up to two gigabytes.
0:18:40 And I don’t know if I’ve actually, I haven’t confirmed that, but like,
0:18:41 that sounds incredible.
0:18:44 Like if you can actually like upload incredibly large documents to this thing
0:18:47 and have it evaluate them or summarize them or, you know,
0:18:50 be a last questions of that document, that’s, that’s huge.
0:18:53 I’m not sure how big chat GPT is, but I feel like, yeah, I don’t know
0:18:57 the upload number on chat GPT, but I know Bill of all also shared something
0:19:01 either yesterday or today around Gemini 1.5 where he uploaded an entire
0:19:05 like interview and had Gemini create the timestamps for it.
0:19:07 So it actually listed the whole interview made timestamps.
0:19:08 And he said they were super accurate.
0:19:13 I’ve tried doing that with chat GPT and A, you have to sort of export
0:19:15 the transcript first, right?
0:19:18 If you’re using Gemini, apparently you can just upload the video
0:19:20 and it’ll just pull the data from the video.
0:19:25 Chat GPT, you have to grab the transcripts, copy, paste them into chat GPT,
0:19:26 tell it to write you timestamps.
0:19:30 It will actually get the sort of notes for the timestamps, right?
0:19:31 But it’ll have the timing way off.
0:19:36 Like the actual like time that it wants you to jump to will not correlate at all.
0:19:41 So that’s another big leap is supposedly it’s getting better at actually,
0:19:46 you know, the multi-modality, multi-modality so far has mostly meant
0:19:50 we can upload text, we can upload images, we can talk to it.
0:19:52 But all it’s doing behind the scenes is transcribing it.
0:19:54 So it’s still just kind of text, right?
0:19:56 Yeah, there’s video understanding.
0:19:58 So just to put some numbers out there, right?
0:20:02 So with Gemini, I commonly cited, although it sounds like the release
0:20:06 is a little smaller than this, but the one million token context window,
0:20:08 like one of the comments that we get a lot is like, OK, tokens,
0:20:12 like, why did you invent like a new unit to count things by?
0:20:15 So token is like three fourths of a word, basically.
0:20:16 It’s like three or four characters.
0:20:20 And if you if you have a million tokens, then you should assume that’s
0:20:25 like 700,000 words in like one lengthy conversation with with Gemini.
0:20:28 You can feed 700,000 words into Gemini, right?
0:20:33 That also roughly equates to one hour of video, about 10 hours of audio
0:20:35 in terms of understanding there.
0:20:38 So to your point, to Bill of All’s experiment, right?
0:20:41 It’s like, man, you can just toss an hour video in there
0:20:42 and it’ll have full understanding.
0:20:45 And this is where that needle in the haystack test gets really important.
0:20:50 So stepping back for a second, the needle in haystack test, as Matt was describing,
0:20:53 is yeah, you just bury some piece of information just like randomly
0:20:57 in this huge amount of text, like somewhere in 700,000 words.
0:21:00 You just drop in that random thing and just see if it can actually detect it.
0:21:04 The reason this is important is when anthropic first released
0:21:06 Claude, like an upgraded version of Claude.
0:21:12 They had this thing where like, oh, like we can do like 128,000 tokens,
0:21:14 which at that point was like pretty significant, right?
0:21:17 Like compared to chat to BT, which is in the five digits.
0:21:20 And then someone ran a test that this needle in haystack test.
0:21:24 And then it turns out that like the more information you were giving Claude,
0:21:28 it would get worse and worse at finding the needle in the haystack.
0:21:32 And there was something they found something where it was
0:21:35 just wasn’t just wasn’t performing correctly, according to spec.
0:21:39 And so this is sort of like a test to show it’s like not only
0:21:44 do we have the ability for you to just like submit one million tokens
0:21:46 worth of stuff, but it actually works, right?
0:21:47 It’s like it’s not going to drop it.
0:21:49 It’s not going to progressively get worse.
0:21:53 It’s like this sort of like a uptime guarantee or like a quality
0:21:56 guarantee of your of your of your token context window.
0:21:58 So there’s so much here, right?
0:22:02 Like I think with one hour, like 10 hours of audio, there is so much
0:22:05 you can do with that that, you know, we were trying to come up with
0:22:07 use cases or like products that could exist.
0:22:11 It’ll be things like, OK, now there’s like some speaking coaches, AI powered
0:22:14 apps that’ll do speaking training and coaching.
0:22:17 It’s like as you’re on a call or like you can upload a video and then
0:22:21 it’ll analyze the transcript and then pick out the ums and the us and
0:22:25 all the repeat words and all these things and tell you like, you know,
0:22:28 you should be using, you should be speaking with more authority.
0:22:31 You’re, you’re, you’re hedging too much or you’re using ums and us too much.
0:22:36 Now imagine it’s actually analyzing the video of you talking.
0:22:39 So in like a public speaking setting and giving you feedback on your stage
0:22:41 presence, right?
0:22:43 Because actually you can see the video now and like what, what you’re
0:22:45 actually doing in the video.
0:22:48 So it’s like these types of things that I think are really interesting.
0:22:51 And I’m just so excited about all the apps that are getting built.
0:22:54 Now that again, developers have wide access to this, public access to this.
0:22:57 You can imagine that a lot of builders who are like, oh man, like there’s
0:23:01 this thing that I’ve been wanting to build and like it requires video access.
0:23:02 And now have video with Gemini.
0:23:05 And so we get kind of get to experiment with that.
0:23:07 So I’m just super excited about all those ideas that are going to come out.
0:23:14 Well, and also there’s Hume as well, H-U-M-E where it actually understands the context.
0:23:17 I don’t know if it actually reads your face right now, but it listens to your
0:23:20 voice to try to understand whether you’re sad, happy, angry, right?
0:23:24 It tries to understand the emotion in your voice and supposedly down the line,
0:23:29 it’ll actually look at your face and try to, you know, pick up micro cues in
0:23:31 your face to figure out the emotions as well.
0:23:33 So that even takes it to another level.
0:23:36 You were talking about like being on stage and it giving you feedback about
0:23:37 your stage presence.
0:23:41 Imagine if it can tell when you’re more nervous, when you’re more excited,
0:23:44 when you’re more happy and even give you feedback on that stuff on stage.
0:23:47 Yeah, yeah, that’s coming, I’m sure.
0:23:48 Yeah.
0:23:54 Nathan, is there anything you’re excited about with, be it Gemini or Llama 3 or
0:23:55 any of these models that are coming out?
0:23:57 Anything that you’re particularly looking forward to?
0:24:01 On the open source side, you know, Elon Musk is saying that Grock is going to
0:24:04 be as good as GPT-4 or better very soon.
0:24:08 So to have that kind of open source version where you can kind of ask it
0:24:11 whatever and not worry about how it’s filtering, you know, cut.
0:24:15 And if you saw some of the stuff with Claude, how, how they’re filtering
0:24:18 what you say to it, like where they actually, they don’t only actually tell
0:24:22 Claude, they don’t tell the LLM what you’re actually saying even.
0:24:25 They pre-filter what you type.
0:24:28 And so they don’t, they don’t even give that your actual message to Claude.
0:24:34 So basically they’re, you write it a prompt, it goes and rewrites the prompt
0:24:36 for you in a way that it thinks is probably going to get you a better
0:24:40 response, but also in some ways may potentially censor you, right?
0:24:43 It is like making sure that it’s, you’re saying things appropriately or whatever.
0:24:45 It is doing that.
0:24:47 So it’s like, and it’s not even showing the LLM that that’s like their
0:24:49 solution is like to literally change your own words.
0:24:53 I just, I think it is good to have a kind of an open source alternative where
0:24:57 like, okay, if open AI and all those go way too hard on that side, that
0:25:01 there’s an open source alternative where you can ask it things and not have
0:25:01 to worry about that.
0:25:04 We’ve talked about all these large language models from Google, from open
0:25:07 AI, from Anthropic, from all of these places.
0:25:11 And Grock didn’t come up once until just now when Nathan mentioned it.
0:25:12 Yeah.
0:25:16 Grock has an open source model that anybody can use if they want.
0:25:19 I mean, I don’t know if most people could run it on their computers
0:25:23 with the size of it, but anybody can access it however they want.
0:25:30 And supposedly he’s about to drop a model that is as good as GPT-4 that is
0:25:32 going to be widely available to users of Twitter.
0:25:37 So it’s just kind of interesting to me how much Grock gets sort of left out
0:25:40 of the conversation when it comes to large language models.
0:25:42 Well, it’s just not as capable as far as I understand it, right?
0:25:42 Yeah.
0:25:46 Like in open source, you have the meta stuff with LLMA.
0:25:49 You have mixed role that’s doing really well.
0:25:52 And then you have like the likes of Grock, et cetera.
0:25:55 But then when you look at the surveys, for example, I think it was
0:25:59 in Dresden Horowitz that did a survey of enterprise buyers and practitioners
0:26:04 for generative AI, basically what they were saying is that the three or four
0:26:10 players that any sort of practitioners considering using model wise is open AI
0:26:15 because they have the leading model is meta LLMA because it’s open source.
0:26:19 And it is just like the only sort of option, basically.
0:26:22 Like why would you give up anything on the capability front when you
0:26:25 just need your best, the singular best open source model?
0:26:31 And then you have Google and Claude or anthropics models similarly
0:26:34 because of sort of the latest capability reasons, right?
0:26:39 And so from that perspective, it’s sort of like, why would you consider
0:26:45 even even Mistral stuff, even Grock stuff if it’s not going to hold up to meta LLMA?
0:26:47 Like, I think that’s frankly the answer.
0:26:51 I don’t use Grock because I just don’t think it’s as good as until it’s good.
0:26:53 I know I see no reason to actually switch over.
0:26:57 I know that like the fact that Elon Musk is behind the project
0:27:00 and actually really pushing it, I think long term he is going to catch up
0:27:02 and surpass a lot of people. I do believe that.
0:27:06 One other interesting thing about Grock, right, is that it does have access to Twitter.
0:27:10 So it’s trying to use like all the conversations on Twitter as part of the context.
0:27:13 But we also saw last week that that can be an issue, right?
0:27:17 Because last week there was an issue where Grock reported some news
0:27:18 that wasn’t even real news.
0:27:21 They just pulled from like memes on Twitter thinking it was real news
0:27:23 and then fed it back to somebody as if it was real.
0:27:27 So, you know, maybe feeding all of the information off of Twitter
0:27:29 isn’t that great of an idea after all.
0:27:32 But that is one of the sort of benefits of Grock as well.
0:27:34 Is that Twitter?
0:27:36 That’s fascinating. I’ve never even thought about that.
0:27:39 So you you can like make up memes and then Grock thinks it’s real
0:27:41 then actually get spread to real news.
0:27:44 It reminds me of I live in San Francisco.
0:27:47 And there’s a lot of these like shenanigans that happen with self driving cars.
0:27:50 So like people people will take like a traffic cone
0:27:53 and like put it on the sensor and prevent it from moving.
0:27:57 Yeah, a more violent one, unfortunately, is I think it was like two months ago.
0:28:01 People literally blew up a self driving car Chinatown in Chinatown here.
0:28:02 Yeah, yeah, in Chinatown.
0:28:04 And it kind of feels like this, you know what I mean?
0:28:08 Like I could totally see if Grock ends up being very good, widely used
0:28:12 and it’s very well publicized that pulls from Twitter that there’s constant
0:28:15 like meme campaigns to throw it off, you know what I mean?
0:28:17 And just to get it to say whatever.
0:28:20 So I don’t know, like never underestimate the Internet, you know what I mean?
0:28:23 Like there are going to be shenanigans abound with this stuff.
0:28:29 Why do you think there’s such an aversion, a hatred towards the self driving cars?
0:28:33 Is it like the people that are Uber drivers and Lyft drivers feeling threatened?
0:28:36 Is it people that are just anti AI?
0:28:38 Is there like any feeling of why that’s happening?
0:28:45 My feeling is if you take the same person and put them in any other city
0:28:48 that’s not San Francisco and has this exposure to tech whatsoever.
0:28:52 The arrival of self driving cars is like magic, you know what I mean?
0:28:58 Like when when I first got access to Waymo, for example, in its beta
0:29:01 and I hopped in, everyone immediately around me was like, oh my gosh, what is that?
0:29:03 Like that is so cool.
0:29:05 They were taking pictures and everything.
0:29:10 And I think there’s just something about how much experimentation there is
0:29:14 in San Francisco that for some reason, the the new fangled thing,
0:29:16 like having a new fangled thing is just baseline.
0:29:21 And in some respects, you get desensitized to just how cool it can be.
0:29:26 And as a result, I think you just get tired of just how much change there is
0:29:31 or like you feel this narrative of like big tech kind of invading you
0:29:34 because you just have to deal with this change all the time.
0:29:38 And it happened when there were these delivery robots that I think it was
0:29:42 DoorDash, I’m not too sure it was these food delivery companies that were piloting,
0:29:45 you know, sending these little rovers around to go pick up your food.
0:29:47 And then you wouldn’t need a person anymore.
0:29:48 You can just like have this whole thing.
0:29:52 And then all of a sudden, like, you know, pedestrian safety became an issue, right?
0:29:56 It’s like, oh, what happens if someone trips and like falls over one of these things?
0:29:58 And these are like good concerns.
0:30:00 I think there’s like the right questions to ask.
0:30:03 But just like this overall feeling of just like the first instinct you have
0:30:07 when you receive something new and innovative, I think you get reminded
0:30:11 a little bit of all the other times where it’s failed or it’s been annoying
0:30:15 or been a nuisance, the scooter wars were a very big example like line scooters,
0:30:19 bird scooters, those invaded San Francisco, polluted the sidewalks.
0:30:21 I think that was a little bit too much by the end of it.
0:30:25 But you kind of just get reminded of that kind of era a little bit.
0:30:26 I think that’s that’s where that comes from.
0:30:30 When you’re there, like it’s so expensive to live there in San Francisco.
0:30:34 And the people who are in tech are making crazy amounts of money.
0:30:39 And the people who are not, they’re barely scraping by in San Francisco on average.
0:30:42 And so when I was there, like when I first moved there, it wasn’t that bad.
0:30:46 But it seemed like every year there was more and more hate towards people in tech in general.
0:30:50 And I think a big, a big part of it is like that difference of income and things like that.
0:30:52 You would even see it like in like tech journalists.
0:30:54 So like a tech journalist are not liking people in tech.
0:30:57 And it was like, become this really odd dynamic there where it’s like,
0:31:00 you got tech entrepreneurs and people are writing about them who don’t like them.
0:31:02 It’s like, this is very odd.
0:31:07 I feel like a lot of journalists seem to hate like tech AI specifically.
0:31:11 If you see anything in like a big media platform these days,
0:31:18 it usually is taking a more negative towards the AI tech than it is the positive slant, right?
0:31:23 So like people like us who are creating content that’s more on that positive spin
0:31:27 of where all of this is going, we seem to be having our heyday right now
0:31:29 where people are paying attention to what we’re talking about
0:31:31 and they don’t trust the big media anymore.
0:31:36 So I mean, I think there is a path to be optimistic and still get views
0:31:38 and still get clicks.
0:31:43 I just think the media is so used to this like old way of doing it of like drama cells, you know?
0:31:49 I do think with some of these conversations, as much as optimists as we are,
0:31:52 I do think it’s important to be sympathetic to some of these points of criticism.
0:31:58 For example, this week, along with all the LLM stuff that’s been going on,
0:32:01 music is like on top of everyone’s minds, right?
0:32:04 So I mean, today specifically, there’s the audio release.
0:32:06 There’s like another one coming out of YC called Sonato.
0:32:08 I think it was last week.
0:32:12 You had 200 of the top artists sign a letter saying, hey, look,
0:32:15 like we generally think AI is good for creative work
0:32:18 and it can be built in a way that supports creative work.
0:32:23 However, we like strongly, strongly, strongly encourage that tech companies, investors, et cetera,
0:32:28 the entire ecosystem working on AI do not build AI that replaces
0:32:33 or diminishes humans or their compensation that are in the industry, right?
0:32:38 Now, the exact achievement of that, like how to actually strike that balance
0:32:40 is like very, very hard in my opinion.
0:32:43 Like there’s going to be coming and going like with all of this, right?
0:32:50 But if you were an artist on that petition, signing that letter
0:32:56 and your latest understanding was Suno or one of these other tools,
0:32:57 frankly, I don’t even think it was Suno.
0:33:01 Frankly, like a lot of these musicians or a lot of people outside of our bubble
0:33:03 aren’t even exposed to this.
0:33:07 And it’s like only truly the early tinkerers on your Discord servers,
0:33:09 et cetera, that are really playing around with this.
0:33:14 Like you can’t be happy seeing the Sonato from today or Udio today,
0:33:18 which are even steps ahead of Suno, right?
0:33:20 You can’t, I don’t think that that makes it happy.
0:33:25 Even if from Silicon Valley’s perspective, it does not like,
0:33:28 I think everyone here is sort of aligned on this general idea of like, OK,
0:33:32 well, like the intent here is to create tools that promote creativity,
0:33:35 that democratize access, that democratize output.
0:33:36 That all sounds great, right?
0:33:40 But if you are in the industry and you’re watching all these developments
0:33:45 happening, I think it’s very fair to assume that like that person is is just
0:33:47 like they’re coming for us, you know what I mean?
0:33:49 Like that’s sort of the mindset.
0:33:52 And it’s like, like if this is happening this quickly and in a matter of a couple
0:33:56 of months, the models have gone this far, then what’s next?
0:34:00 Like the only thing I see left remaining is for this to fully replace me
0:34:02 or or outright replace my job.
0:34:03 That can’t feel good, right?
0:34:08 And I’m very sympathetic to that because to I feel like to brush that
0:34:11 conversation, that narrative under the rug, I think is missing the point.
0:34:16 I think that we’re going to have a much harder time convincing people to one,
0:34:21 use these tools and then to convince them that the proliferation of these tools
0:34:23 is net good for society.
0:34:26 We don’t at least acknowledge that sort of thinking, right?
0:34:28 So I think it’s it’s important to be nuanced in these things.
0:34:32 It’s just there are a lot of when I notice the animation communities,
0:34:36 the gaming communities, the music communities that are all deeply creative
0:34:40 people and there is a lot of anti AI sentiment, right?
0:34:43 I think like all three of us have probably been subject to some amount
0:34:47 of like pseudo cancellation on Twitter or YouTube comments or whatever it is.
0:34:51 As a result of this, Matt, maybe, I don’t know, you’ve been you’ve been pretty
0:34:53 your comment section always looks pretty positive.
0:34:55 You’re not seeing it all then.
0:35:00 I do think like, you know, like we have to acknowledge it, right?
0:35:03 And it’s not just a matter of like, oh, these people like they’re just
0:35:04 going to get left behind.
0:35:06 It’s like, no, like, I think these are pretty legitimate concerns.
0:35:11 And we just have to figure out a way economically, legally and technologically
0:35:15 that we can sort of just have everyone come along.
0:35:16 It’s not going to be perfect, right?
0:35:18 But I feel like there’s got to be a solution there.
0:35:21 I’m an optimist about that, at least that there is sort of a way
0:35:23 forward that that kind of fits everyone in the future.
0:35:24 That’s fair.
0:35:26 Like, I think this will, you know, there will be some jobs that will
0:35:27 dramatically change.
0:35:29 Some will disappear, new ones will appear.
0:35:32 Most jobs will change in dramatic ways.
0:35:36 And there will be some aspect of like, OK, how do you make sure the artists
0:35:37 get compensated?
0:35:38 And and it’s tricky.
0:35:40 I think there is going to be a big legal battle, right?
0:35:44 Because like, I’m sure OpenAI has trained on so much data that maybe you
0:35:46 would say they shouldn’t have been able to do.
0:35:51 And I don’t think they would have done that unless they had talked to lots
0:35:54 of lawyers and decided that, OK, this is actually a gray area.
0:35:58 Because when you’re training on this data, since this is such a new thing,
0:36:01 you can say that this is basically the same way that a musician would
0:36:03 learn from listening to another musician.
0:36:05 Did that person steal that person’s work?
0:36:09 You know, it’s like, you know, and so I think that’s really the argument
0:36:12 they they plan to make and they just not wanting to publicly state that yet.
0:36:16 But I do believe that is the argument that legal argument is coming,
0:36:18 because it sounds like they’ve actually scraped YouTube’s videos
0:36:22 possibly and that’s how they’ve, you know, built up their video.
0:36:24 Yeah, I mean, I take that same approach.
0:36:28 I try to be very empathetic, sympathetic to those concerns.
0:36:31 I try to take them all into account when I am sharing the news.
0:36:34 I always try to kind of talk about both sides of the equation.
0:36:38 You know, I really do feel for a lot of these people.
0:36:41 Like when it comes to music, I don’t think it should be OK
0:36:44 to just go and create a song in the style of Drake.
0:36:47 It sounds like the beatstrake would make sounds like his style,
0:36:51 sounds like his voice and you just took him out of the loop completely.
0:36:52 I don’t think that’s cool.
0:36:53 Honestly, I don’t.
0:36:55 I don’t think people should be allowed to do that.
0:36:58 But saying that there is a very, very fine line.
0:37:02 What if somebody has a voice that sounds very similar to Drake?
0:37:05 Are they not allowed to make music that sounds similar to Drake?
0:37:09 Like there’s some weird, like nuances, gray areas there
0:37:12 that are really, really hard to deal with.
0:37:14 But I think at the end of the day, a lot of it comes down
0:37:18 to the sort of monetary aspect of it, right?
0:37:21 If somebody is like who dedicated their life’s work to becoming
0:37:25 the best musician that they can possibly be, the best game developer
0:37:28 they can possibly be, the best artist they can possibly be.
0:37:32 And now people can create stuff at the level that they were able to create it.
0:37:36 Well, how do they sort of still make money?
0:37:38 How do they still have a livelihood?
0:37:40 Like it all comes back to the monetization.
0:37:44 And I think something needs to happen in the AI world
0:37:48 where there is some sort of like Spotify style split.
0:37:50 It is insanely complicated to figure out.
0:37:53 And I do not envy the people that have to figure that out.
0:37:56 I just get to be here on the sidelines, armshare, quarterbacking.
0:38:00 But, you know, I do think there needs to be some sort of way
0:38:05 to compensate the people who maybe with the data was trained on.
0:38:09 I think if I was a musician right now and I, you know, you could see my background.
0:38:12 I’ve got guitars. I’ve played in bands before I’ve recorded.
0:38:13 I’ve gone on tour.
0:38:16 I’ve done the whole music thing in the past.
0:38:19 But if I still played music as often as I used to,
0:38:22 I would actually be really excited about what’s going on because I can be like,
0:38:26 man, if my voice is popular, I can do the Grimes thing and license it
0:38:30 and have a million people going out there and creating music on my behalf.
0:38:35 And I can somehow earn from those people using my voice and using my style of music.
0:38:39 To me, that gets me excited, but I also understand the flip side of that coin.
0:38:42 I think that has to be coming because, I mean, Will, I am invested in Udio.
0:38:44 That was announced like two hours ago.
0:38:49 And and so, you know, alongside Andrews and Horowitz and a bunch of other great investors.
0:38:52 And I don’t think he would be investing if they didn’t have some kind of plan
0:38:55 for revenue share with artists. I could be wrong, but I think I think the fact
0:38:59 that he’s involved shows that there’s probably they’re probably not just
0:39:00 going to be saying F you to all the artists.
0:39:02 I think there’s going to be some kind of plan to compensate people.
0:39:06 Yeah, well, if you look at Udio and you look at Suno, you can’t go into either
0:39:11 one of those apps and say, generate a song for me that sounds like Katy Perry.
0:39:13 It’ll just say, sorry, can’t do that.
0:39:18 That other one, Sonato, I actually did a live stream and I was playing with it
0:39:20 and just out of curiosity on the live stream, I was like, generate a song
0:39:23 that sounds like Blink 182. And it’s like, here you go.
0:39:28 We use the lyric, the voice style of Mark Hoppus and the guitar style of Tom DeLong
0:39:30 to make this sound like a Blink 182 song.
0:39:35 So Sonato actually does let you plug in a band’s name and generate a song
0:39:37 of like wherever you want.
0:39:40 I was feeling that feature will probably get phased out.
0:39:44 I don’t know. I think they need to nail that because that’s like consumer wise.
0:39:45 That’s where you’d actually get traction.
0:39:49 Like if you can’t say artist name, you’re not you’re not actually like a major,
0:39:52 you know, mass adoption of these tools, I don’t think.
0:39:54 But consumer wise, that’s where it’s at.
0:39:57 Legal wise and not wanting to get sued up
0:40:01 as by every music industry person in the world.
0:40:03 I don’t know if that’s the approach you want to take. I don’t know.
0:40:07 Yeah. What what is your guys’s take on
0:40:10 Udio and Suno? How do they sort of compare to each other?
0:40:12 I’ve been playing around with a couple of them.
0:40:14 I think Udio is very clearly the best one.
0:40:17 But what have your experimentations led you to?
0:40:18 Personally, I’m just having fun with them.
0:40:21 Like I love them. I think they’re both so much fun.
0:40:25 I’ve made probably like 20 potential jingles for future tools, right?
0:40:29 Like write me a jingle about why future tools is the place to go
0:40:33 to learn about AI and the latest tools and it makes like really fun songs.
0:40:35 And I work them into my YouTube videos.
0:40:37 For me, that it’s just kind of like a fun toy.
0:40:41 I don’t really see a whole lot of practical applications for it yet.
0:40:43 But I do think they’re really, really fun to play with.
0:40:48 Yeah. Right. I think Udio probably overall is better for the most part.
0:40:52 But I did give it a prompt like generate some like electronic dance music for me.
0:40:58 And sooner did that style of music better than than Udio did.
0:41:02 So I think, you know, genre dependent, but I think as a whole, Udio is probably better.
0:41:06 I think the the overall direction that I’m excited in is one for
0:41:08 for you to refine what you just generated, right?
0:41:13 And so instead of having this be like a one shot sort of it just generates.
0:41:14 And that’s pretty much all you get.
0:41:17 And you just have to keep on trying over and over again to go in
0:41:18 and change the little aspects of it.
0:41:22 Maybe there’s like individual parts of it that you can you can tease out
0:41:24 the little the individual stems and edit them.
0:41:30 And I think that is generally the direction that I also think is helpful
0:41:35 to turn this into more of like a creative tool, more than an outright
0:41:38 sort of prompt to full song type of thing.
0:41:41 Because Matt, to your point, like I am a little bit questioning the use case
0:41:45 beyond doing your your jingles for YouTube videos.
0:41:49 I’ve also done this for for jingle purposes and like jingles aside.
0:41:51 Like jingles are not a big market, you know what I mean?
0:41:54 Like they’re not going to like make all these companies
0:41:57 enough revenue to compete over the long term.
0:42:00 However, if you’re to build more tooling around it, which is like,
0:42:04 I think this is maybe Sonato’s approach here, which is can we allow
0:42:08 a little bit more directed iterations over time?
0:42:09 That is interesting, right?
0:42:14 Like that that now is opening up a new canvas for for new artists
0:42:19 to explore and creatives to really get in there and master the individual pieces.
0:42:22 And that can interest a new way of creating music.
0:42:24 So generally excited about it.
0:42:28 But yeah, I’m very curious to see once we’ve solved that sort of like
0:42:33 can it can create a 30 second clip of like pretty coherent, pretty good
0:42:36 music end to end, then it’s like then what right?
0:42:38 Like then then what levels you go to?
0:42:41 Which direction do you head in research wise?
0:42:46 They saw a Twitter post ex post earlier today of somebody saying
0:42:51 like I’ll be impressed with music when an AI song tops the charts or something
0:42:53 like that. Is that like the goal of some of these tools?
0:42:57 Are they trying to like get you to make songs that’s going to be top of the charts?
0:43:01 Like, I don’t know, as a human myself, maybe you guys can relate.
0:43:05 But as a human myself, I enjoy listening to music created by other humans.
0:43:10 Like there is something about listening to like a Led Zeppelin song
0:43:14 and knowing that like Jimmy Page played those notes on his guitar.
0:43:19 And that was a very complex thing that he played that makes it that much
0:43:23 more pleasing on the ears, just knowing the human element was attached to it.
0:43:27 I don’t feel that sense of accomplishment when I listen to an AI song
0:43:31 that I generated, right? I don’t really like I might go share it on Twitter
0:43:32 because I think it’s funny.
0:43:37 But I’m not going like this is the next, you know, dance club hit or anything.
0:43:41 I feel like the AI music right now, what it’s best for like a real
0:43:44 world use case is actually take the lyrics out of it completely
0:43:48 and use it as like a sound bed for your your videos or for your podcast
0:43:50 or things like that.
0:43:53 But once you add the lyrics into it and you’re trying to make your own
0:43:58 like pop songs with singing, I actually don’t know the real use case for it right now.
0:44:01 Was it musically that started off as and then it turned into TikTok?
0:44:02 Or maybe I’m good. Yeah.
0:44:05 And so like, you know, I think you could have some kind of new music app
0:44:07 where like, yeah, if somebody could just generate any kind of song
0:44:11 as part of some kind of social app and then and then that thing could go viral
0:44:14 and you know, and that could become a top chart things like some some songs
0:44:17 from TikTok have like become very popular because they were big on TikTok.
0:44:20 Right. Once again, I do think it’s sort of genre dependent.
0:44:23 What I get excited for is do you guys remember that clip of Maggie Rogers
0:44:27 when she was still in school and Pharrell came and listened.
0:44:30 It was like a music production class of some sort.
0:44:31 And it was like she blew up.
0:44:35 Like that’s exactly how she blew up was off of a video recording
0:44:39 of Pharrell listening to her music that she produced all by herself.
0:44:42 She’s saying just everything was hers, right?
0:44:44 And I think Pharrell’s commentary was just like, I have no notes
0:44:48 because this is so authentically you that like I can’t come in
0:44:54 and sort of give you any sort of advisor criticism or editing off of this.
0:45:00 And what I think about is if you took Maggie Rogers then
0:45:05 and gave her a set of new, like these AI powered tools,
0:45:09 how much more powerful of that experience would that experience have been?
0:45:11 Right. And it’s not about to say like, oh, Maggie,
0:45:15 you only give us the lyrics and you sing and sort of AI will do everything
0:45:18 around you and sort of little one click sort of like spinning into different genres.
0:45:22 There are probably a lot of smaller technical things
0:45:26 that she probably did not do in the first pass, didn’t know how
0:45:30 or just didn’t see it creatively or whatever it is because she’s young
0:45:34 and sort of early on in your music career, getting you that first draft of it
0:45:36 sparks the creativity, if anything, right?
0:45:40 And you can you can have her be what she was in that moment,
0:45:43 which was the vibe curator of the song, right?
0:45:46 Like she brought the spirit of what was the message
0:45:49 that she wanted to send the emotional state that she wanted to conjure.
0:45:51 But the rest of the technical pieces,
0:45:55 maybe there’s help that we can give the next Maggie Rogers, right?
0:46:00 That would be cool to me to your point around the pop factory songs.
0:46:00 You know what I mean?
0:46:04 Like these are all the same set of musical artists that are writing songs for each other, right?
0:46:08 They will literally sit in a room and be like, OK, today we’re writing for Shawn Mendes.
0:46:10 And like, you know, eight of them will get together
0:46:12 and just sort of all throw songs together.
0:46:15 One of them will hit and it will become Shawn Mendes’s next song, right?
0:46:18 There’s something about that that feels a little bit different, perhaps.
0:46:22 And, you know, it is a much more mechanical process, I suppose.
0:46:27 But when I think about the essence, like the spirit of like what we’re trying
0:46:30 to do with these creative tools, or at least the companies doing them,
0:46:33 I would hope is to find that next student, right?
0:46:38 And to give them, you know, that much an easier of a time to express the spirit
0:46:42 and the concept of what they’re doing, that feels very powerful to me.
0:46:46 I do actually have like a couple like sort of last questions I want to ask.
0:46:50 One of them is, do you have any advice for business owners
0:46:55 in like the current landscape of AI with the existing suite of tools that are available?
0:47:00 What do you tell business owners who maybe are hesitant to get into using AI?
0:47:02 Where where would you send them first?
0:47:08 If I had one line, one line, it’s to give your entire business
0:47:13 access to one of the paid plans of chat between Claude and just let him figure it out.
0:47:16 Honestly, it’s $20 per user per month.
0:47:20 For most businesses, this is more than affordable, right?
0:47:24 Like just one month, 30 days and be serious about it, right?
0:47:28 Like give everyone access and tell them I want you to figure out where it can be more effective.
0:47:34 Arm the people with a mandate to go figure it out because people will, again,
0:47:35 given the nature of the product, right?
0:47:40 Again, we’re focusing a lot of the benefit comes from the bits and pieces
0:47:41 of all these different processes.
0:47:44 Who knows them better than your employees that are dealing
0:47:46 with these processes every single day, right?
0:47:50 This is not a top down thing where we’re saying, OK, we are now going to apply
0:47:52 AI in like our marketing department.
0:47:56 It’s like, no, there are all these like little things that there are enough
0:47:59 early adopters at your company that will take the time and experiment
0:48:03 and go figure out if you let them and give them the tools and the support
0:48:05 from leadership to do it, right?
0:48:09 Host the hackathon, encourage sharing, reward the behavior, right?
0:48:15 So typically a lot of the patterns that I’ve seen is there’s just like one person
0:48:19 in this one department that spent their off hours, one or two hours every now
0:48:24 and then playing around with a prompt or like a Zapier integration or something
0:48:29 like that, and all of a sudden it’s saving the entire department 100 hours
0:48:32 a week because of how many people they scaled the entire operation to.
0:48:36 Maybe again, something as simple as everyone here.
0:48:40 I sent the I sent the prompt over email, copy and paste the prompt from the email.
0:48:43 Go to chat to your cloud or whatever it is and then use this when you’re doing
0:48:44 this thing, right?
0:48:49 That’s already impactful and it had nothing to do with hiring an outside
0:48:51 consultant or some other vendor or whatever it is.
0:48:55 It was all just harnessing the, the inherent sort of interest
0:49:00 and attention there is around AI and grasping onto those early adopters
0:49:03 who would like to figure it out and giving them the environment for them
0:49:04 to do it for you, right?
0:49:05 So that’s the main thing.
0:49:08 You’ll be surprised at how many businesses have not taken this.
0:49:11 Like I’ve talked to a few enterprises that like from the outside looking in,
0:49:15 they really should have had some movement on this in the last year or so.
0:49:16 And they just haven’t touched it at all.
0:49:19 Not one bit because they’re, they’re just like, okay, this is cool.
0:49:19 We get it.
0:49:21 But like, ah, like we’re fine.
0:49:21 You know what I mean?
0:49:23 It’s like, no, take it seriously, right?
0:49:27 Like go through the exercise, get it in people’s hands and it’ll come.
0:49:28 And it’ll come.
0:49:30 So if, again, that’s, that’s the one thing.
0:49:33 There’s a whole bunch of other ways you could do it, but like, this is
0:49:34 the most dead simple one.
0:49:37 You already have the people and the resources just given them the environment.
0:49:40 Thank you so much, Pete, for, for hanging out with us today.
0:49:40 That’s awesome.
0:49:43 You are non-mer Pete over on X.
0:49:46 The newsletter is the neuron.
0:49:51 Where should people go to go learn more from you and to, you know, get
0:49:52 inside of your ecosystem?
0:49:52 Yes.
0:49:54 Find me on Twitter, nonbear Pete.
0:49:57 Go to the neuron daily.com for our newsletter.
0:49:59 And then find me, Pete Huang on LinkedIn.
0:50:00 Thanks again for joining us.
0:50:01 This has been such a fun conversation.
0:50:02 Thanks guys for having me.
0:50:03 This is so much fun.
0:50:13 [Music]
0:50:23 [Music]