Learning Library

← Back to Library

AI Hype, Market Slump, Skepticism

Key Points

  • The panel unanimously rejected the notion that AI companies are responsible for the recent downturn in the U.S. economy, viewing AI as a “cherry on top” rather than a macro‑economic driver.
  • Recent market volatility was discussed, with participants attributing the swings more to traditional factors (e.g., Fed policy, exotic financial positions) than to hype surrounding AI investments.
  • The conversation highlighted the cyclical nature of AI hype and Wall Street’s rapid swings, stressing that sustainable value hinges on clear differentiators and strong moats—not just hype.
  • The acquisition of Character AI was examined as a case study, prompting a reminder that firms must understand their unique value proposition and guard against competitors replicating their offerings.
  • AI governance and responsible deployment were identified as crucial themes, with experts emphasizing the need for structured research, ethical oversight, and long‑term thinking amid the current excitement.

Full Transcript

# AI Hype, Market Slump, Skepticism **Source:** [https://www.youtube.com/watch?v=1tTXTDzaPyY](https://www.youtube.com/watch?v=1tTXTDzaPyY) **Duration:** 00:32:04 ## Summary - The panel unanimously rejected the notion that AI companies are responsible for the recent downturn in the U.S. economy, viewing AI as a “cherry on top” rather than a macro‑economic driver. - Recent market volatility was discussed, with participants attributing the swings more to traditional factors (e.g., Fed policy, exotic financial positions) than to hype surrounding AI investments. - The conversation highlighted the cyclical nature of AI hype and Wall Street’s rapid swings, stressing that sustainable value hinges on clear differentiators and strong moats—not just hype. - The acquisition of Character AI was examined as a case study, prompting a reminder that firms must understand their unique value proposition and guard against competitors replicating their offerings. - AI governance and responsible deployment were identified as crucial themes, with experts emphasizing the need for structured research, ethical oversight, and long‑term thinking amid the current excitement. ## Sections - [00:00:00](https://www.youtube.com/watch?v=1tTXTDzaPyY&t=0s) **AI Hype, Mergers, and Economy** - A panel of technologists debates Wall Street hype, the strategic merit of acquiring Character AI, and whether AI companies could threaten the U.S. economy, concluding they likely won’t. ## Full Transcript
0:00Wall Street Spooks pretty easily and 0:02Hypes pretty easily and they're also on 0:05a cycle that research certainly is not 0:08structured outputs probably the most 0:10sexy release of this summer you're kind 0:12of breaking this fcking Bronco that uh 0:15just came out of the blue does the 0:17acquisition of character AI make any 0:18sense at all you have to know what's 0:20your value ad and how much of that is a 0:23differentiator with high Mo so others 0:25can't just come in and do what you do 0:27all that and more on today's episode of 0:29mixture of 0:35experts I'm Tim Hong and I'm joined 0:38today as I am every Friday by a genius 0:40panel of technologists Engineers and 0:42more to help make sense of another 0:43hectic week in AI land on the panel 0:46today we've got three guests Marina 0:47danki as a senior research scientist 0:50Kush vne IBM fellow working on issues 0:52surrounding AI governance and chit vary 0:54senior partner Consulting on AI for US 0:57Canada and Latin America 0:59[Music] 1:04all right so uh let's just get into it 1:06first story of the week uh is a big one 1:09um but I want to start with kind of a 1:10round the horn question let's just start 1:12with a quick yes or no and it's a very 1:14simple question to kind of kick off to 1:16discussion which is are AI companies 1:18going to bring down the American economy 1:20uh Kush yes or no what do you think uh 1:23no uh no and Marina no okay we have 1:30uniform skepticism at that position and 1:33I think that's actually what I wanted to 1:34get into so uh if you've been keeping 1:36your eyes on the financial news this 1:38week markets were massively down um 1:41across the board 1:42internationally and uh there was a lot 1:44of speculation as to why this was the 1:47case um people were proposing you know 1:49the unwinding of exotic Financial 1:51positions uh concerns about the FED not 1:53cutting rates but one thing that a 1:55number of people kind of argued was 1:57should we blame AI like the hyper AI for 2:00this um and part of this claim was based 2:02around the idea that the companies 2:04really leading the downturn um and 2:06arguably a big drag on you know indexes 2:09like the S&P 500 were tech companies 2:11that have made big bets on AI in the 2:13last 12 to 24 months and so want to get 2:16the kind of panel's opinion and Kush 2:18maybe we'll toss it over to you first is 2:21you know do we buy this as a theory like 2:23why should we or shouldn't we believe 2:24that AI is kind of a contributor to this 2:26downturn um and and is kind of a a 2:28popping or at least kind of increasing 2:30skepticism around AI um having these big 2:33macro effects I'm curious why you said 2:34no in the the first question there yeah 2:37I mean uh there's clearly I mean hype 2:40Cycles with everything uh but I think 2:43the economy has a lot more to offer I 2:46mean it's a it's a very broad-based sort 2:47of thing AI is kind of the the cherry on 2:50top or the the icing on the cake so um 2:53uh I mean yes it affects uh perception 2:57uh and less uh 3:00I mean of the view that it uh is really 3:03about the fundamentals at this point I 3:05think that'll change over time but but 3:06not right now well I think there's been 3:08I mean part of this I think is also 3:10following on the tales of I think we've 3:12been talking about it for the last few 3:14episodes is kind of these reports coming 3:15out of Banks and other uh you know 3:17Financial firms kind of raising some 3:19skepticism around kind of like the 3:21excitement around AI so you know there's 3:23the Goldman Sachs one that we talked 3:24about a few weeks back and also the sequ 3:26report that some people might have seen 3:29um it is true though that the tech 3:30companies have made genuinely a really 3:33big bet on the market for AI um and I 3:36guess I'm kind of curious you know maybe 3:37show bit I'll throw it to you is you 3:39know are you seeing you know clients 3:41kind of following those Jitters are they 3:42reading these reports and saying well 3:44you know maybe AI is not providing you 3:46know what we thought it would you know 3:47should we be a little bit more cautious 3:49about how we make these Investments so I 3:51don't think the clients have to and I'm 3:52talking about the 1400 companies 500 3:55they don't have to read these reports to 3:56realize that certain U areas AI has been 4:00over prised in certain areas they have 4:01they are under utilized right so there's 4:03absolutely no confusion about the fact 4:05that AI is going to have is having a 4:07seismic impact on the businesses going 4:09forward so no CEO can can say that the 4:12next 5 years are not going to be uh 4:14massively impacted by what AI can do 4:16it's a question of how do you applying 4:18AI surgically in the processes and how 4:20do you think about a strategy of data 4:22that then leads to an AI strategy that 4:24then delivers value for you right so the 4:27conversation has changed more and all 4:29right after experimenting for 2 years we 4:31have a good sense of where Ai and geni 4:34are working well we now need to make 4:36sure that we have a good mechanism to 4:38figure out the high value unlocks in the 4:40business appreciate that it's a it's a 4:42combination of AI Automation and 4:44generative AI it's not all gen handling 4:47the entire process into end and we need 4:49to make sure that our data real estate 4:50and the people and and the processes are 4:52aligned to unlock that value right so I 4:54think there's a significant appreciation 4:56of the value it can bring but also the 5:00fact that it's a journey and you need to 5:01do you need to do make steps along the 5:03way to make sure that you're getting 5:05that value unlocked that's very clear to 5:07all my for 100 500 clients yeah I think 5:09that's maybe one thing that our 5:11listeners would benefit a lot from your 5:12expertise on show bit is I think you 5:14kind of put out the idea that there's 5:15like underhyped areas of AI right um and 5:19I'm kind of curious if there's when you 5:20say that if you've got kind of 5:21particular areas in mind where you're 5:23like this is where businesses aren't 5:24looking right like you know there's a 5:26lot of hype in the space but like this 5:28this seems to be where some of the 5:29Hidden are I'm curious if you can speak 5:31to that a little bit yes I think it's a 5:32it's a mundane tasks it's the stuff that 5:35uh how do you how do you make sure that 5:37every employee across organization can 5:40experiment in their day-to-day workflows 5:42with AI with generative AI in a very 5:44secure and governed way so within IBM 5:47Consulting for example we have 160,000 5:49Consultants who wake up in the morning 5:51and doing all kinds of varied tasks 5:53there are a small subset of people who 5:55are AI gurus right they are they feel 5:58that if it's been 20 minutes since llama 6:003.1 landed and we have not had it 6:02running locally you are an embarrassment 6:04to society right there's a small portion 6:05of those but 85% of the other uh part of 6:09Consulting they're doing things like I'm 6:11going to do code creation I'm a tester 6:13for the last 11 years I've been doing 6:15marketing campaigns I'm going to do 6:17Finance workflow so I'm going to get a 6:19invoice I'm going to marry it against 6:21the contract the the purchase order and 6:23I'm going to approve it or disapprove it 6:25right so those kind of mundane workflows 6:28have a human in the loop and you need to 6:30figure out how Excel got embedded in 6:33those workflows you're now at the point 6:34where you're having AI generative AI get 6:36embedded everybody figured out how to 6:38use Excel to improve their day-to-day 6:40workflows right we're at that same point 6:42today so you need to get to a point 6:44where every IBM consultant we call that 6:46Consulting assistant as an example it 6:48could be a co-pilot from Asher could be 6:50Amazon Q's Google of the world but you 6:53need to democratize people actually 6:55messing with their day-to-day figure out 6:57that oh this email that I write ,00 7:00times a month can be automated and 7:02that's the value unlock get get CL get 7:04your end customers to start your end 7:07employees to start experimenting in a 7:09governed way so that kush doesn't have a 7:10heart attack just make sure doing this 7:12in a way that we don't get ourselves 7:13into trouble um yeah I think that's 7:16actually in some ways like it has ended 7:18up being what you're describing show bit 7:19the kind of like 800 pound gorilla of 7:22the AI world you know I love this kind 7:24of joke that like you start open AI 7:25because you really want to create like 7:26AGI but like just slowly but surely the 7:29gravity ational well of being like B2B 7:31SAS uh and offering that as a service is 7:33like really where the gigantic amount of 7:35money uh is um Marine I did have a 7:38question for you based on kind of what 7:39sha just talked about here I know you 7:42know sha you kind of made the difference 7:43between like people saying okay if you 7:44can't Implement Lama 3 on day one and 7:47revolutionize all your business you know 7:48processes in the first day you're a 7:50waste of society um I'm kind of curious 7:53so there's one standout company here 7:55which is NVIDIA uh which is Hardware um 7:58and that is a company that has been hit 7:59in the stock market rather hard um and I 8:01think was one of the examples that 8:02people said see this is why AI is hyped 8:05um do you do you buy that I mean is 8:07NVIDIA indeed the most valuable company 8:09uh in the whole world uh and you know 8:11how should we think about sort of 8:13Hardware in this picture like will 8:14Hardware continue to be sort of the most 8:16valuable kind of piece of this AI Pi at 8:19least as far as like the stock market is 8:20concerned I mean it's a dependency so 8:23just talking of pure engineering terms 8:24you are pretty uh much tied to it 8:26because it's very much a dependency I 8:28will say as far far as Nvidia going up 8:30in value crashing in value um Wall 8:32Street Spooks pretty easily and Hypes 8:35pretty easily and they're also on a 8:37cycle that research certainly is not 8:40they want to know all right q1 what do 8:42you got Q2 what do you got Q3 what do 8:43you got it's not the rate at which 8:45research actually happens so when you 8:47have preliminary results Wall Street 8:49will get over excited and then the 8:51results next time are not as good and 8:52then they get over depressed and we 8:55actually have the same thing in research 8:56where I'm like I can't guarantee you 8:58that the research breakthroughs are 9:00going to happen in three months on the 9:01dot this is not how this you got to 9:03deliver your Q2 breakthrough Marina 9:05right I can't I can't promise my Q2 9:07breakthrough so um I would also say that 9:10this is also to some extent uh a 9:13mismatch between the schedule of Wall 9:15Street in the schedule of research in a 9:17in a new area in an area that we don't 9:18yet understand very well and that's I 9:20think a lot of what we're actually 9:21seeing here yeah that's fascinating is 9:24almost you're saying like we should not 9:26be looking to the stock market to judge 9:28the value of the AI space in part 9:30because like the market doesn't know how 9:32to Value it at the moment is kind of 9:33what you're saying is like I don't think 9:35it's very clear yet I I don't know if K 9:37but you disagree but I actually don't 9:38think we actually know very well yet how 9:39to Value AI properly I'm I'm with Marina 9:42on this uh and I I don't think that the 9:45common uh stock investor understands the 9:47impact especially in Enterprise space 9:49and what can do for people so we've been 9:52we've been just dunking on AI stocks 9:53saying that hey you are leading to the 9:55downfall of economy we look at the 9:57positive that it has done right it's 9:59also contributing insanely towards the 10:01overall economy right so you should give 10:03AI enough credit to lift the entire 10:05stock market up as well not just the 10:07training week and say hey U the the 10:10market is down X points because of the 10:12large Nvidia swing which like just look 10:15at the world we live in right in the 10:17last few months Nvidia has swung a 10:21trillion dollars in market cap a 10:24trillion this just pause and realize how 10:26much of an impact that's having on on 10:28people right it is people reacting to oh 10:31my God I don't want to miss 10:33out but also not knowing at what point 10:36are you investing in the fundamentals or 10:38are you pulling out of the stock too 10:39early right like even massive companies 10:42like Arc invest ended up missing the 10:45boat on 10:46Nvidia lost a billion dollars of 10:48opportunity there right so you need to 10:50understand the fundamentals and stay 10:52long in the market versus going and 10:55reacting to these quarterly ups and 10:57downs 10:59I don't know I'm you're not arguing this 11:01but I think it's almost like you could 11:02make out the argument that like you know 11:03what the biggest meme stock in the whole 11:05world is it's Nvidia you know it's it's 11:08not it's not GameStop it's not anything 11:10like that I was just gonna agree with 11:11Marina I mean uh the fact is that uh and 11:16what chth was saying as well I mean this 11:17is a it's a long game uh we don't really 11:20know how to Value things uh yet uh it's 11:24not like some commodity where you can 11:26like grab it and hold on to it and see 11:28what it's doing so um uh I think we'll 11:31we'll get better um just like we've had 11:33trouble valuing data as well um uh 11:36valuing the models and what we can do 11:38with them is going going to be part of 11:40this as 11:41[Music] 11:44well so I'm going to move us on to our 11:47second segment uh of the day um so open 11:50AI this week announced the new feature 11:52uh they call structured outputs um and 11:56uh this is huge uh although it might not 11:58seem like it on the surface for people 12:00who are like not in the day-to-day work 12:02of of AI um effectively what they're 12:05offering is for the very first time uh 12:08model developers uh are allowed to 12:11basically work with their system um to 12:13constrain their outputs to match 12:16specific schemas that are defined by 12:18Engineers um and this is a little bit 12:20nerdy but I think it's actually worth 12:22kind of walking through the technical 12:23points here because I think it's one of 12:24the areas where if you dive a little bit 12:26into the technical kind of understand 12:28what's going on you may recognize why 12:30out of a summer of lots and lots of 12:32announcements of AI this may actually 12:34end up being the biggest announcement of 12:36the summer in some ways um so I'm going 12:38to try to explain this and then I think 12:39Marina you'll keep me honest you should 12:41be like that's completely wrong Tim 12:42you've completely misunderstood what 12:43they're trying to do the way as I 12:45understand it is that language models 12:47are of course Very powerful they can do 12:49all sorts of remarkable things but the 12:51problem is that they kind of output in 12:53sort of non determinative ways they like 12:56produce outputs that are kind of 12:57difficult to kind of constrain and 12:59standardize and this has been a really 13:01tough problem because you know you have 13:03to take Ai and then you have to connect 13:04it to all the other all these other 13:06traditional systems that are expecting 13:08structured data right like there's a 13:10computer just being like Oh well I'm 13:12expecting a table that has like the 13:14following elements within it and it's 13:15been very hard to kind of like integrate 13:17language models with that um and is what 13:20open AI is saying here that you can 13:22finally for the first time do that 13:23reliably uh correct me if I'm wrong I'm 13:25just kind of thinking through this the 13:27thing I'm actually going to push back on 13:28is this whole finally for the first time 13:30thing this is not for the first time the 13:32fact that we before were like all right 13:34structured outputs semi-structured 13:36outputs are where it's at we used to say 13:38what you do with unstructured data this 13:39is work that I've done for years is you 13:41try to turn it into something more 13:42structured so it's features and you can 13:44feed it into a you know classifier feed 13:46it into ML and and go from there then 13:49everybody said oh Foundation models all 13:51right now we can doesn't matter we no 13:54more structure is needed no more data is 13:55needed nothing is needed we're just 13:56going to go and have unstructured data 13:58is going to everything you go and you 14:00work with that for a while and you go no 14:02guess not all right we're going to go 14:04ahead and walk it back a little we're 14:05going to walk it back a little let's go 14:06back to the fact that especially if 14:07you're trying to mix and match a 14:09heterogeneous system you do need 14:11structure output because these things 14:12don't know how to talk to each other so 14:13I'm going to pretty strongly push back 14:15on the for the first time and go back to 14:17no now that we're trying to be practical 14:19about it we've gone back to the fact 14:21that you need to impose a bit of 14:22structure I would also say that this is 14:24with like the success of uh code models 14:26where we see that there already is a lot 14:28more structure imposed on what kind of 14:30things can go in and can go out there's 14:31some lessons being learned there again 14:33going oh maybe we don't do just 14:35generally unstructured text and we're 14:37going to go back to having a bit of a 14:39mix um Kush would you agree with that 14:41particular we're kind of back yeah yeah 14:44no I mean I think that's exactly right I 14:46mean one way to look at it is I mean 14:49you're kind of breaking this bucking 14:51bronco that uh just came out of the blue 14:53in the last couple of years and bringing 14:55it back to uh where it should be right I 14:58mean the the control and the governance 15:01I mean all of that is part of making 15:03these things practical right and um I 15:05think another way to look at it is uh 15:08and one good thing about these language 15:10models is uh that they're very creative 15:12they're coming up with all sorts of uh 15:14different things but it's really a 15:16tradeoff safety versus creativity and um 15:20the control the the constraint is bring 15:23us back to to that safety aspect and um 15:26uh if you're inspiring a poet I mean go 15:29ride that Bronco it's all good um but uh 15:31I mean for all of the Enterprise use 15:33cases that uh uh that we care about that 15:35are going to make the the productivity 15:37differences and all that sort of stuff 15:39then uh that extra control is is where 15:41it's at yeah for sure so show bit am I 15:44um am I just being an open AI shill 15:47here just really hyping this feature 15:50where I guess Marina is just telling us 15:51like this has all been said and done 15:53before you know they're just selling 15:54something that everybody has known how 15:55to do for a long time so so Tim hot take 15:57on this this is the first time open AI 16:00is now appreciating and admitting that 16:03the whole workflow end to end won't be 16:05done by an llm they have admitted by 16:08releasing this that at step number three 16:11somebody's going to call an llm and 16:13expect it to behave in a structured 16:15manner so it can be a part of a team 16:18that does an end to and flow other 16:20aspects will be automation RPA there'll 16:22be some regular AI they'll be just PL 16:24old API calls but now llm they have 16:27admitted to this by releasing this that 16:29it's now down to a subtask level versus 16:32being the llm that's going to do the 16:34entire process end to end right so I 16:36think it's it's a really hard take on 16:37what they're doing for for practical 16:39deployments for me in the field we uh we 16:42are we are the launch partners with open 16:45Ai and whatnot right we do a ton of open 16:46AI with clients in our workflows last 16:50week um on Monday actually we were 16:52working with a large Healthcare client 16:54where we're reading greams of different 16:56documents and stuff and we extracting 16:57things from those documents right so 16:58talking about my HealthCare coverage I 17:00need to know what's in network what's 17:02out of network what's family coverage 17:03what's single coverage and so and so 17:05forth so using an llm to go extract 17:07things out from it every time we run 17:09this against our rubric of uh checking 17:12the accuracy there quite often response 17:15back with a blurb instead of giving me 17:17the in network and out of network so the 17:19way we used to solve this historically 17:21we would ask questions in a manner and 17:23then we provided some coaching saying 17:24just respond with the actual dollar 17:26amount the problem there used to be it 17:29responds back with saying 17:3214.9 and in three out of 10 cases it'll 17:34forget to put million in front of it 17:37right there's like practical issues with 17:39you having with leveraging these large 17:40language models and then we like okay 17:41fine just give me the entire thing and 17:43then like to Marina's point I'll just 17:45use a small regx somewhere to extract 17:47what I need from it and then I'll plug 17:48it back in that was a horrible way of 17:50doing things in production yeah that's 17:52awful having a commitment now saying 17:54that this is the JS I'm going to get and 17:56if you can't fill that number if you 17:58don't don't know what the single 18:00coverage is for outof Network it'll be 18:02null it'll be blank then I can do 18:04something in a structured manner raise 18:05some some alerts and have a workflow 18:08accordingly I think it's brilliant 18:09they're allowing you to do this this 18:11combined with the price drop that we got 18:1450% decrease in inputs 33% in the 18:16outputs makes it very very easy for us 18:19to plug it in the 40 Mini price is just 18:22Rock Bottom it's slow it's very 18:24inexpensive to deploy mini even the 18:27fine-tune versions of mini now they're 18:29allowing you to go fine tun these models 18:31very very easily have a structured oper 18:33around it so they've understood the fact 18:35that instead of doing a generic top down 18:38I'll take care of the entire thing all 18:40the way down to a subtask level it has 18:42to be fine-tuned for that task it has to 18:44be super inexpensive and has to be a 18:46good contract on what the input and the 18:47output structure coming out right in 18:49other words like a good a good tool to 18:51be used in the 18:52Enterprise um so the super interesting 18:55takes on this it definitely went in a 18:56direction that I wasn't expecting but I 18:58think is like very helpful in kind of 19:00thinking through why open AI did this I 19:02think the final aspect of this I want to 19:04touch on is it was very funny I mean as 19:05someone who you know is a software 19:07engineer kind of turned into a lawyer 19:09you know I like read this very long blog 19:10post about structured outputs and then 19:12at the very end it's like oh by the way 19:14it's not eligible for zero data 19:15retention which I think was a very 19:17interesting part of the announcement was 19:19basically like normally the promises 19:21that open AI will not train on any data 19:23that you send in through the API on the 19:24Enterprise basis but in this one case if 19:27you send in a schema they're going to 19:29they're going to train on that um and I 19:31guess for our listeners I think it'd be 19:33useful for them to hear a little bit 19:34some intuitions for why it is that open 19:36AI sees this data as so uniquely 19:39valuable right that they're going to say 19:40we've got this General policy of zero 19:42data retention but for this tiny little 19:44segment we're going to cut out a hole 19:45and if you send us our schemas we 19:47definitely want to train on that um 19:48Chris I see you nodding but I don't know 19:49if you want to speak to speak to why 19:51they would do something like this yeah I 19:53mean uh I was reading the announcement 19:55as well and uh I think the there's 19:57they're taking two different technical 19:59approaches to make this work right one 20:00is just training on more and more of 20:03these schemas the second is constrain 20:05decoding using this uh context free 20:07grammar to really make sure that um uh 20:11what comes out is uh is really I mean 20:13matching the the schema and stuff so on 20:16the first of the two I mean it's really 20:18hard to to get this sort of variety of 20:21uh what kind of schemas are going to be 20:24out there this is not something you can 20:25just download from the web and uh I mean 20:28in some of our work we also I mean look 20:30at very unique Enterprise sort of um uh 20:33policy documents or or other stuff like 20:35that and it's just not easy like um I 20:38was uh talking with one of my group 20:40members yesterday we were trying to 20:42figure out what are like different 20:43policies for um or guidelines for 20:45different professions and I was looking 20:47like can I get the New York State barber 20:49license uh guidelines like what does a 20:52barber need to do to do their job and 20:54like there's tons of stuff like that 20:56that um is like really not out there so 21:00I mean just the the uniqueness of it is 21:03is the key I think I think that's that's 21:04absolutely right and I think that will 21:06be coming sort of the increasing battle 21:08it seems like right as like all of the 21:10easy to get data is now accessible now 21:13the kind of question is like who's got 21:14these kind of access to like very hard 21:16to get data and this kind of these 21:17schemas are they're they're valuable 21:19tokens right they're they're unique 21:21tokens um in a lot of ways so this has 21:23been a big struggle for us with our 21:25clients in Enterprise settings we go 21:26through Enterprise security govern 21:29when we take a new product and we have 21:31to make sure that it's being used in a 21:32particular way everybody signs off on it 21:34and so on so forth right so we we're 21:36struggling with this with our 21:37Enterprises when when you Outsource your 21:40API calls to a third party then every 21:42time the API calls change or they do 21:45something differently or now in this 21:46case there's the retention issue with 21:48with the schemas right you need to go 21:51back through the whole process and I 21:53don't think Enterprises have a good 21:54mechanism to understand capture and then 21:57act on each one of these incremental 21:59updates that happen so it scares me a 22:01little bit that enterprises will end up 22:03approving a product in a particular 22:05state but it so rapidly evolves with 22:07features and stuff that you won't be 22:08able to go back in time and say I have 22:10to this small incremental thing has to 22:12be done differently the data scientists 22:14will start getting super excited about 22:16these function calls and and about these 22:18structured outputs and start using it 22:20and then that's where Kush and team are 22:21going to come in and say guys time out 22:23there has to be a good discipline around 22:24how you govern incremental updates that 22:26are happening to these so you don't get 22:28yourself into trouble so I think that's 22:30a very unaddressed issue with at least 22:32my Enterprise 22:33[Music] 22:37clients so I'm going to move us on to 22:39our final uh story of the day um it was 22:42announced last week that Nome shazir who 22:44was the CEO of character AI was going to 22:46rejoin Google along with a core team 22:48from his company um and also that Google 22:50was going to acquire a license to all 22:52character IP um this is widely seen 22:55though it's disputed as an acquisition 22:57ultimately of character um which had 23:00raised something like $150 million and 23:02was basically building sort of 23:03personalized companion AIS um and so I 23:06really want to go into this story 23:07because it's very interesting and part 23:08of a trend of uh Acquisitions in the 23:11space if you will um that I think are 23:13very interesting and I think get us to 23:14thinking a little bit about how this 23:16Market's going to evolve and what we 23:17really anticipate from AI startups the 23:19next 12 to 24 months Kush I wanted to 23:22turn to you first is you know why is a 23:24company like Google interested in a 23:26company like character AI at all you 23:28know like it feels like Google's got all 23:30the resources in the world to do all the 23:32AI um why are they acquiring companies 23:34at all at Great cost like it feels like 23:36couldn't they just build kind of a 23:37character product on their own um and 23:39we' love to get your thoughts on what do 23:40you think is motivating this in the 23:41first place yeah I think that's the 23:44similar question like why does IBM 23:46research exist versus um why don't we 23:49just tell me a little more about that 23:50yeah yeah I mean why don't we just keep 23:52acquiring a lot of startups I think uh 23:55there's always going to be a balance 23:56between kind of organic growth and and 23:58uh kind of uh the acquisition sort of 24:00thing um uh there's always a spark of 24:04some idea it you can't assume that uh 24:07you're going to have all of them and uh 24:10uh I mean in these cases there is 24:12something unique there's something where 24:14there's a market that they've touched on 24:16and something that I think only a 24:18startup can can maybe tap into because 24:20um they have a different pulse of the 24:22scene so I think it it makes sense to 24:25for a company like Google to to have a a 24:28mix of ways that they they grow yeah for 24:30sure I to push you a little bit further 24:32on that do you think it's cuz like is 24:34there some kind of compliment like 24:35what's the angle that I think you think 24:37Google's trying to chase after here 24:38because I I mean it's a search company 24:40right like ultimately um this feels like 24:42very consumer uh in some ways of what 24:45they're trying to do yeah I mean maybe 24:46they don't think they're they are a 24:48search company going forward I don't 24:49know um maybe they're uh edging on to 24:53there's I mean more things or or other 24:55things but uh I think just uh once you 24:58get I mean something interesting 25:00something exciting uh that just draws 25:02customers to you draws consumers to you 25:05and then uh uh you can keep them and get 25:07them into other stuff so yeah as part of 25:10a pivot for sure um so maybe we could 25:13take the other angle at the story I 25:15think which is you can see it from the 25:16perspective of the acquirer why would 25:18Google do something like this but I 25:19think it's also worth investigating it 25:21from the perspective of the startup um 25:23you know Marina there was a bunch of 25:25commentary online where people were 25:26saying look you've seen Adept go through 25:29a similar transaction there's another 25:30company called inflection that went 25:31through a similar transaction these are 25:33companies that have raised an enormous 25:35amount of money um and by all accounts 25:37would be very successful right like 25:39maybe some of the most successful 25:40startups in the AI space um but as yet 25:43the founders are choosing to to sell um 25:46effectively right they're they're 25:47choosing to go and join the big tech 25:49companies um do you have a theory for 25:52that like why would you I mean if I'm 25:54sitting there I'm n shazir you know I've 25:55raised $150 million that's certainly 25:57more money than Ever Raised right um 25:59what is what is motivating these kind of 26:01Founders to say okay well actually want 26:03to kind of throw in with the big 26:04companies rather than trying to make it 26:06on my own and does it suggest you know 26:08problems in the startup Market do you 26:09think I mean even 150 million can be 26:11burned through pretty quickly if you're 26:12doing a whole bunch of your own training 26:14what is 26:1515 everything else there might be a a 26:18case here of again if there's an 26:20understanding that you want to have a 26:23sort of a pre-baked user base or a 26:26pre-baked you know set of being able to 26:28use a whole bunch of um of resources 26:31which a company like uh Google company 26:33like meta they're going to be uh really 26:34quite good with that um again 26:36potentially other people to collaborate 26:37with I really will second what kush said 26:39which is you've had one or two or three 26:41good ideas it doesn't mean that you're 26:42going to have 40 and they really are a 26:44ton of extremely interesting smart 26:46people who are working in these 26:47companies so it may be that there's a 26:49desire to to also do that and as well 26:52and and have that partnership be a lot 26:54more close in order to be able to to see 26:56that that yeah I mean zooming out to the 26:58macro level I mean cha do you think that 27:00um like what does this prage I guess for 27:04kind of like startups in the AI space in 27:05general like are you seeing more AI 27:06startups over time because I think 27:08there's almost one way of reading this 27:09which is well even if these companies 27:11that have raised so much money can't 27:12make it 27:13independently uh you know like no one 27:15can make it right like we're about to 27:17see a lot of consolidation in the AI 27:18startup space I think the the core 27:20values the fundamentals haven't changed 27:22you can't have a thin wrapper around an 27:24open AI API call and expect it to keep 27:27keep drawing more right so you you do 27:29realize that the intellectual property 27:31that you've built is what people are 27:32going to pay for and the talent that you 27:33have that you have assembled that 27:35particular team that's what is is uh 27:37golden now big companies will try to 27:40walk around Acquisitions and come get 27:44very creative to work around any of the 27:46antitrust rules and things of that 27:47nature as well right so in this case 27:49they're not acquiring it they are 27:51getting hiring some people or they're 27:53licensing some terms and so on so forth 27:54right so you can see that there are 27:56there some motivation on not just 27:57outright acquiring it but on the flip 28:00side just like any startup environment 28:01you'll also see big companies like whz 28:05which Google was trying to acquire and 28:07uh whz walked away from 28:09$23 billion 28:12offer and uh this I'm just laughing 28:14because it's like that's like a 28:15literally hilarious amount of money 28:17right that is insane and O who's the the 28:20co-founder of BZ he wrote a very 28:22humbling letter to all the employees 28:24explaining them why you're not getting 28:25rich today right Essen explain to them 28:28why I'm I'm not taking this offer it's a 28:29very humbling offer but here are the 28:31reasons why we believe that going IPO is 28:33a big bigger value ad and so on so forth 28:35right historically we have seen a lot of 28:37misses and hits and misses Yahoo trying 28:39to sell itself to Google or like Netflix 28:42to Blockbuster all of these have been 28:43multiple reminders that you you have to 28:46know what's your value ad and how much 28:48of that is a differentiator with a high 28:50Moe so others can't just come in and do 28:52what you're doing right so it takes a 28:54while to understand the rhythm of where 28:55you lie where you lie in the competitive 28:57landscape and R forecast I think we put 28:59undue pressure on co-founders on the on 29:01the founders who who are just passioned 29:03about building a product but now all of 29:05a sudden we are we are surrounding them 29:08with Venture capitals who have different 29:11objectives than what you mean I need to 29:13build a 29:15business yeah I think they need to bring 29:18back Silicon Valley as a as episodes in 29:20today's world with llm that's right yeah 29:23it's for sure yeah I saw this great 29:25Twitter thread that was on like if we 29:27modernized you know Silicon Valley what 29:29would it be and just like everybody's in 29:30AI basically um I mean it goes to a 29:33point that Marina raised earlier in our 29:35first segment though is like it almost 29:36kind of feels like this is almost like 29:38the micro version of the market being 29:40not able to price these startups 29:41properly like it feels like in a lot of 29:43these cases like these big companies 29:45like goog are like ultimately acquiring 29:47the talent versus necessarily like the 29:50product um I guess character you can 29:52maybe debate because it actually had a 29:53big install base but it feels like at 29:55the core of it is just simply like 29:57here's a team of people who seem to be 29:58able to get what they want out of the AI 30:01and like that actually ends up being 30:02like this huge value that's almost 30:04separate from like did you have a 30:05blockbuster AI product release um and 30:08yeah it kind of goes to these 30:08interesting questions that I'm thinking 30:09about now about like how do you how do 30:11you actually value these companies right 30:13because it's just like so unclear in 30:14such a fluid 30:16environment um any final thoughts on 30:19this um super super interesting and I I 30:21think again I mean to argue against 30:22myself you know this is also during the 30:25same week we saw a bunch of top 30:26leadership from uh opening I leave right 30:28and so it's not necessarily all 30:30consolidation it's possible that you 30:31know people are moving between big 30:33companies and also creating like new 30:34startups onto themselves um so any final 30:37thoughts to round this out for today um 30:39just one um I mean conversation I was 30:42having with my brother-in-law last week 30:43not related to this but uh I mean the 30:46difference between running your own 30:47business versus doing a job in a big 30:50company right and the lifestyle sort of 30:52issues there and um I think like I mean 30:55the point you were making before Tim 30:57like uh if you just want to make one 30:59product versus building a business I 31:01think maybe a lot of the folks that are 31:03um getting into this right now um are 31:06not in it for maybe that lifestyle or 31:09for that uh uh business building sort of 31:12uh sort of way of of going about it so 31:16maybe it's just a way for them to return 31:19back to their natural sort of State um 31:21so so that could be driving it as well 31:23kind more of the lifestyle issue yeah I 31:25believe that for sure yeah it's I mean 31:27personally crazy to do a startup so and 31:30then got a friend who was a Founder who 31:31is like it's literally an irrational act 31:33to do a startup 31:35so um well great on that note uh no 31:38shade to anyone else who has already 31:40been on mixture of experts as a panelist 31:41but I have to say this is my favorite 31:43panel the marina Kush show bit you know 31:46power Trio is basically like we just get 31:48the best conversations all the time so I 31:50appreciate all three of you coming on 31:51the show and for all you listeners 31:53thanks for joining us this week uh if 31:55you enjoyed what you heard you can get 31:56us on Apple podcast Spotify and podcast 31:58platforms everywhere and we will see you 32:01same time next week