WEBVTT - Monologue: AI Hits A Wall

0:00:02.800 --> 0:00:07.280
<v Speaker 1>Zoe Media. Hello and welcome to this week's Better Offline Monologue.

0:00:07.320 --> 0:00:17.919
<v Speaker 1>I'm your host ed Zetron. It's going to be a

0:00:18.000 --> 0:00:21.000
<v Speaker 1>long week, after a long month, after a long few years,

0:00:21.040 --> 0:00:23.360
<v Speaker 1>so I'm recording this monologue a little earlier than usual

0:00:23.440 --> 0:00:26.200
<v Speaker 1>on Tuesday, to be specific, which means I don't have

0:00:26.280 --> 0:00:29.120
<v Speaker 1>thoughts on in Video's earnings and won't have them unless,

0:00:29.160 --> 0:00:32.479
<v Speaker 1>of course, things go really haywire. Now I've said that

0:00:32.479 --> 0:00:34.640
<v Speaker 1>that sounds possible, but I think it's likely the no

0:00:34.800 --> 0:00:37.479
<v Speaker 1>it loads refused cash dump will continue at least for

0:00:37.520 --> 0:00:40.920
<v Speaker 1>another quarter there In any case, we have a far

0:00:40.960 --> 0:00:44.040
<v Speaker 1>more pressing problem, the deterioration of open Ai and the

0:00:44.159 --> 0:00:48.040
<v Speaker 1>AI industry hitting a wall. Last week, open Ai completed

0:00:48.040 --> 0:00:51.560
<v Speaker 1>the seven billion dollar employee share sale, but unlike most

0:00:51.600 --> 0:00:54.760
<v Speaker 1>employee share sales, open Ai used its own cashed by

0:00:54.800 --> 0:00:58.240
<v Speaker 1>its employee stock the previously raised valuation of eight hundred

0:00:58.240 --> 0:01:01.400
<v Speaker 1>and sixty five billion dollars. A few days later, both

0:01:01.480 --> 0:01:04.839
<v Speaker 1>chief operating officer and former CFO Brad Lightcap and Chief

0:01:04.880 --> 0:01:09.440
<v Speaker 1>revenue officer Denise Dresser left the company. Dresser had been

0:01:09.480 --> 0:01:11.760
<v Speaker 1>at the company less than a year and likely walked

0:01:11.800 --> 0:01:14.319
<v Speaker 1>away from tens of millions of dollars worth of unvested

0:01:14.360 --> 0:01:17.400
<v Speaker 1>open Ai shares. In April, she said that she had

0:01:17.440 --> 0:01:20.520
<v Speaker 1>and I quote CNBC never seen this level of conviction

0:01:20.640 --> 0:01:23.920
<v Speaker 1>spread so quickly and consistently within the enterprise industry, though

0:01:23.959 --> 0:01:26.680
<v Speaker 1>it appears that said conviction didn't spread far enough to

0:01:26.720 --> 0:01:28.760
<v Speaker 1>keep her at a company that definitely plans to go

0:01:28.800 --> 0:01:32.160
<v Speaker 1>public some time in twenty twenty seven, unless, of course,

0:01:32.240 --> 0:01:37.240
<v Speaker 1>it's not quite so definite anymore. Anthropic has dramatically accelerated

0:01:37.280 --> 0:01:40.039
<v Speaker 1>its plans to go public, with Bloomberg reporting that is

0:01:40.080 --> 0:01:42.840
<v Speaker 1>aiming for a September October listing, though we're yet to

0:01:42.880 --> 0:01:45.959
<v Speaker 1>see any reporting on its actual s. One Now, this

0:01:46.000 --> 0:01:48.200
<v Speaker 1>is a problem for open ai in many ways, but

0:01:48.320 --> 0:01:50.960
<v Speaker 1>let's get through all the news first. At the very

0:01:51.080 --> 0:01:53.640
<v Speaker 1>end of last week, Bloomberg also reported that open ai

0:01:53.880 --> 0:01:56.440
<v Speaker 1>was on track to hit forty billion dollars in annualized

0:01:56.480 --> 0:02:00.360
<v Speaker 1>revenue seven billion dollars in annualized run rate, less than

0:02:00.400 --> 0:02:03.440
<v Speaker 1>Anthropic had claimed it hit in May. And yes, you

0:02:03.480 --> 0:02:05.400
<v Speaker 1>heard that right, I said on track, which does not

0:02:05.520 --> 0:02:07.240
<v Speaker 1>mean it hit the number. And no, it did not

0:02:07.320 --> 0:02:10.280
<v Speaker 1>define run ray because Bloomberg neglected to do so either,

0:02:10.360 --> 0:02:15.520
<v Speaker 1>And well, who needs definitions when you're just talking about bullshit, right,

0:02:17.200 --> 0:02:19.639
<v Speaker 1>it could mean the last four weeks multiplied by twelve

0:02:19.680 --> 0:02:22.960
<v Speaker 1>or thirteen, or really anything. We truly don't know. Everyone

0:02:23.000 --> 0:02:26.280
<v Speaker 1>seems to use a goddamn different definition of run ray anyway.

0:02:26.360 --> 0:02:29.720
<v Speaker 1>Then on Monday, Bloomberg reported that Anthropic had hit sixty

0:02:29.720 --> 0:02:32.680
<v Speaker 1>five billion dollars in annualized run ray, which it defined

0:02:32.720 --> 0:02:35.320
<v Speaker 1>I shit you not as a metric that projects full

0:02:35.440 --> 0:02:38.400
<v Speaker 1>year revenue from a shorter period with no further explanation.

0:02:39.720 --> 0:02:42.799
<v Speaker 1>Who fucking cares, right, it's only about what's ostensibly meant

0:02:42.800 --> 0:02:44.840
<v Speaker 1>to be the largest AI company in the world, with

0:02:45.200 --> 0:02:48.800
<v Speaker 1>three hundred billion dollars worth of commitments. No need to

0:02:48.840 --> 0:02:52.080
<v Speaker 1>actually know what you're talking about, No need to say

0:02:52.120 --> 0:02:55.720
<v Speaker 1>anything further about a market moving number, no need to

0:02:55.760 --> 0:02:59.079
<v Speaker 1>do that. Why fucking bother, right, Just put the story

0:02:59.120 --> 0:03:01.360
<v Speaker 1>in the bag, the hype in the bag. Check please.

0:03:01.639 --> 0:03:04.720
<v Speaker 1>We wouldn't really say that at grocery store anyway. In

0:03:04.760 --> 0:03:07.560
<v Speaker 1>other words, Open Ai had taken four months to not

0:03:07.639 --> 0:03:10.440
<v Speaker 1>quite catch up with Anthropic, only for Anthropic to quickly

0:03:10.520 --> 0:03:14.440
<v Speaker 1>drop that the race wasn't even close. Regardless, AI boosters

0:03:14.440 --> 0:03:17.080
<v Speaker 1>immediately got a little anxious because many had been convinced

0:03:17.120 --> 0:03:20.240
<v Speaker 1>by anonymous sources and questionable data blocks like yipit that

0:03:20.320 --> 0:03:23.120
<v Speaker 1>annualized run rates were somewhere between seventy billion and one

0:03:23.160 --> 0:03:26.400
<v Speaker 1>hundred billion. I guess the voices in their heads weren't reliable.

0:03:26.880 --> 0:03:29.040
<v Speaker 1>Did they try sacrificing an animal? I don't know what

0:03:29.080 --> 0:03:32.400
<v Speaker 1>these fucking people do. Isn't real. It's just listening to

0:03:32.440 --> 0:03:35.640
<v Speaker 1>the voices that say so, Yes, one hundred billion annualized

0:03:35.640 --> 0:03:38.240
<v Speaker 1>by end of year. You only the way they will

0:03:38.280 --> 0:03:42.520
<v Speaker 1>be able to invest in this IPO. Yes anyway said.

0:03:42.560 --> 0:03:45.280
<v Speaker 1>Anxiety is spread across the AI industry, with even the

0:03:45.280 --> 0:03:48.160
<v Speaker 1>most mold poisoned boosters now fretting that both Anthropic and

0:03:48.200 --> 0:03:51.040
<v Speaker 1>open AI's revenues had slowed down, mostly because they have.

0:03:51.840 --> 0:03:53.800
<v Speaker 1>I will also add that these numbers, which come from

0:03:53.840 --> 0:03:56.560
<v Speaker 1>the end of July are still from the era of tokenmaxing,

0:03:56.640 --> 0:03:59.360
<v Speaker 1>where companies were encouraging employees to burn as many tokens

0:03:59.400 --> 0:04:02.400
<v Speaker 1>as possible with no regard for costs, meaning that growth

0:04:02.440 --> 0:04:05.440
<v Speaker 1>will invariably slow from here as enterprises cap spending, as

0:04:05.480 --> 0:04:09.520
<v Speaker 1>Walmart and Stripe and multiple others have done. A day

0:04:09.560 --> 0:04:12.320
<v Speaker 1>after Anthropic crept in its coffee, open Ai made the

0:04:12.320 --> 0:04:16.080
<v Speaker 1>incredible decision too, and I quote pause some frontier reinforcement

0:04:16.200 --> 0:04:19.039
<v Speaker 1>learning training to ensure it can meet the appropriate alignment.

0:04:19.360 --> 0:04:21.520
<v Speaker 1>A complex way of saying that open ai will stop

0:04:21.560 --> 0:04:23.640
<v Speaker 1>training its latest models for at least the next few

0:04:23.640 --> 0:04:26.800
<v Speaker 1>weeks to increase safety. Connected to the recent incident where

0:04:26.800 --> 0:04:29.480
<v Speaker 1>an open ai model in an incorrectly configured server was

0:04:29.520 --> 0:04:31.599
<v Speaker 1>able to hack hugging Face, which is less of a

0:04:31.640 --> 0:04:33.240
<v Speaker 1>case of it going rogue and more of a case

0:04:33.279 --> 0:04:35.440
<v Speaker 1>of open ai being bad at its job. If you

0:04:35.520 --> 0:04:37.480
<v Speaker 1>tell software to do something, it's going to try and

0:04:37.480 --> 0:04:40.080
<v Speaker 1>do it. Oh, it isn't lined right. You program the

0:04:40.120 --> 0:04:43.760
<v Speaker 1>software poorly. You're wrong here. This isn't the magic of Ai.

0:04:44.240 --> 0:04:47.760
<v Speaker 1>This is the tragic of open Ai. It's unclear what

0:04:47.800 --> 0:04:50.599
<v Speaker 1>open ai hope to achieve here, rather than vaguely saying

0:04:50.640 --> 0:04:53.360
<v Speaker 1>we care about safety. But from everything I'm seeing on

0:04:53.400 --> 0:04:56.080
<v Speaker 1>social media and in my inbox, everybody is interpreting this

0:04:56.200 --> 0:04:58.680
<v Speaker 1>as open ai has stopped training new models, which is

0:04:59.120 --> 0:05:02.080
<v Speaker 1>also the truth. Be only they claim for two weeks.

0:05:02.400 --> 0:05:06.239
<v Speaker 1>Maybe if you believe them. This is a really fucking

0:05:06.279 --> 0:05:08.479
<v Speaker 1>stupid idea, by the way, or at least was a

0:05:08.520 --> 0:05:10.800
<v Speaker 1>really fucking stupid idea to announce at a time when

0:05:10.800 --> 0:05:13.880
<v Speaker 1>everybody has noticed that open Aiy's revenue growth is accelerated.

0:05:14.640 --> 0:05:17.039
<v Speaker 1>New models and features are effectively the only way for

0:05:17.080 --> 0:05:19.240
<v Speaker 1>these companies to keep growing, and open ai is far

0:05:19.279 --> 0:05:21.960
<v Speaker 1>behind Anthropic, which makes the idea of not doing anything

0:05:22.000 --> 0:05:26.159
<v Speaker 1>new or much new all the more questionable. Open Ai will,

0:05:26.160 --> 0:05:28.320
<v Speaker 1>of course try and spin this as it's new models

0:05:28.320 --> 0:05:31.640
<v Speaker 1>having incredible new abilities it must prepare for. Oh god,

0:05:31.680 --> 0:05:34.480
<v Speaker 1>are you gonna see look at anyone saying that you

0:05:34.640 --> 0:05:36.560
<v Speaker 1>just you just look at them. You can see the

0:05:36.600 --> 0:05:39.200
<v Speaker 1>three finger thing from Chlorius bast It's like, Yep, there's

0:05:39.200 --> 0:05:41.160
<v Speaker 1>someone who doesn't give a shit. There's someone who will

0:05:41.520 --> 0:05:46.280
<v Speaker 1>scarf down the marketing slot when fed to them. Mmm yummy. Anyway,

0:05:46.480 --> 0:05:48.599
<v Speaker 1>this is all very rumant, but the word on the

0:05:48.640 --> 0:05:50.880
<v Speaker 1>street is that this is open ai hitting a wall

0:05:50.880 --> 0:05:53.320
<v Speaker 1>and running out of compute to serve inference and train

0:05:53.400 --> 0:05:57.000
<v Speaker 1>new models, with one person suggesting that Anthropic could follow suit.

0:05:58.200 --> 0:06:00.840
<v Speaker 1>This does make sense for Anthropic, who got so desperate

0:06:00.880 --> 0:06:03.960
<v Speaker 1>that it's renting capacity from Elon, Musk, SpaceX and potentially

0:06:04.000 --> 0:06:06.640
<v Speaker 1>even Meta. But I'm surprised to hear it about open Ai,

0:06:06.720 --> 0:06:09.160
<v Speaker 1>who boasted earlier in the year about its massive compute

0:06:09.160 --> 0:06:12.800
<v Speaker 1>advantage over its competitors, and again it's also very possible

0:06:12.839 --> 0:06:15.599
<v Speaker 1>that compute capacity is just not coming online very fast.

0:06:16.400 --> 0:06:18.960
<v Speaker 1>Sources of Oracle tell me that stargay Abelaine, which was

0:06:18.960 --> 0:06:21.640
<v Speaker 1>meant to be finished in June or maybe the beginning

0:06:21.640 --> 0:06:23.160
<v Speaker 1>of the year or maybe the end of the year,

0:06:23.200 --> 0:06:25.880
<v Speaker 1>it's really unclear at this point, only has three out

0:06:25.880 --> 0:06:28.360
<v Speaker 1>of eight buildings for the operational with a fourth in

0:06:28.440 --> 0:06:31.000
<v Speaker 1>some sort of strange limbo where it's sort of ready

0:06:31.000 --> 0:06:33.720
<v Speaker 1>but not quite isn't clear if it's generating any money,

0:06:33.960 --> 0:06:36.400
<v Speaker 1>and you need money to pay for goods and services,

0:06:36.440 --> 0:06:39.800
<v Speaker 1>and it's called revenue. Not sure any of these companies

0:06:39.839 --> 0:06:42.719
<v Speaker 1>fully understand business, but anyway, this could also be a

0:06:42.720 --> 0:06:45.880
<v Speaker 1>way for open ai to save money. By the information,

0:06:46.040 --> 0:06:48.760
<v Speaker 1>open ai spent eight point six billion dollars on research

0:06:48.760 --> 0:06:51.400
<v Speaker 1>and development so model training in the first quarter of

0:06:51.440 --> 0:06:53.719
<v Speaker 1>twenty twenty six, a number that likely increased in Q

0:06:53.880 --> 0:06:56.880
<v Speaker 1>two with the launch of its GPT five point six sole,

0:06:57.000 --> 0:06:59.960
<v Speaker 1>lunar and cyber models. And also for all the compute

0:07:00.000 --> 0:07:04.600
<v Speaker 1>it was spending accidentally hacking hugging face I swear that's

0:07:04.640 --> 0:07:08.640
<v Speaker 1>like a felony, right, like, we don't. We don't really

0:07:08.680 --> 0:07:16.000
<v Speaker 1>have like laws anymore, I realize. But anyway, in any case,

0:07:16.400 --> 0:07:18.720
<v Speaker 1>if Anthropic and open AI have truly hit the limits

0:07:18.760 --> 0:07:22.000
<v Speaker 1>of compute, there's very little can accelerate it coming online,

0:07:22.120 --> 0:07:24.600
<v Speaker 1>as every AI data center is the equivalent of cramming

0:07:24.640 --> 0:07:27.360
<v Speaker 1>an entire city's power, usually spread over a billion or

0:07:27.400 --> 0:07:30.560
<v Speaker 1>billions or more square feet into a round a million

0:07:30.600 --> 0:07:33.000
<v Speaker 1>square feet built in the space of two years, with

0:07:33.080 --> 0:07:37.200
<v Speaker 1>constantly increasing costs around memory, specialist talent, specialist materials, and

0:07:37.200 --> 0:07:39.440
<v Speaker 1>the oil necessary for the gas, turbines and the trucks

0:07:39.480 --> 0:07:41.119
<v Speaker 1>to bring all this shit to the plot of land.

0:07:42.480 --> 0:07:45.200
<v Speaker 1>And all of this is happening as revenue growth deccelerates.

0:07:46.120 --> 0:07:49.360
<v Speaker 1>If open AI and Anthropic can't or won't train new models,

0:07:49.680 --> 0:07:52.880
<v Speaker 1>that means they'll effectively become model wrapper companies, building products

0:07:52.880 --> 0:07:55.120
<v Speaker 1>on top of clawed or GBT models in the hopes

0:07:55.160 --> 0:07:57.600
<v Speaker 1>that they canduce further revenue growth, which is not going

0:07:57.680 --> 0:08:00.400
<v Speaker 1>to help very much given them. Basically every company that

0:08:00.400 --> 0:08:03.360
<v Speaker 1>builds models businesses on top of their models is both

0:08:03.440 --> 0:08:06.480
<v Speaker 1>unprofitable and can barely reach ninety million a month in revenue,

0:08:06.480 --> 0:08:08.560
<v Speaker 1>which wouldn't even cover a day's worth of inference if

0:08:08.600 --> 0:08:12.120
<v Speaker 1>it were profitable, which it would not be. I'll add

0:08:12.120 --> 0:08:15.120
<v Speaker 1>that if Anthropic beats open Ai to IPO, it's unlikely

0:08:15.160 --> 0:08:17.560
<v Speaker 1>open Ai will be able to float or at least

0:08:17.920 --> 0:08:21.480
<v Speaker 1>have anything approaching an IPO that anyone would actually want

0:08:21.520 --> 0:08:25.920
<v Speaker 1>to invest in. While both are unprofitable horribly so, Anthropic

0:08:25.960 --> 0:08:29.080
<v Speaker 1>has the revenue advantage and slightly better unit economics, which

0:08:29.080 --> 0:08:32.120
<v Speaker 1>is all it needs to make Sam Mortman's public offering untenable.

0:08:33.280 --> 0:08:35.679
<v Speaker 1>I think it's time to start having a reasonable discussion

0:08:35.720 --> 0:08:38.760
<v Speaker 1>about the potential collapse of open Ai and the ramifications

0:08:38.760 --> 0:08:42.360
<v Speaker 1>for counterparties like Google, Microsoft, Amazon, Core Weave, Cerebrus, and

0:08:42.440 --> 0:08:46.600
<v Speaker 1>of course Oracle and all associated venture capitalists, in particular Thrive,

0:08:46.640 --> 0:08:48.920
<v Speaker 1>who has billions of dollars tied up in this company,

0:08:49.120 --> 0:08:52.480
<v Speaker 1>and of course SoftBank, which needs that liquidity otherwise well

0:08:52.720 --> 0:08:55.840
<v Speaker 1>they're treat will fucked. If you're a journalist listening to this,

0:08:55.960 --> 0:08:58.920
<v Speaker 1>you need to start seriously thinking about and writing about

0:08:58.960 --> 0:09:01.760
<v Speaker 1>this because the current stay at this company suggested it

0:09:01.760 --> 0:09:05.240
<v Speaker 1>has few viable paths to avoid pertition. It can't raise

0:09:05.320 --> 0:09:08.320
<v Speaker 1>much more money. Ninety percent of its last funding round

0:09:08.400 --> 0:09:11.840
<v Speaker 1>came from Nvidia, Amazon and SoftBank, none of which are

0:09:11.920 --> 0:09:14.120
<v Speaker 1>really going to do a return trip. Soft Bank literally

0:09:14.120 --> 0:09:16.920
<v Speaker 1>can't afford it. Venture capitalists have never offered more than

0:09:16.960 --> 0:09:19.640
<v Speaker 1>twelve billion dollars of any Open Ai round, and even

0:09:19.679 --> 0:09:21.520
<v Speaker 1>if we are the World Telethon isn't going to be

0:09:21.640 --> 0:09:25.160
<v Speaker 1>enough to keep this pig inflated. What a disgusting way

0:09:25.200 --> 0:09:28.680
<v Speaker 1>of referring to a company. Anyway, continuing, there is not

0:09:28.840 --> 0:09:31.520
<v Speaker 1>really a viable path for open Ai here. I don't

0:09:31.520 --> 0:09:33.840
<v Speaker 1>know what they do, but it's kind of looking like

0:09:33.840 --> 0:09:35.680
<v Speaker 1>they're going to die. As I wrote my news there

0:09:35.800 --> 0:09:37.079
<v Speaker 1>this week, and I'll put a link to that in

0:09:37.600 --> 0:09:41.280
<v Speaker 1>the notes. Either way, I'm off on a flight tomorrow.

0:09:41.480 --> 0:09:44.079
<v Speaker 1>It's a crazy week. You'll know why soon. It's all

0:09:44.120 --> 0:09:47.079
<v Speaker 1>good stuff. I love you all, Thank you for listening.

0:09:47.240 --> 0:09:49.720
<v Speaker 1>I'll be back next week with the wonderful calnw Port.

0:09:50.080 --> 0:09:50.720
<v Speaker 1>Catch you then,