Media, Hello and welcomes a better offline.

I'm, of course your host ed Zitron.

We're in the third episode of our four part series where I give you a comprehensive explanation as to the origins of the AI bubble, the mythology sustaining it, and why it's destined to end really, really badly.

Now, if you're jumping in now, please start from the very beginning.

The reason why this is a four part my first ever, is because I want it to be comprehensive, and because this is a very big subject with a lot of moving parts and even more bullshit.

A few weeks ago, I published a premium newsletter that explained how everybody is losing money on generative AI, in part because the costs of running AI models is increasing, and in part because the software itself doesn't do enough to warrant the costs associated with running them, which are already subsidized and unprofitable for the model providers.

Outside of open and to a lesser extent, Anthropic, nobody seems to be making much revenue, with the most successful company being any Sphere, makers of AI coding tool Cursor, which hid five hundred million dollars have annualized so forty one point six million in one month a few months ago, just before Anthropic and open ai jacked up the prices for priority processing on enterprise queries, raising their operating costs as a result.

In any case, that's some pissport revenue for an industry that's meant to be the future of software.

Smart Watchers are projected to make thirty two billion dollars this year, and as I've mentioned in the past, the Magnificent Seven expect to make thirty five billion dollars or so in revenue from AI this year, and I think in total, when you're throw in core even all them, it's barely fifty five billion dollars in total.

Even Anthropic and open Ai seem a little lethargic, both burning billions of dollars while making by my estimates, no more than two billion dollars in Anthropics case this year so far and six point six two six billion dollars in twenty twenty five so far for open Ai, despite projections of five billion dollars and thirteen billion dollars respectively.

Outside of these two AI startups are floundering, struggling to stay alive and raising money in several hundred million dollar versus their negative gross margin businesses flounder as they dug into.

A few months ago, I could find only twelve AI powered companies making more than eight point three million dollars a month, with two of them slightly improving their revenue, specifically AI search company perplexd, which is now here one hundred and fifty million dollars an ur in or twelve point five million dollars a month, and AI coding startup Replayer, which has hit the same amount.

Both of these companies burn ridiculous amounts of money.

Paplexd burned one hundred and sixty four percent of its revenue on Amazon web services, open Ai and Anthropic last year, and while replet hasn't leaked its costs, the information reports its gross margins in July but twenty three percent, which doesn't include the cost of its free users, which you simply have to do with llms, as free users are capable of costing you a shit ton of money.

And some of you might say that's how they do it in software, Well, guess what software doesn't usually connect you to a model that can burn I don't know ten cents twenty cents every time they touch it, which may not seem like much, but when you're making three dollars on someone and they don't convert, it does problematically.

Your paid users also cost you more than they bring in as well.

In fact, every user loses you money in Generative AI because it's impossible to do cost control in a consistent manner.

A few months ago, I did a piece of Anthropic losing money on every single claud code subscriber.

And now I'm going to walk you through the whole story in a simplified fashion because it's quite important.

So claud Code is a coding environment that people use used, or I should really say, try to use to build software using generative AI.

It's available as part of Anthropics twenty dollars, one hundred dollars and two hundred dollars a month claud subscriptions, with the more expensive subscriptions having more generous rate limits.

Generally, these subscriptions are all you can eat.

You can use them as much as you want until you hit limits, rather than paying for the actual tokens you burn.

When I say burn tokens and someone reached out saying I should specify this, I'm describing how these models are traditionally built.

In general, you'll builded a dollar per million input tokens as in user feeding in data and output tokens the output created, so you wouldn't get one token built, so every million you get charged.

So, for example, Anthropic charges three dollars per million input tokens and six million output tokens to use its clauds on it for model, and it's about I think, well, a word before tokens should really look that up.

It's it also gets more complex as you get into things like generating code.

Nevertheless, claud code has been quite popular, and a user created a program called cc usage which allowed you to see your token burn the amount of tokens you were using.

You were actually burning using Anthropics models while using clawed code versus just getting charged a month and not knowing, and many were seeing that they were burning in the excess of their monthly spend.

To be clear, this is the token price based on anthropics own pricing, and thus the cost of Anthropic are likely not identical.

So I got a little clever using anthropics gross profit margins, I chose fifty five percent, and then a few weeks solved my article sixty percent was leaked.

I found at least twenty different accounts of people costing Anthropic anywhere from one hundred and thirty percent to three thousand and eighty four percent of their subscription.

There is also now a leader board called vibrank, where people compete to see how much they burn with the current leader burning and I sheit you not fifty two hundred and ninety one dollars of the course of a month.

Anthropic is, to be clear, the second largest model developer and has some of the best AI talent in the industry.

It has a better handle on its infrastructure than anyone outside of big tech and open AI, and it still cannot seem to fix this problem even with weekly rate limits brought in at the end of August.

While one could assume that Anthropic is simply letting users run wild, my theory is far simpler.

Even the model developers have no real way of limiting user activity, likely due to the architecture of generative AI.

I know it sounds insane, but at the most advanced level.

Even there, modeled providers are still prompting their models, and whatever rate limits may be in place appear to at times get completely ignored, and there doesn't seem to be anything they can do to stop it now.

Really, Anthropic counts amongst its capitalist apex predators one lone Chinese man who spent fifty thousand dollars to their compute in the space of a month fucking around with glord code.

Even if Anthropic was profitable, it isn't, and we'll burn billions of dollars this year.

A customer paying two hundred dollars a month ran up fifty thousand dollars in costs, immediately devouring the margin of any user running the service that day, that week, or even that month.

Even if Anthropics costs are half the published rates, they're not.

By the way, one guy amounted to one hundred and twenty five US is worth of monthly revenue.

This is not a real business.

That's a bad business without of control costs, and it doesn't appear anybody has these costs under control and face with the grim reality ahead of them, these companies are trying nasty little tricks on their customers to douce more revenue from them.

A few weeks ago, Replet, an unprofitable AI coding company, released a product called Agent three, which promised to be ten times more autonomous and offer infinitely more possible abilities, testing and fixing its code, constantly improving your application behind the scenes in a reflection loop.

Sounds very real, sounds extremely real, It's so real, but actually it isn't.

In reality.

This means you go and tell the model to build something, and it would go and do it, and you'll be shocked to hear that these models can't be relied upon to go and do anything.

Please note that this was launched a few months after Replet raise their prices, shifting to obfiscated effort based pricing that would charge the full scope of the agent's work.

And if you're wondering what the fuck that means, so are their customers.

Agent three has been a disaster.

Users found the tasks that previously cost a few dollars were spiraling into the hundreds of dollars, with the register reporting one customer found themselves within one thousand dollars bill after a week, and I quote them, I think it's just launch pricing adjustment.

Some tasks on new apps ran over an hour and forty five minutes and only charged four to six dollars, but editing pre existing apps seems to cost most overall.

I spend one K this week alone, and they told that to the register.

By the way, another user comp that costs skyrocket without any concrete results, and they quote the register here.

I typically spent between one hundred dollars and two hundred and fifty dollars a month.

I blew through seventy dollars in a night at Agent three launch, and another redditor wrote alleging the new tool also performed some questionable actions.

One prompt brute forced its way through authentication, redoing auth and hard resetting users password to what it wanted to perform app testing on a form.

The user wrote, I realized that's a little nonsensical, but long story short, it did a bunch of shit.

It wasn't asked to.

As I previously reported, in late May early June, both open ai and Anthropic cranked up the pricing on their enterprise customers, leading Replet and Cursor both shifting their prices upward.

This abuse is now trickled down to the customers.

Report has now released an update.

Unless you choose how autonomous you want Agent three to be, which is a tacit admission that you can't trust coding elms to build software replets.

Users are still pissed off, complaining that report is charging them for an activity when the agent doesn't do anything, a consistent problem I've found across redditors.

While Reddit is not the full summation of all users of every company everywhere, it's a fairly good barometer of user sentiment and man a user's piss and now here's why this is bad.

Traditionally, Silicon Valley startups have relied upon the same model, have grow really fast and burn a bunch of money, then turn the profit lever.

AI does not have a profit lever because the raw costs of providing access to AI models are so high and they're only increasing that the basic economics of how the tech industry sell software don't make sense.

I'll reiterate something I wrote a few weeks ago.

A large language model users infrastructural burden varies wildly between users and use cases.

While somebody asking chat gpt to summarize an email might not be much of a burden, somebody asking chat gpt to review hundreds of pages of documents at once.

A core feature of basically any twenty dollars a month subscription could eat up to eight GPUs at once.

To be very clear, a user that pays twenty dollars a month could run multiple queries like this a month and there's not really a way to stop them.

Unlike most software products, any errors in producing an output from a large language model have a significant opportunity cost.

When a user doesn't like an output, or the model gets something wrong which it's guaranteed to do, or the user realizes they forgot something, the model must make a further generation or generations, and even with caching which anthropic is added are told to there's a definitive cost attached to any mistake.

Large language models are for the most part lacking in any definitive use cases, meaning that every user is even with an idea of what they want to do, experimenting with every input and output.

In doing so, they create the opportunity to burn more tokens, which in turn creates an infrastructural burn on GPUs, which cost a lot of money to run.

The more specific the output, the more opportunities there are of a monstrous token burn.

And I'm specifically thinking about coding with l elms.

The token heavy nature of generating code means that any mistakes, suboptimal generations, or straight up errors will guarantee further token burn.

Even efforts to reduce compute cors by, for example, pushing free users or those on cheap plans, the small or less intensive models have dubious efficacy.

As I talked about in a previous episode, open ai split a model in the GPT version of CHET.

GPT requires vast amounts of additional compute in order to route the user's request or the appropriate model, with simpler requests going to smaller models and more complex ones being shifted to reasoning models, and it makes it impossible to cash part of the input.

As a result, it's not really clear whether it's saving open ai any money, and indeed, kind I suggest it might be costing them more.

In simpler terms, it's very, very very difficult to imagine what one user free or otherwise might cost, and thus it's hard to charge them anything on a monthly basis or tell them what a service might actually cost them on average.

And this is a huge, huge problem with AI coding environments.

But let's talk about claud Code again.

Anthropics code generate a tool.

According to the information claud code was driving nearly four hundred million dollars in annualized revenue, roughly doubling from a few weeks ago on July thirty first, twenty twenty five.

The annualized revenue works out to about thirty three million dollars a month in revenue for a company that predicts it will make at least four hundred and sixteen million dollars a month by the end of the year, and for a product that has become for a time the most popular coding environment in the world from the second largest and best funded AI company in the world.

Is that it is that fucking it is that all that's happening here thirty three million dollars, all of which is unprofitable after it felt, at least based on social media chatter and discussing with multiple different engineers, that claud code have become ubiquitous with anything to do with LLLMS and coding.

To be clear, Anthropics, so on It and Opus models are consistently some of the most popular for programming an open router, an aggregator of LM usage, and Anthropic has been consistently named as the best at coding.

Whether or not I feel that way is irrelevant.

Some bright spark out there is going to send it.

Microsoft's get hub copilot at one point eight million paying subscribers, and guess what that's true?

In fact, I reported it.

Here's another fun fact.

The Wall Street Journal report that Microsoft loses on average twenty dollars a month per use, with some users costing the company as much as eight bucks.

And that's for the most popular product.

But wait, wait, wait, wait, hold up, wait, I read some shit in the newspaper.

Aren't these LLLM code generators replacing actual human engineers?

And thus, even if they cost way more than twenty dollars one hundred dollars or two hundred dollars a month, they're still worth it.

Right, They're replacing an entire engineer.

Oh my sweet summer child.

If you believe the New York Times or other outlets that simply copy and paste whatever anthropic CEO Warrio Ama Day says, you'd think that the reason that software engineers are having trouble finding work is because their jobs are being replaced by AI.

This grotesque, manipulative, abusive, and offensive lie has been propagated through the entire business and tech media without anybody sitting down and asking whether it's true, or even getting a good understanding of what it is that elms can actually do with code.

Members of the media, I am begging you stop stop doing this, Stop publishing these fucking headlines.

You're embarrassing yourself.

Every asshole is willing to give a quote saying that coding is dead and that every execut if he is willing to burp out some nonsense about replacing all of their engineers.

But I'm fucking begging you to either use these things yourself or speak to people that do.

I am not a coder.

I cannot write or read code.

Nevertheless, I'm capable of learning, and I've spoken to numerous software engineers in the last few months, and basically I've reached a consensus of this is kind of useful sometimes.

However, one time, a very silly man with an increasingly squeaky voice said that I don't speak to people who use AI tools.

So I went and spoke to three notable experienced software engineers and ask them to give me the straight truth about what coding lllms can do.

Now, for the purposes of brevity, I'm going to use select quotes from what these people said.

But if you want to read the whole thing, you can check out the newsletter first.

I'm going to read what Carl Brown of the Internet of Bugs said, and I had him on the show a few months back.

He's fantastic.

So most of the advancements in programming languages, technique and craft in the last four years have been designing safer and better ways of tying these blocks together to create large and larger programs with more complexity and functionality.

Humans use these advancements to arrange these blocks in logical abstraction layers so we can fit an understanding of the lairs interconnections in our heads as we work.

Diving into blocks temporarily is needed.

This is where AIS fall down.

The amount of context required to hold the interconnections between these blocks quickly grows beyond the AI's effective short term memory, in practice much smaller than its advertised context windows size, and the AIS like the ability to reason about the abstractions as we do.

This leads to real world code that's illogically layed, hard to understand, debug, and maintain.

Carl also said code generation AIS, from an industry standpoint, are roughly the equivalent of a slightly below average computer science graduate fresh out of school without any real world experience, only ever having written programs to be printed and graded.

That's bad because, as he pointed out, whereas llms can't get past this summer, in turn stage, actual humans get better, and if we're replacing the bottom rung of the labor market, there won't be any mid level or senior developers later down the line.

Next, I asked Nick Sharesh of I will fucking pile drive you if you mention AI again what he thought.

Llms, he said, will sometimes solve a thorny problem for me in a few seconds, saving me some brain power.

But in practice, the effort of articulating so much of the design work in plain English and hoping the LM emits code that I find acceptable is frequently more work than just writing the code.

For most problems, the hardest part is the thinking, and lllms don't make it that part any easier.

I also talked to Colvogi of no AI is not making AI engineers ten X is productive.

We also had in the show recently, and he said this, llms often function like a fresh summer intern.

They're good at solving the straightforward problems that code has learned about in school.

But they are unworldly.

They do not understand how to bring lots of solutions to the small, straightforward problems together into a larger hole.

They lack the experience to be wholly trusted and trust this is the most important thing you need to fully delegate coding tasks.

In simpler terms, lms are capable of writing code, but can't do software engineering because software engineering is the process of understanding, maintaining and executing code to produce functional software, and lms do not learn, cannot adapt, and to paraphrase something Carl Brown said to me, break down the more of your code and variables you ask them to look at at once, so you can't replace a software engineer with them.

If you are printing this in a media outlet and have heard this sentence, you are fucking up.

You really are fucking up.

I'm really neat members of the media here in this You need to change.

You need to change on this one.

You are doing software engineers dirty.

Look, and I understand why too.

It's very easy to believe that software engineering is just writing code, but the reality is that software engineers maintain software, which includes writing and analyzing code, amongst a vast array of different personalities and programs and problems.

Good software engineering harkens back to Brian Merchant's interviews with translators.

While some may believe the translators simply tell you what words mean, true translation is communicating the meaning of a sentence, which is cultural, contextual, regional, and personal and often requires the exercise of creativity and novel thinking.

And on top of that, while translation is the production of words, you can't just take code and look at it.

You actually need to know how code works and functions and wide functions.

In that way, using an LLM, you'll never know because the LM doesn't know anything either.

Now, my editor Matt Hughes gave an example of this in his newsletter, which I think i'll paraphrase.

He used to live in France and the French speaking part of Switzerland, and sometimes he will read French translations of books to see how awkward bits of prose are translated.

Doing those awkward bits requires a bit of creative thinking.

And I quote take Harry Potter in French, Hogwarts is boudlard, which translates into bacon lice.

Why did they go with that instead of a literal translation?

Of Hogwarts, which would be Verus Spork.

I'm sorry to anyone who can actually read languages, no idea, but I'd assume it is something to do with the fact that Poolard that Poudlard sounds a lot better than Veru Spork, and both of them, I can say flawlessly.

Someone had to actually think about to translate that one idea.

They had to exercise creativity, which is something that an AI in is inherently incapable of doing.

Similarly, coding is not just a series of texts that program as a computer, but a series of interconnected characters that refers to other software in other places that must also function now and explain on some level to someone who has never ever seen the code before why it was done in this way.

This is, by the way, while we're still yet to get any tangible proof that AI is replacing software engineers, because it isn't replacing software engineers, and now we need to understand why this is so existentially bad for generative AI.

Of all the fields supposedly at risk from AI disruption, coding fields or felt the most tangible, if only because the answer to can you write code with LMS wasn't an immediate unilater or no The media has also been quick to suggest that AI writes software, which is true in the same way that chat GBT writes novels.

In reality, lms can generate code and do somewhere some sort of software engineering adjacent tasks, but like all large language models, break down and go totally in saying hallucinating more and more as the tasks get more complex, and software engineering is extremely complex.

Even software engineers who can read code and have done so for decades will find problems they can't solve just by looking at the code.

And as I pointed out earlier, software engineer is not just coding.

It involves thinking about problems, finding solutions to novel challenges, designing stuff in a way that could be read and maintained by others, and that's ideally scalable and secure.

The whole fucking point of an AI is that you handshit off to it.

That's what they've been selling it as.

That's why Jensen Huang told kids to stop learning to code.

As with AI, there's no point and it was all a fucking lie.

Generative AI can't do the job of a software engineer, and it fails.

While also costing an abominable amount of money.

Coding large language models seem like magic at first because they, to quote a conversation with Carl Brown, make the easy things easier, but they also make the harder things harder.

They don't even speed up engineers.

There's a study that showed that make them slower YEAT coding is basically the only obvious use case for lms.

Oh, I'm sure you're gonna say, but I bet the enterprise is doing well, and you're also very, very wrong.

Microsoft, if you've ever switched on a TV in the past two years, has gone all in on generative AI, and despite being arguably the biggest software company in the world at least in terms of desktop operating systems and productivity software, has made almost no traction in popularizing generative AI.

It has thousands, if not tens of thousands of salespeople and thousands of companies that literally sell Microsoft services for a living, and it can't sell AI.

I've got a real fucking scoopyeo, I'm so excited, and I buried it in the third part of a four pot episode.

AAH and truly twisted.

But a source that has CM materials related to Sales has confirmed that as of August twenty twenty five, Microsoft has around eight million active license so paying users of Microsoft three sixty five Copilot, amounting to a one point eight one percent conversion rate across four hundred and forty million Microsoft three sixty five subscribers.

Must be clear that three sixty five is their big cash cow.

This would amount to if each of these users paid annually at the full rate thirty dollars a month, to about two point eight eight billion dollars an annual revenue for a product category that makes thirty three billion dollars a fucking quarter.

It's productivity and business unit for Microsoft, and I must be clear, I am one hundred percent sure these users aren't all paying thirty dollars a month.

The Information reported a few weeks ago that Microsoft has been reducing the software's price, referring to Microsoft three sixty five with more generous discounts on the AI features.

According to customers and salespeople, heavily suggesting discounts have already been happening.

Enterprise software is traditionally sold at a discount anyway, or put a different way, with bulk pricing for those who sign up a bunch of users at once.

In fact, I found evidence that they've been doing this for a while, with a fifteen percent discount on annual Microsoft three sixty five Copilot subscriptions for orders of ten to three hundred seats mentioned by an IT consultant back in late twenty twenty four, and another that's currently running through September thirtieth, twenty twenty five, with another Microsoft Cloud Solution Provider program.

Yeah this, I've found tons of other examples too.

A Microsoft three sixty five is the enterprise version where they sell things with like Word and PowerPoint and sometimes teams as well.

This is them probably the most popular product, and by the way, they even manipulate the numbers a little bit there.

An active user is someone who has taken one action on any Microsoft three sixty five app with Copilot in the space of twenty eight days, not thirty twenty eight.

That's so generous, now, I know, I know that word active.

Maybe you're thinking ed, this is like the gym model.

There are unpaid licenses that Microsoft is getting paid for.

Fine, fine, fine, fucking fine.

Let's assume that Microsoft also has based on research that suggests this can be the case for some software companies another fifty percent four million paying Copilot licenses that aren't being used.

That's still twelve million users, which is around two point seven percent conversion rate.

That's piss, poor buddy, that's piss, Paul, that's pissy.

It sucks.

It's bad, Doodoo.

Well I just said pp I guess anyway, very serious, very serious podcast.

But why aren't people paying for Copilot?

Well, let's hear from someone who talked to the information and I quote, it's easy for an employee to say, yes, this will help me, but hard to quantify how.

And if they can't quantify how it will help them, it's not going to be a long discussion over whether the software is worth paying for.

Is that good?

Is that good?

Is that what you want to hear?

It isn't.

It isn't.

That's that's the secret.

It's not.

It's bad.

It's really bad.

It's all very bad.

And Microsoft through sixty five Copilot has been such a disaster that Microsoft will now integrate Anthropics models to try and make them better.

Oh one other thing too.

Sources also confirm GPU utilization, So how much the GPUs set aside for Microsoft through sixty five?

Yeah, their enterprise codpile.

It's barely scratching the sixty percents.

I'm also hearing the share Point, which is an app they have with over two hundred and fifty million users, has less than three hundred thousand weekly active users of their copilot features, suggesting that people just don't want to fucking use this.

Those numbers that from August, by the way, and it's pathetic, and it must be clear.

If Microsoft's doing this badly, I don't know how anyone else is doing well, and they're not.

They're all failing.

It's pathetic.

But I've spent a lot of time today talking about AI coding, because this was supposed to be the saving grace, the thing that actually turned this from a bubble into an actual money minting industry that changes the world.

And I wanted to bring up Microsoft through sixty five because that's the place where Microsoft should be making the most money.

It's the most ubiquitous software, it's their most well known software, and they're not eight million people eight million people.

I've run that by a few people and everyone's made the same Oh God noise.

It's quite weird, the old God noise and the numbers.

But this just isn't happening.

Things are going badly and it really only gets worse from here, and I'm going to tell you more tomorrow in the final part of our four part Thank you for your patience and thank you for your time.

Thank you for listening to Better Offline.

The editor and composer of the Better Offline theme song is Matasowski.

You can check out more of his music and audio projects at Matasowski dot com, M A T T O S O W s ki dot com.

You can email me at easy at Better Offline dot com or visit Better Offline dot com to find more podcast links and of course my newsletter.

I also really recommend you go to chat dot Where's youreaed dot at to visit the discord, and go to our slash Better Offline to check out our reddit.

Thank you so much for listening.

Better Offline is a production of cool Zone Media.

For more from cool Zone Media, visit our website cool Zonemedia dot com, or check us out on the iHeartRadio app, Apple Podcasts, or wherever you get your podcasts.