Anthropic Cut 80% of Claude’s Prompt, Young Coders Vanish, Microsoft Snubs OpenAI

Hey everyone, happy Monday. Welcome to Techstrong gang. And so it begins again.

" But no, it's not going to be a slow week. No week is a slow week these days. But we've got some really good people to talk about everything that's happening.

Let me introduce you to them. I've got my friend from Hudson, Ohio, Steven Foskett. Looks like she's back home in Ontario, rejoining us after her holiday, our own Garima Bajpai.

And back from his trips to the wilds of Montreal and Quebec, our own chief content officer, Mike Vizard. Bonjour, Mike. How are you?

I'm doing really well, but my French is not any better. Well, okay. Didn't think it would be.

Anyway, it was an exciting week while you were away, Mike. We talked about a bunch of stuff on here. But then towards the end of the week, like last Thursday, Friday, and then over the weekend, just a new crop of really juicy stuff came out that we're going to hopefully, it'll take us all week to discuss, but let's jump into it.

Mm-hmm. Our first one is, this comes out of Anthropic and Claude about a new kind of engineer, because the world needs a new kind of engineer, don't we? Sure.

Just when you thought prompt engineering was going to be it, it's not. Well, and let's just dive into it, because we've been talking about context engineering off and on, but I don't think we've kind of sat down and put a bow around it. But Anthropic is basically talking about how if you improve your context engineering skills, you can reduce the amount of code that you need to write by as much as 80%.

Now, other folks have been talking about this issue for a little while now, and everybody who makes a graph or a database or some sort of indexing engine is now talking about their role as a repository for context engineering. And the idea is to rely less on the LLM for the processing, because then the AI agent doesn't have to stuff as much stuff into memory, and you might get better answers as a result because the AI agent is more focused on the task at hand versus all the scut work that goes with creating the task at hand. And then it also has the added benefit of reducing costs because now I'm not using as many tokens because I'm leaning more on some sort of platform out there to provide some context to the AI agent that hopefully is pre-processed.

In a lot of ways, Alan, this is not a new idea, right? It's really kind of like pre or co-processing in some ways for AI, and we've kind of seen this model before, but now we're calling it context engineering, and we're kind of moving beyond prompt engineering and I don't know. It looks to me like maybe this is where the weight of this argument's going to start to move and where the weight of the market's going to be.

It's a funny thing. I've made a study of looking at people's prompts. We'll have the same end goal.

We're going to write something about this, or we want to code that. How do you write the prompt? Right?

And I first, quite frankly, a year and a half, two years ago, I was introduced to the idea of a pre-prompt, right? Have the AI write a pre-prompt of the prompt you want it to write, right? And then write the prompt based upon that pre-prompt.

And I look at my output from AI versus other people's output from AI and Mike, I discovered this months ago, that I did not generally give the AI... I don't like giving it the formulas. I don't like spelling out every single thing.

I don't like the prompt this long, unless it's truly I'm giving it my words that I want in or something. And I think that's what the Claude people are talking about here. They didn't reduce 80% of the code.

They reduced 80% of the standard prompt that Claude code, being a program, that Claude code operates on. It's system prompt. And it's substituting that specific prompt for contextual background, right?

Things that are in its environment and everything else that you don't specifically have to load as that prompt in terms of, to your point about burning the tokens and how it works. I think in the long run, it's going to make it easier. You don't have to do these arcane prompts where you say, "I want you to do one and two and three and four, and use this and five for that and six.

" I think going in the future, that's not how people are going to prompt. And it's a good thing. Let me ask Garima this.

Part of this whole conversation about AI has always been around the notion that we're going to democratize this and anybody can do it. And yet, the more I look at this context engineering, the more it seems to me you need to know more about whatever it is you want to automate or whatever task it is you want to complete, because you have to understand where the context is in the first place to show that to the AI agent. So, is this really a skill or a new type of engineering, or how do you view context engineering?

To answer your question, first of all, you can't remove the software engineering part from software. It's very delusional to say that we will not be requiring software engineers, right? The nature of the job would change.

And I'll tell you what was the problem last year. Last year, we were having a problem of deficit of, let's say, software engineering skillset. So we have these agents who are heavy lifting and writing some kind of commoditized code.

We generated a different kind of problem. In the longer run, what we saw that the quality of the code was dropping. Then people started to front load and build some guardrails around the particular problem so that the quality doesn't drop and it doesn't result into 10,000 odd bugs at the end of the day.

So that is where we saw that when the front loading was happening, there was the cost of writing the code became a point of consideration. And a lot of enterprise teams, and I run with a lot of enterprise teams, were doing the same job of harnessing the agents to write the quality code, right? So that problem was visible at least in the beginning of the year.

Now, what Anthropic has done is that they are advocating, and there are several parts of this. So they are advocating that context engineering becomes a premium element. It's no longer that you should front load a lot of guardrails and a lot of heavy data and heavy kind of documentation upfront.

So the agents can do their job, but they are also providing some services to cleanse the code. So if you think about auditing services or cleaning services or the doctor service, which they are. So it's the rise of the next generation of services, and it's also the natural maturity of where software engineering would go from here.

Again, to answer your question that I personally don't believe that this will lead to removal of software engineering. It's just that the job is changing and context engineering has become part of your software engineering teams, right? So this is great.

I feel that this resolves the redundancy of creating these kind of structures in multiple enterprise teams, and it provides some kind of a scalable method to go about it. Stephen, you want to weigh in here? Yeah, absolutely.

And it's a really interesting story because I agree with the fundamental thrust here, which is that these models have really changed in terms of capability this year. And that means that the nature of prompting has dramatically changed as well. Let me give you some context here, though.

The reduction in the system prompt actually-- Well, first off, it's really specific to Claude Code, and it actually happened before Opus 5. 8 transition, which happened a few months earlier than this. 8, but it's way smaller than it used to be.

And the reason for that is that they got rid of a lot of the examples that they had in there and sort of the-- Well, I guess examples is the right way to put it. That prompted Claude Code on how to do things and how to respond, and instead, they're relying on the model to already have that capability in it. But I think it's important to note that the consumer version of Opus 5, which I think we're going to talk about here in a little bit, actually has a much larger system prompt.

So this is specific for coding. The consumer version actually has grown to as much as 200,000 characters as opposed to about 8,000 for Opus 5 on the Claude Code side. So that is a massive, massive prompt.

And if you look at what they did there, it's because they've included all sorts of safety instructions. So they decided if we kind of look at this, what we can see about the new model, the latest models from Anthropic, is on the one hand, they don't need as much examples. They can pretty much do things on their own.

On the other hand, Anthropic is putting way more effort on safety and putting constraints around these models than ever before. And I think that's pretty interesting, too. So Alan, I don't know about you, but I'm getting inundated by vendors who are all positioning whatever it is that they have as a context engine or in a context engine platform of some type or another.

So do you think that this is going to be a market, a standalone kind of thing that sits outside of AI, or is this just an extension of the AI models themselves? Or how are you viewing this whole thing? Well, I think vendors have stopped pitching me because I just don't answer them.

No, I'm only kidding. " Well, no. I think when they're talking about context, I think they're talking about what is sort of innate, built into the model, the environmental variables Anything else is a prompt.

And yes, there are vendors who are jumping on the new term because what vendor doesn't like to own a shiny new term? Mm-hmm. Right?

But I think context engineers are not prompt engineers with another name. They figure out how to use the context that the models have built in to accomplish the mission with shorter prompts, quicker, better, faster, more secure. I think Steven's point to the security of the, or the guardrails, I guess I'd call it, guardrails to the consumer version's a good thing.

I think that's separate from this context. I think that's just Claude and the Anthropic folks trying to avoid lawsuits and getting shut down maybe. But I don't think these companies who are pitching context platforms, I don't see them staying around.

I'm not sure I agree, and here's my other thought process on this, and I'll throw this to Garima. But, if I lean too heavily on an AI model for all the context, I'm going to get further and further locked in. And if I have some other platform out there that's providing me some context, I can mix and match my LLMs a little more easily because I don't have to make sure that each LLM has the same level of context about a thing.

I'm actually exercising more control over that, and I can bounce between LLMs as I need to, maybe. Yeah, I think that is correct. And context engineering has to be made interoperable.

It doesn't have to be tied up with one model, one... Because we also, as enterprise teams, are evaluating not only the models like LLMs, just capability and all that, right? But we also are evaluating our context, and which context works well with which model.

Sometimes a proprietary model would work better than a generic commoditized version, right? And we probably wouldn't like to feed everything in the commoditized version of an LLM. So this is a choice which every organization is making, and I think it's a good segue also to the next section, which we will actually discuss about what Microsoft's strategy is to kind of stay in-house and be more proprietary based.

Right? So we will talk about that, but it also reflects that challenge. Mm-hmm.

As we look at all this for a minute too, it's... And let me toss this over to Steven a little bit. If I look at all these players and how they're going to connect, there's this phrase out there you start to hear now where it's LLM routing or AI model routing, and is this part of this whole motion here?

Because now I'm constantly finding ways to separate my context and my content from the underlying LLM, and this too is how I'm going to exercise some control over all this stuff. Oh yeah. I'm glad you brought that up because the model routing is just a huge topic.

I've been using Open Router myself and experimenting with some of the open source routing concepts as well, because frankly, it's really the only way to- Mm ... to go when it comes to the financial aspect here, but also in terms of the sustainability aspect. And I don't mean that in an ecological sense.

I mean that in a sort of a reliability sense. 5 Flash. Well, that doesn't exist anymore.

So now, we've got to use something else. So a model router helps me sort of abstract away, sort of like virtualization abstracted away from the server. It allowed me to abstract away from specifically calling this model.

And then the other thing that I'm using it for is having it make a judgment on which model do I need for this particular type of task. Is this a simple tool use task that we don't need to use a frontier model? Is this like a summarization?

Well, then we can use kind of a flash, short, quick, easy model. Is this more something that needs some research? Well, then we need to use a model that has more capability to look things up.

And I think that that is really going to change the way that people use AI models going forward. And frankly, what that means is that context is the only thing that's going to matter. Because as you're using all sorts of different models, you have to make sure that the model that you're using has the right information in order to give you the right result, or it's just not going to work.

So, but what we're talking about here, let's be clear, these are harnesses, right? Mm-hmm. They're contextual harnesses which allow you to do that.

I spent a lot of time this weekend thinking about, do I want to start using some of these open weight models? What will it take to actually set it up, or do I use it from a hosted system, or what have you? Unfortunately, we're out of time for this segment, but Mike, that's something we probably want to come back with, is there's a richness of models out there now.

A lot of them are open weight. How does, not the big enterprise who has, or an engineer who could set it up, but how does the average Joe- Wow ... going to use this?

This is true, but I think collectively, when I look at all this stuff, I feel like enterprises are getting savvy quicker than usual about not getting locked into an AI model provider. I know. They're getting clobbered with token costs.

That'll smarten them up real quick. But we really got to jump, Mike. Speaking of models, that's the next segment here, isn't it?

I think so. Let's go to the next segment here. So as we look at what's going on here with a bunch of other things, but first we're going to talk a little bit about coding and the agent, a blog post that you have, and then we'll get to Microsoft and- Oh, okay ...

and some of the other folks that are going around out there. But I wanted to kick this one to Garima first. Allen has a provocative post talking about how, well, there are no under 25s starting to show up in the application development world.

And is that going to be a crisis later on when we don't have enough people who know what's going on? Or is the fact of the matter is, is that professional developers have just taken more responsibility for the entire workflow and when the new ones come in, they're going to have to learn the same thing a lot faster than they have historically? I guess, what's happening with young developers, Garima?

What are you seeing? So I think, in the short term, Allen's article is spot on because we are seeing a lot of de-skilling with the emerging talent. The talent pipeline is shortening because there is a few factors to it.

The first is that the technology is moving faster than what we can think about, consume, and adopt. The second aspect is that our education system is not modeled for this change, and I'll talk this in a while. And the third aspect is that industry is kind of gearing up for optimization, so they are optimizing their tech talent based on what work agent can do.

The agents can write boilerplate code. So the commoditized or the heavy lifting of how code was written in the past is going to the agents. So there is definitely kind of a shortening talent pipeline, which is required from the industry perspective.

So these three factors influence how we see the ecosystem shifting. So there is not enough interest in the younger people to go into this kind of profession because they probably see that there might be the lucrativeness of the software engineering career is kind of in a limbo. Now, from my perspective, and this is only my take on this, that this is a short-term or a transit kind of view on this.

I think what could happen, and again, I am a software engineer at core. I do not see that software engineers would be removed from the ecosystem. Is this just that the job is changing?

And trust me, I started with software engineering when I used to look at Fortran and COBOL people and aspire to become like them. But my job has changed. I took that job from there to where GitHubs and DevOps and cloud transformed a lot of our jobs.

And now AI and AI agents are doing that. So I feel that what will happen is that the nature of the job is in itself transforming. So we'll have to accept that fact.

We also have to introduce people earlier to software engineering so that they understand the basic pieces of the puzzle and then start teaching. In academics, there has to be a larger shift that what software engineering could look like. And probably I would love to design a course for the next generation of software engineers, where the code generation to goal orchestration becomes one subject or one area of due diligence, how intent and constraints define how you write code, how you validate code, how you execute code.

Then context engineering. We have been talking about this in the previous section, how prompt overloading is not required anymore, so how context engineering will transform this landscape. How agent-first infrastructure would look like.

How agent orchestration environment or agent execution environment could look like. Those are kind of things and there might be many more. So there has to be an academic shift and the talent pipeline has to be geared towards being goal orchestrators, reviewers.

The nature of the job is changing. The commoditized boilerplate code, nobody cares about. So the agent can write it.

So that's how I feel that the transformation would happen. Again, I would love to bring a course which probably talks about this and get the next generation of pipeline prepared for this. So Allen, you wrote this piece.

Are we looking at what maybe what Garima's point is? Is this just a lost generation of developers and the next wave coming out will have more of these prompt and context engineering skills? Well, and so let's call them software engineers and not coders.

I think that's the key piece of this. They're not coders, they're software engineers. There's two things here that I find fascinating.

One is how quickly word gets out to the kids. All of a sudden, I'm reminded of the Little Rascals. " And Stymie tells, Spanky tells Alfalfa, Alfalfa tells the other, and before you know it, none of the kids are eating the porridge.

Well, word's gotten out Don't do that computer science stuff because those jobs aren't there. And so at the same time that you're seeing this lack of hiring of under 25s, the under 25s are responding by going elsewhere, right? To whatever the new hot thing is.

I don't know if they've quite caught onto contextual engineering, quite frankly. But this goes beyond software. Mm.

" As AI, I'm not saying people losing their jobs or firing you because you're being replaced by AI, but as AI takes on the role of the junior, or not the role, but takes on the task of a junior person, you hire less junior people to do those tasks because AI does it, and it's literally hollowing out the economy. And so the older people who are there, they continue to go on with their careers, but eventually they die out. And then there's just a hollow nothing in the middle there to replace them.

And I think that right there is the issue. Right now, we're seeing it with software as the tip of the spear, but you're going to see this in any vertical that you look in. You're going to see it with lawyers.

Right. You're going to see it with accountants. So- Doctors ...

" So he's just ignoring it and going to work as usual. Right. There is, look, people, especially young people, because you know they're going to live forever anyway.

And so they don't worry as much about that day-to-day stuff, perhaps. Yeah. And especially, it's out of their control.

You can't worry about things that are not in your control. But as you get older, you sit here and pontificate on this kind of nonsense, and you put your thoughts into it. I do think that this hollowing out of the economy is an important thing we've got to be thinking about.

So Steven, I don't know what your take on this is, but as I look at it, I sit there and I wonder, am I looking at a world where everybody's a developer, but nobody's a software engineer and so we just wind up with a lot of more code than ever being created that no one knows how to push into production? Well, it kind of seems like that, doesn't it? You look at the amount of code that's being created, and it looks like a, yeah, absolute flood of code, but very few people who actually are focused on efficiency or even functionality.

Essentially, you've got AI writing the code, AI testing the code. Now you've got AI writing the specs. But it's a really funny situation here.

There's another angle that I want to call out. " But it's interesting that Patience actually has the same point that Allen has, in that, yes, AI isn't coming for your job yet, and maybe never will, except if you are an absolute rank junior developer. Which he says, yeah, it is absolutely coming for that job, and that seems to be what the data is showing.

Yes. But I actually want to point out that there's another thing that happens here, and that is that senior developers, and they may want to throw things at me when I say this, but senior developers may not be as good as they think they are anymore. Now, there's a lot of research that has been done in many different fields, specifically a lot of scientific research in the aviation and medical fields, showing that supervisors lose the ability to properly do the work once they transition 100% supervisory.

Now, I don't want to throw any stones here at Allen or anybody else on this, but I personally feel that way as well. I have forgotten a lot of the things that I used to know how to do, and that's actually happening in software development, too. And what that means is at the same time that we're losing the people who are doing the real nuts and bolts work, we're relying increasingly on people who may have forgotten how to do that work.

And before you think I'm blaming old people, like I said, this is very well documented in very empirical research. It actually came out of, I don't know if you guys remember, the Air France, I think it was 447 crash of 2009. When a jet crashed and they said, "How could this have happened?

" And the answer was they forgot the basics of avionics. We lost the young people who were supposed to know this stuff, and so nobody knew it. And I'm really worried that that could happen in software development as well.

I got one thing. Steven, get off my grass and go play in front of your own house. Hey, I'm using science here, buddy.

No, but wait. Go on, Tim. " Which means they stopped coding at some point in their lives.

Yes. And then they all came back in and started coding again because they felt like AI would let them do that. " I want to add something here because this surge of software development you saw in the last decade.

This code abundance came through because everybody became software savvy. But software engineering as a discipline, as a field, computer science, computer engineering, is much more than that. So I think we should not confuse ourselves with the transformation of the job, the code abundance, and the commoditized version of software developers.

I think there was a rise, there was a phase where it was needed, and then there will be a decline. And then history reminds us when COBOL and Fortran were languages, why did C++ and Java come in? Because the complexity of the software was increasing.

So we need to have another kind of dimension to how do we do software. So similarly, I think things will change. And these professions will be changing, but they're not completely displaced.

We need software engineers, computer scientists, to kind of shift the needle in the right direction. So let me just ask you one last question on this, Garima, because you alluded to it, but I'm curious. Essentially, the education system that we're using for bringing new people into this field, does that need to be revamped?

Do we need to go back to all these universities because they've been teaching them yesterday's game, and we need to just revamp that entire program? Definitely. And I think some of the universities and researchers are looking at it.

There is a new term which has been surging in the market, which is high-leverage symbiotic software engineering. This is a new discipline, which is rising from here. And definitely we need some people to come together to look at the academic side of it.

And I again say that we'll have to introduce software engineering and the coding skills from the beginning, initial stages, so that the commoditization and the commoditized version is everybody's aware of it. So it becomes a no-brainer. It becomes the literacy quotient for everyone, right?

Mm-hmm. Guys, we're over time, Mike. I got to keep us moving.

I'm sorry. No worries. But some kid somewhere is going to be suing some university soon because he can get an ROI on his education, so there you go.

Ah, they have all the disclaimers on that. All right. So shifting gears here now, and this is a continuing conversation for the gang, and a lot of different people have been talking about it for the last week.

" And that seems to have also led to Anthropic kind of rolling out some new models that are, shall we say, more competitive. And that as we look at that, they're easier to use, they're less expensive, and I don't know where that tagline underneath came from, but we'll worry about that another day. But so my question to you, and I'll start with Steven, is are we on a full-boat kind of war here for market share and old-fashioned slug 'em, knock 'em out kind of competition?

Absolutely, and I'm here for it. I'm excited. This is awesome.

Let me say, this is so great. We talked previously about Meta rolling out lower-priced models that are competitive with the Chinese models. Obviously, the Chinese companies have been really the cost leaders here in competitive models.

Google has been right there for a long time. We've got Microsoft, and now we've got Anthropic in the mix. I got to say, I've been using Opus 5 myself since it was announced, and this is a very, very good model, and I love knowing that it's a lot cheaper.

And why is it cheaper? Well, all these models are cheaper because they simply use less tokens. 5 Pro, showing that it's using as much as 89% less GPU power to generate essentially the same or better result than previous models.

This is incredible. This is great. This is exactly what the industry needs to do because a lot of the...

Well, some of the concern about AI has been the incredible impact of developing and deploying and using these models in terms of power, space, cooling cost, et cetera. And that eventually is going to roll downhill and is going to come to consumers. And so when people, whether they're business or individuals, when people go to use these models, eventually they're going to have to pay the full boat price, and they're not going to like what that price is because it's pretty hefty.

Unsubsidized AI model cost is pretty high because these things use a lot of power. They take a lot of equipment to run. And so I think all of these companies are scrambling to reduce the impact of these models in order to better deliver something that matches the consumer's demands.

If a company made, I don't know, a really good car that cost them a million dollars to make and they sold it for $30,000, everybody would run out and buy one. But the company can't exist doing that for very long before they run out of money. Oh, they make it up in volume.

What happens when they try to cut costs, et cetera, eventually it has to reflect the real cost, and that's going to happen to AI. And before that happens, they better figure out how to make it something that consumers would be willing to pay on a cost-benefit analysis because otherwise this whole industry just goes p**f. So Mike, Steven, Garima, this is exactly what my upcoming book, "The Indispensability Trap," is about.

In order for something to become indispensable, it has to be so widely available. Even like we talk about electricity, you had to run electricity out to every farm. It couldn't just be in the city.

Same thing with phone service, same thing with railroads. It had to crisscross the whole country here in the US, Canada. It's the same thing with AI, right?

It's Jevons paradox, which is the more popular something gets, the more it gets used, right? And in order for that to happen, the price has to come down, because if the price stays high, it'll never reach its goals. So that's exerting this force on it.

The other force that's being exerted is specific to AI and technology, software, is the open... It's not open source, it's open weight. It's open weights.

But these open weight models are changing the game. They're changing the game. They're forcing Anthropic and OpenAI to say, "Hey, we've got to compete," because Steven, as you said, for 90%, 95% of the tasks we use AI for, I don't need Fable 5.

I don't need Five Six Soul. And then the other piece of it is, quite frankly, the open weights are as good. There's this much difference between the couple of models that came out of China last week and OpenAI's flagship model.

So are you saying that in the absence of some open weight models, that they're essentially acting as a counterbalance against providers of preferred- Well, there's four or five counterbalances. One is the open weight, open source. The second is government.

The government is saying, "I will bring an antitrust suit. I will take 50% of your stock, in spite of you offering me 5%. I will declare you a monopolist.

I'll do things. " So the government is a huge one. There's open source.

There's the competitiveness of the market, right? Microsoft's a great example. They own a huge chunk of OpenAI.

They signed a deal with Anthropic to distribute that, and now they're saying, "Yeah, we have those two, but let me give you mine here. It's cheaper. " That's the market at play, right?

Just regular market forces. And then you have the technology moving forward. We look at what we were using from Anthropic and OpenAI a year ago, and we thought it was cool back then, but it's not so cool today in light of what we've got.

Yeah. So all of these things put pressure. And then there's Jevons paradox.

The pressure is the more we use it, the more we're going to use it. So Steven, back in the day, if I wanted a better price from somebody, I put a mug from their competition on my desk. So is the same thing going to play out here?

Are we all going to get open weight coffee mugs and put them on our desk and force Microsoft and Anthropic to give us a better deal? Well, I don't know about you, but I would never fall for a logo coffee mug. Well, I think that the interesting thing is back in the day, if we wanted to switch providers, especially in the IT field, it was heavy lifting.

Hard. It was very difficult and that's one reason that companies like Oracle are still so dominant and they're still making so much money, because it just would be absolutely impractical not to use them. Now, even setting aside anything about the relative merits of the products, but with model routing, we are actually in a very different world.

These are commodities, and especially if you've got a router built in, these commodities are... You can switch instantly. You can switch transparently.

You can even switch... Imagine if you could have, I don't know, a bicycle, a moped, a motorcycle, a cheap car, an off-road car, a minivan, a pickup truck, and you could just literally drive the bicycle to the store and drive the pickup truck home. That's what you can do with these models, and that's pretty amazing.

And that gives the user a lot of power, and I think that also undermines the whole situation for a lot of these companies, because if there's very little friction in switching, and there's very little friction in choosing a better model, then people will. Especially when, like I said, when the bill comes due. That's why a lot of companies are doing what Microsoft and OpenAI are doing here, which is that they are, I'm sorry, Anthropic.

They are specifically building a router into their product. So Microsoft, when you go to use it, it's going to decide which back-end model to use dynamically, and they're hoping that you won't make that choice on your own. The same thing with Claude.

ai, you make a request. But yet, Satya called Claude out on that. Yeah.

It's going to be using a different model depending on what you're requesting, depending on what its choice is. And again, that's more about optimizing the back-end cost Not the front-end price. And so keep that in mind too, that a lot of the time when we talk about these things, there's a difference between cost and price.

In fact, when it comes to AI, there's a huge- But I want to insert Paul as well on this, because this is good that they've realized that routing and creating an ecosystem of different tiers of model would be great. But I am also seeing that there will be a surge in services companies where they will offer outcome as a service, which is basically, now we are paying for these massive models and all these things without any kind of promise on outcome. If your car is having a mileage of 60 kilometers per hour, it has some kind of weightage on it, right?

But what if we are using these models? There is no commitment, there is no SLA, there is no outcome. So I think this is a great opportunity for services companies to pitch in and say, "Okay, you don't care about which model we should use for what kind of services.

" So I'm simple-minded, and so let me ask Alan this question. But so let's say I'm Anthropic and- Go into the simple mind here. Go ahead.

But if I'm Anthropic and I'm open AI, and I'm investing in all these more advanced models, and they're expensive to build and train and yada, yada, yada, how do I justify that investment if, on the other hand, there's pressure to cut costs and price is starting to increase dramatically, to Steven's point, and people can route between models and anything they want. So how sustainable is it to be at the top end of the AI model market when maybe you just want to be at the good enough AI model market? So let me give you my best "The Graduate" movie impersonation.

Mike, I got one word for you. Oh. Upstack.

Both OpenAI and Anthropic know they got to move upstack. Models will be commoditized. I think the real question we got to ask ourselves is why the hell are we spending $8 trillion- Oh ...

on data centers to house these models that are being commoditized when most people are probably going to want them running on the computer on their desk or their phone in their pocket or somewhere else, and we're spending $8 trillion, the biggest capital raise investment that we've ever done as a civilization, into something like this. Right? And it's going to be commoditized because it doesn't reach its goals without being commoditized.

The real money sits on top of that. What are we going to do with all this intelligence? And that's why you see Claude coming out with Claude Code and Claude Design and OpenAI Office and OpenAI Health and all these other things.

They're desperately trying to climb out of that quicksand that is the indispensability trap. It traps you down here. On that note, though, we're out of time.

So, great conversation though. Garima, it's a pleasure having you back. We missed you.

I know you had a good holiday, but we missed you here. Steven, it's always great having you here Monday as usual, any other day we could get you, and Mike, we're glad you're back, and though you may not have gotten any better in French, you still do a good job here on the Gang. We hope you've enjoyed watching Techstrong Gang Taz.

Looking at the numbers, it looks like we had a nice little audience, and we'll continue to get more. tv, our Techstrong TV YouTube channel, our Techstrong TV OTT app, which is on iOS and Android and Roku, Apple and Amazon, any screen you like to watch this stuff on. You can also catch our Techstrong TV.

Steven, what about a Tech Field Day? Well, it's the summer. I know it's the dog days of summer, yeah.

So we don't have as many Tech Field Day events going on, but we do have some coming up, and I am pretty excited. We've got Cloud Field Day coming up here in September. We've got a really jam-packed networking field day.

In fact, I don't want to give any spoilers, but that may be a more than three-day event even. And we're also going to be at some events coming up. I'm personally looking forward to Supercomputing, to seeing you all at KubeCon and re:Invent and things like that.

So yeah, it would be really great to connect there. And of course, it'd be great to have Garima, Mike, and Alan joining us on the Tech Field Day delegate panel as well. Always.

Tech Field Day must be an NFL team. They only play in September, right? Training camps are open.

It seems like that right now. They're open. All right.

Until tomorrow. Thanks for watching. You've just watched the Gang.