Bloomberg Audio Studios, Podcasts, radio News.
Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Wassental and I'm Tracy Alloway. So, Tracy, you're cool like if I, like, you know, just start doing this part time as I like build out my software business, right, Like, you're cool about that.
I was gonna say, I've been thinking about AI and productivity, and so far your productivity has gone down, Joe, because instead of doing Odd Lots things, you're coding your own software.
Except that I'm creating content for the Odd Lots newsletter about coding, and that is productivity a creative debatable debatable, But but you're cool with that. You're cool with like me like, oh, I'm just gonna like check in part time on Odd Lots when we have a recording.
Well, so of course not okay, good, of course not.
Yeah, that's the right answer. I want you to yes, I want you to be really sad. But like a few other people, you know, I have like caught the sort of like bug of like AI coding and I'm totally blown away. I've like played with it from the big I started playing around with it last year, but then over the holidays. I've been writing about this in the newsletter. Suddenly, like my Twitter feed is like clud code, clug code, clud code. I do just cursor before, which
I was very impressed by at the time. And so when I got home from vacation, one of the first things I did is like figure out how to install clug code on my computer. And I was like, oh, I am like hooked. This is actually like I see why I have my Twitter feed is just like people posting about this.
All right.
So I have to say I have not tried it because I only have a work computer and I can't install new software, and I probably definitely cannot install new software that then makes changes to.
Your existing software.
I don't think Bloomberg would like that, but I have seen the hype. Lots of people talk talking about it. Have you seen claud Cowork Have you heard.
Of Oh yeah, yeah, yeah yeah.
So one of the criticisms of claud code was that you know, like, okay code, but you still need some background knowledge in coding, because like, you know, the interface is kind of like it pies and all of that or nineteen nineties Cowork apparently like goes a step further for normal people in quoting. It makes it super super easy. And the funniest thing is that apparently cloud code actually coded Cowork.
So this is like really relates to my experience last year and then this year, which is that even last year, like trying to use the AI coding tools, it was an annoying process because there are various things that you had to do in the actual command line of the computer that were like I didn't I don't know command line vernacular, and you have to install these libraries and stuff like that. So there was this sort of like
barrier that existed. But what's really changed in the last year or with the with Claude code, which has actually been around for a while and I should have like played with it before, is that like, because it sits on your computer, it sort of takes away it de abstracts it and so when you talk.
About like that it actually does the stuff.
It does it it just like oh, it's like, oh, we're gonna need to install this open source natural language processing library. It just does it automatically. Instead of me trying like figure out like what are the right keystrokes to pull that in or why is this not going
into the right file? Folder or whatever, and so like what like Cowork, it's like all like all of these sort of like little frictions, like these technical things like command line user very rapidly are like dissipating and so that like then you have something like Cowork where it's just like they know they're taking care of that, and so you get this like user interface that's just like it's just getting easier and friendlier. There's almost no technical frictions at all anymore.
Also, it feels very iterative, like the code is improving upon itself at this point I think was one of Claude's main selling points.
Well, this is like you've seen like people talk about like, oh is AGI here? And this is like part of the debate because the premp one of the ideas I guess behind AGI is like, well what happens when you have software that can train itself and so forth? And I don't really know if I buy that, but you do just see like how fast the iteration cycles are, and I think we want to get into this. In part, they're fast because a bunch of people are suddenly getting excited.
So then the human provides this sort of like we're sowing the seeds of our own demise because we're so enthusiastically participating in the evolution. But I just like it's suddenly clear, like, oh, this is going to change I think computing. And the other thing is the code works, like it creates code that like this is like there's no bugs. You know it works.
Did you see speaking of automating yourself, you see there was a post on Reddit from a lawyer who said he's basically used claud code to automate like his entire job, and he hasn't told anyone.
I'm not exactly surprised because the other thing that I experimented with is and I haven't one hundred percent verified this, but on jobs Day last week, I downloaded the full pdf and I just typed into the cloud code like find the most interesting details and make some charts based on and it did it in like a couple of minutes. I have no like ability to like I've never like
built charts myself or whatever, like designing or whatever. And I didn't totally confirm yet that the data was all correct, but I'm pretty sure it was because everything I spot checked, so I didn't just that crucial I didn't tell you I know I didn't. That's why I didn't want to like, oh, like here's what, here's the today's jobs reporting charts.
But I did.
So what application did it actually build it in the charts?
I don't know. I just had a file like that's the thing. I had a file on my computer at that point. What kind of file, like a P and G file, like an image file, that's the crazy thing. I don't know. And so there was just this image that had a bunch of charts, and my spot checks did suggest like I didn't see anything off. And people get paid money to like build that kind of stuff for like analysts and stuff like that.
And right, so this is the other big question. If everyone can build their own software, what actually happens to software? And I was reading something I forget who it was by, but someone used claud code to create they wanted a website that would basically make the money for doing nothing, and that was the prompt And did.
They do it?
Yeah.
So the idea that the model came up with was you can sell prompts, packages of good prompts and sell them for like forty bucks and you'll make tons of money. And I was thinking about that like, Okay, it's possible to make money that way, but also why wouldn't I just use claud code to do the same thing.
There are many big questions that we use an economy are going to have to think about, and I think my main takeaway is we're gonna have to think about these sooner red than later. But what is clud code? Why is everyone so hyped about it? Like what is it about this particular piece of software versus what exists from open Aye and Gemini and all this stuff, Like why is this captured everyone's imagination? We really do have the perfect guess because it's someone who, unlike me, has
been getting their hands and the stuff for longer. One of the few people that I know who is into lollms before chat GPT existed and was actually using them via the API, and was actually talking about their technical capacity to do things like coding even before November of twenty twenty two. So truly the perfect guests We're going to be speaking with Noah Bryer. He is the co founder of Alefic, which is a consultancy that helps big
companies deal with AI stuff. So, uh, Noah, thank you so much for coming on odd lots.
Thank you for having me.
What is eleven? What's the deal? How are you like using llms before chat GPT existed? I don't know. I know very few people who were doing that.
I had the good fortune of shutting down a startup in twenty twenty two, and so I had a lot of free time on my hands.
And then how are you using it though?
Like?
How did you look at your like how did you aware that there was this thing that could be of potential used to.
What was So my very first thing I was doing was using gethub Copilot, which at the time was built into VS code, and it was autocomplete inside VS code. So it was a nice and pretty immediately realized that there were certain coding tasks that it could just handle completely. Anything that was very pattern based. So if you write code, you write a lot of tests. If you write tests, every test kind of follows the same pattern, and you
want it to follow the same pattern. You're looking for that structure, and over time, because it was looking at your code base, it was able to basically autocomplete it. I also started playing with the GPT three API, which had come out. I think that came out in November of twenty twenty one, and that was the first time it was publicly available to everybody, and they had a large language, Boddel as we know it today, available to them.
So I was just testing and building things, and I pretty immediately realized the very first thing I did where it just blew my mind was I built a web scraper. So I was just trying to pull pricing data from a website. And I've done a lot of this in my career. It's maybe the most annoying task you have to do in all of coding, because HTML is the
most miserable language to have to parse. And I just had this thing where I took the page, I took the content, I took the text, and I gave it to the AI and I asked it to give me back the pricing table and give me back the pricing table. And I just thought, I'll never do it the other way again.
That's it.
Yeah, that HTML mentioned just brought up like memories of me in like the mid nineties on HTML goodies. Do you remember that side? Yeah? I wonder if it's still Is it still up? That would be wild? Does cloud code? Does that count as AGI? This seems to be the debate, right is it AGI?
I try not to wait into what's AGI and what's not. I think my guess on AGI, for what it's worth is that it's probably going to be a conversation like the Turing Test, where everybody thought it was really really important for a really long time. We thought the Turing Test was the biggest thing for seventy years or whatever, and then CHATCHYBT very clearly passed the Turing test, and now everybody pretends, like we it's not just that they forgot,
they pretend that it never mattered. Oh and so I am kind of guessing that that's going to be what the conversation. It's like, it's just going to be a sort of forever moving goalpost, because it turns out that the idea we had for what general intelligence looks like is not quite that. But I also think, you know, the computer scientists and the sort of serious AI researchers would say that much of what's going on inside claud
code is not the model itself. It's the model paired with a human and I think that is a pretty important distinction.
But I don't know about AGI.
Well, okay, so you were using GPT to code prior to the release of jag GUPT. So therefore, coding models have been around a long time. So what is for those who haven't played around with it? What is claude code? Because again, coding models have been around for a long time. Well maybe have heard of cursor, copilot or some of these other harnesses, et cetera. What is cloud code?
So if we back up first and we go to copilot.
So Copilot was the first sort of commercial application of a largelanguage model by most accounts, And what Copilot did in its initial instantiation was just.
Auto Microsoft product.
It's a Microsoft product, so Microsoft doones, get hub, getthub, develop copilot. It was Microsoft had the partnership with open Ai, and so they built it in and what it was doing was doing autocomplete. So if you're writing code, a lot of writing code is boiler plate or trying to
remember the name of a function. And you know, the reason stack overflow existed was because you can never remember the exact name of that function or the exact rejects that you need to use in order to find and replace something, and so you would go search for it. And they realized that you could just build that into the ID your code editor and and have it autocomplete
for you, and it was pretty amazing. Then, uh, Chatchipt came out, and even before that, I had built a simple chatbot for myself because I realized that, hey, I could just ask this and instead of going and searching stack overflow. It was totally capable of answering code questions, and it was capable of writing rejects or doing these
things and didn't make mistakes. Yes, but like there's famous mistakes on stack overflow of incorrect rejects that now exists in every code base in the world, and so you know, I think there were a lot of us just kind of playing with these things and realizing they were a huge boon. And so I think, really the next step is Cursor comes out. And the thing Curser realized that co Pilot didn't was that it wasn't good enough to
have autocomplete. You also needed the Q and A because you have these things that you can't just auto complete. You want to be able to ask the question and answer it. And then Chatchipt came out and everybody was switching between ID and then I think, really the next
big piece is that claud code came out. And what claud Code did that was so remarkable was they took the same set of models really and they took them out of the chatbot and they really just gave it some very basic functionality to operate within your machine.
Right.
And so you know, if you really look at kind of what exists within claud code, you're calling out to a model, and they gave it capability around sort of two big things. One is you can read and write files on your computer, and then two is that you can operate Unix the base commands, the bashed commands that
exist in your environment. And again, because these models were trained on the Internet, and there's no greater source of information on the Internet than how to make the Internet, they know how to use Unix commands incredibly well, right, because Unix has existed for whatever it is, sixty years, and the way these commands were designed, they're all designed to be very very simple. There's a find command, and you know, there's some thing called re rep and it
can search to a code base. And Unix has this sort of beautiful way of tying one command to another, so you can take the output of one command and send it to another. And they kind of just gave the model access to these two or three very simple things, and it kind of turned out that it unlocked a whole bunch of functionality. I don't think even the people who built it fully realized. Like one example that I think about a lot is just the challenge you have with all of these.
AI models is that they're stateless.
So every time you talk to chat GPT, it's sending your entire conversation history back to chat GPT because it has no saved history of that chat, right, And that's fine, it's the way it works. It's just fact. But it means that you know, it forgets things. It doesn't no conversation to conversation. And one very easy way to save your state is just write it to a file. And so you give it right access and it can create files.
And now all of a sudden you've overcome this, like probably the single biggest challenge that exists inside these large language models, which is that they're fundamentally stateless.
So Claude writes itself little like memory notes right to remember the entire ca context of the conversation, and that's how it solved that problem.
No, so there's sort of two things going on in claude code.
Beneath the hood.
There's one thing that works exactly like chetchipt or any of these other ones, which is it's maintaining a conversation history. So every message you send it and every action it takes, it's recording to a log, which is just one big file. That's really no different than what chet ChiPT can do.
Where it gets really interesting though, is it can also write files that it can then read, so whereas that conversation history is all saved off, and eventually that conversation gets too long and it needs to do a thing called compaction. And when it compacts it, it tries to sort of just remember the bits because there the total.
Window is.
Large. But I mean it's like one hundred thousand tokens I mean.
By memory notes, right. It compacts the information into the important stuff, then retreat.
It does that.
It only does that at the end, like once it runs out of space, once it runs out of context window. So it has two hundred thousand tokens I think, and two hundred thousand tokens in rough terms is probably one hundred and fifty thousand words. It says, okay, it's time for me to compact all of this stuff, and so it still saves your whole history.
On your computer.
You still have the entire message, but for that session, it just compacts it down to this, you know, maybe twenty five thousand token memory of what it was.
And is this like something that was not obvious before as a solution like this compaction? How important is it for this being? Like okay? As a human I can work on this on a project for a long time, like how much of an an unlock was there?
I'm not sure compaction was the okay walk. I think the compaction functionality is helpful. The way chat GPT does it, for what it's worth is they don't do compaction. They just forget your messages eventually. So if you're in one chat, eventually your oldest message is going to fall off the back. For coding, that's probably less helpful, but there are trade offs. I both techniques work. I think fundamentally, the thing that
is special about cloud code is not the compaction. It's the ability to write and read files on your computer, which means you can always write off memories.
And then what does that mean? Write off memory?
So you could say, hey, it's really important that I remember this thing for future sessions. I want to always work this way. So in a code base of mine, I have a set of documentation that explains how I like to do things. And cloud code makes a mistake, and so the next time I can write a memory, essentially it's written as a thing they call a skill, and you can write it off and you say, hey, whenever you run into this, I want you to operate
in this kind of way. And that existing across every session is really a thing you can only do when you can store it as a file. It's a thing you can't do in quite the same way when you're operating in this environment or it's just going back and forth to the API. So this access to the file system is one really big piece. And then the second
is is just the Unix commands, I mean computers. Every computer program lives on top of these sort of baseline functions and the way that the designers of Unix built
them is really elegant, and they're very small. They all do one thing, and they're all composable, and in coding terms, composable means they can be chained together, right, and so you can say, hey, look for files that mentioned this word, and then from those files, I want you to take this second action, and then from the output of that action, I want you to take a third action.
And that's just built into Unix.
You literally just put a little pipe in between and you just pipe them from one to another and that's it. And so you give it access to write these commands, and all of a sudden, it gets these sort of second and third order effects that are just incredibly powerful and built over a really long time.
So how much of claud code, the way it's different to other models, how much of that was overcoming technological challenges versus just having a good idea, because hearing you describe it, I mean, giving access to a computer seems like kind of obvious, like let's just do that.
I don't have a good answer to that. I think that it was kind of just a good idea. Yeah, I think they did some patterns really well. They're clearly incredibly talented, not just engineers, but kind of thinkers about how to structure it. Like the primitives inside cloud code
are just smart. And then the thing that they've done and Boris Cherney, who's the lead developer on cloud coded anthropic, he talks about light and demand a lot right, and Layton Demand is basically just, hey, look at the ways people are using these systems and then figure out ways to make that a part of the product itself. I think what they've done brilliantly, and this is kind of easy when you have a community of developers who are nerds who want to go talk about all the ways
that they're using these things. Is they have I am amazed at the speed in which you know, I have a small community of fifteen CTOs who all use this stuff religiously. And you know, when we first started that community, it took them a month to I would see it in the chat, and then a month later it would get built into cloud code, and then increasingly.
It's like a day later. It feels like they're just they're just listening to it.
But I think they're just not only tapped in, but they're really fundamentally, you know, they're dog fooding it. They use their own products when you you know, they talk about the productivity engineering productivity at Anthropic. You know, despite growing at a crazy clip, it continues to go up. And you know, anybody who's built had to manage large scale pieces of software, large scale code basis knows that's not the norm.
So VS code and cursor, these are IDEs. Cloud code is not an ID. What it's called a CLI is.
A CLI a command line interface.
Got it and the other labs now they also have CLIs. So why are we all talking about claud code and I chattypts is called codex. I don't know what Geminis is called.
I think it's just called the gem.
Why are we all talking about claud code rather than the other cla is that kind of have the same thing? Like, what is the difference?
I think, first and foremost they were first, okay, so and I think they've they've had a lot more and you know, from my very personal opinion, I think they've done something smarter and better as far as the permissioning model. So you know, one of the really dangerous things is you've got the same running on your computer. You don't want it to go and delete everything. And they have a very fine grain permissioning model where you can say, hey,
I want to allow this just this one time. I want to always allow it.
You know, I always click always allow. I'm living on the edge.
You can you can next time you run it. You can just do a flag that says dangerously skip permissions and and it'll just They call it yolo mode. I think I think more fundamentally though, if I look at at Codex versus Code, I think it's a difference in philosophy around what you want AI to do to me. Codex, which is excellent, is very focused on building an agent that you could just give something to and it'll just go do it. So I want to give it that task.
I don't want to intercede. I don't want to give it any more feedback. And claud code is much more designed to be kind of a pair programmer. And so you know, in engineering, pair programming has existed for a while. It's a really weird sort of productivity thing where you put two engineers on the same problem and it turns out that you can get better code and multiplier Yeah, and it sort of makes up for the fact that
obviously you know you're doubling the staff on it. But because of how many fewer bugs because you've both said svies, it has seemed to work out for many folks. Most companies don't practice it, but I think cloud code fundamentally is much more designed in that way.
It's a pair programmer. It's they. You know.
Whenever I start a project, I start in plan mode. So you start in plan mode. You put together a plan. I really, I mean it has spent a lot of time in plan mode. You go through, it gives you a plan, back, it asks you how you feel you can give it a whole bunch of direction, and then it's only then that it goes off and it goes
into it. So you know, we're working together. And I actually have a whole system now that I've designed where I use a task management system called Linear, So I have claud Code write tasks off to Linear, and then I've worked with claud Code to write a document that helps sort of decide a set of heuristics to decide when you should assign it to Codex versus when you should give it to claud Code. And so if it's tightly defined enough and simple enough, I just send it
off to Codex and it does it totally independently. And then if it's complicated enough that I think it requires my time and attention, then it saves it for me us to do together, and we'll work on it together. And so if it's sort of touching a kind of important enough if it's changing some part of the data model. There's these other kind of you know, fairly basic extet of criteria that I use, But that to me is
the fundamental distinction. And you know, I find cloud code in that way to be just it sort of fits what I want to do and how I want to work much better.
Talk a little bit more about how it actually impacts the workflow of an engineer, because you know, my impression was people can code, right, Like, the coding problem is kind of solved at this point, and even if you can't code, even if you're not a professional engineer, you can hire someone from like India or Indonesia or wherever to just write you a code. Maybe it'll take them a week instead of like two days with claud code. But how much does this actually change the workflow for an engineer.
As completely as it could be changed. I mean, I would say that over the last three months, I've written, personally, I don't know a few hundred lines of code. Like I am mostly a manager of a set of agents who are writing code on my behalf. And you know,
increasingly what I think is interesting. I've been thinking about this a bunch lately, is like, in some ways, it's just bringing me back to the core challenge that has always existed in software development, which is how do you manage all large scale software.
Development project acrossation.
It has become a coordination problem, and I spend a lot of time sort of now designing my CLAUD code system to ensure that code goes through all the proper spec checks and that it has all these things. The other thing that you know makes code a particularly good place to do this is that code is verifiable in a way that you know most other work is not. So you know, with code, you can verify that the build works right. So you can say, hey, I want
to build this, I want to build this package. I want to make sure that it's actually going to build and that there's going to be no failures. That's a very easy check. It's either true or it's not true. There's also coders use linting, and so linting is a way to kind of look at it's static code analysis, so it basically tries to sort.
Of find.
Things in your code base that are not going to work ahead of time where you can predict that. Obviously, you can't predict Alan Turing prove that you can't predict with certainty whether code is going to run, but there's certain patterns and things that it can find. It's essentially does static pattern analysis, and so you know, you have
it run all these things. But the more kind of opinionated you can be about that, and the more steps you can have it go through, so I find, you know, now I'm kind of the designer, which, honestly, as an entrepreneur and as a CEO of company is like that's kind of always been my job. Like I've been not I have less and less been a person who writes code and more and more been a person who designs a system in that case of company with a bunch of people who write code.
One of the funny things it seems to me is that setting aside Claud code, Claude itself has a reputation for it's a nicer chat bot to talk to. People find it, and you know, chat gbt seems to really be psycho fantic. I still think it's I know it's improved,
but I actually don't think it's improved or not. I still people like the prose style of Claud Claude and I'm curious that in the pair trading pair trading, I'm thinking about finance, the pair engineering model, whether there is also an edge there which is like, here is a chat bot that is not annoying to talk to while you're iterating, and whether that is like a meaningful distinction between you know, coding with codex or whatever.
Yeah, I I don't know. It still can be very annoying, I can tell you. And it'll still sometimes be overly, overly effusive with me about a design choice, I mean, or sort of notice something which I could live without.
So I work on this project that's doing this linguistic things and I eventually had to say, like, give it to me straight, how bad is this? And then so I said, I said, Actually, what I said was assume for a moment that you are a quantitative linguistics for the PhD. Give me your honest assessment of where we are with this. And it said like you've developed a nice toy and there's no evidence that it actually does.
And I was like, okay, that's nice to hear. I actually like, you know, I appreciate that, and I thank you very blunt, not you know, it's still like polite, but it was like, this, doesn't you haven't really shown anything. You haven't really established at all that your saltware does what it claims to.
Yeah, I think so. I think stylistically I kind of personally agree. My theory by the way, on Claude versus Opening Eye Chat GPT models is I think Claude is actually better at sort of reflecting what you give it. And so I think part of why we think it's better is it it's better at.
Pretending it's us.
Yeah, And so we tend to like that is this is purely speculation, but that's always been my theory on so.
It flatters you in a different way.
I think it's flattering you in a much more way.
Yeah.
Interest But for a long time, just Anthropic has been producing the best coding models, you know. I mean there's there can be some debate there now, but you know, there's a great curse story from Cursor actually where Cursor basically wasn't that good, and then SunNet three point five came out and all of a sudden, Cursor was amazing, and Cursor became a tool that everybody started using. But it wasn't until this other model came out and they made that the default model, and you know, I for
what it's worth. I think the other takeaway from that, which is a kind of big theme we see in the market as a thing that the cloud Code team has talked about, is you constantly have to be building ahead with AI in a way that is very unique in the world of software, where you kind of always want to build things that are working at like seventy or eighty percent, because if you really spend the time to get it up to ninety or one hundred, you're going to lose all the games you get when the
next model comes out, and the you know, with the amount of capex being spent on these models, like there's a next model that's going to come out that's going to be awesome, and you just kind of want to be downstream from that, and you don't want to waste six months getting an extra three percent when that new model is going to give you an extra seven.
Yeah, this is the only certainty with AI is like there's always going to be a new model.
The worst model ever use is the one that we're using.
That's right, That's right.
Are we all going to become coding illiterate? Are we just going to forget how to code. If everyone's using you know, general language to do.
Forget, I've never learned. Yeah, okay, you know what I've been thinking about.
You know that.
Scott Karp, the CEO Palanteer, here's that line. He's like, when I was young, I was too poor to have a car, or so I didn't get a so I never learned to drive. And now I'm too rich, so I never learned to drive. I feel like when I was young, I was too dumb to learn to code, and now.
You leaped ahead.
Yeah, now I'm too smart to learn Python or HTML or whatever.
I have a couple of takes on this person one personally. So first one is I just think like this is the worry of all technology.
Ever.
There was a paper that came out that showed that people were, uh, you know, they were forgetting more things or something because they were using chat GBD. But you know, uh, in Phadrious Plato was worried that people were gonna forget things because they started writing things down. And you know, I think the trade off there was pretty good. We got the scientific revolution a couple of other things, So, uh, you know, I think that's the sort of natural knee jerk.
With that said, it is. I it's very strange when you have people, you know, the Cloud Code team is talking about how little code they write. Now, I draw a distinction between the sort of Vibe coding and the kind of amateur people who have never written code. And I think that is amazing, by the way, And I think there's a lot of software developers who are really mad about that because there they claim it's for safety reasons or whatever, but I think fundamentally it's just they've
got people on their turf. But I think that's incredible. I mean, my hat, my my nine year old Vibe coded a website wo and for Secret Santa. She's now ten. She would get bad at me if I called her nine, but I think she vibe good at when she was nine. But that that's awesome, right, I don't know, that's amazing. That's a way for people to express themselves in a way that they couldn't before you did your your linguistics process.
That's that's fun and interesting. But yeah, I also think the other the the thing that's happening with professional software developers when you hear from anthropic or you know, when I'm talking about it's you know, the code going through this process, and you know, all the code still gets
reviewed by people. We're not letting it get out the door if it's not at the same level as human and it's just But what's amazing is I'm I'm running five of these sessions at a time, right, and so I've got like software being developed in parallel in a way that is unimaginable. And you know the other thing is just now, the best software engineers wrote the least
code anyway. You know the sort of classic story of like the difference between a junior developer or senior developers that a junior developer gets a problem and they sit down and they put their fingers on the keyboard and they start writing code. And a senior developer gets a problem and sits there for three hours and tries to figure out what the best way to solve it is and then spends five minutes writing code to get it done.
True elegance is restraint. That's what I say.
What are you.
Seeing in the companies you're working for? Like, I find it hard to believe and I was maybe skeptical of this, but it feels like right now we're here with technology, where like if viral like company is like, like I said, you can build charts of data in a way that used to be like someone would have had to get
their hands dirty, ear, et cetera. And the companies that you talk to is right now that having an effect on how they think about what positions they're hiring for and the skills they're looking for and so forth.
I think that it's hard to answer, right, really, I think that certainly. I do think I personally think if I look at the sort of layoffs in the technology industry of the last couple of years, I think some part of that is just looking at the output of these models and saying, hey, these models are able to produce it, you know, the median, and I have a whole bunch of sort of middle managers who are producing
it the sixty fifth percentile. And it's like I can produce median for a dollar fifty per million tokens, or I can produce sixty fifth percentile for hover many hundreds of thousands dollars a year. It's it's a sort of fairly simple trade off, I think. So I do think there's a lot of downstream effects. I think the other thing that's happening, is is kind of like middle management is under threat because it's the realization that hey, like, part of what these models are amazing at is is
I think of them as like a fuzzy interface. They can sort of turn any data into any other data, right. You can sort of transform data from one format to another. You can take a PDF and you can turn it into charts, right, And there's whole people who exist or you know, if you think about what product managers do a lot of what product managers do is they take how people are using a product and they try to transform it into a format that engineers can then use
to figure out what to do. And I think a lot of those kind of a lot of those pieces that used to just be kind of transferring knowledge.
I've always said, Tracy, I think one of the most important roles in any organization is essentially translation work. And you see it in the newsroom where it's like here's a team specialized in emerging market currencies, and then they have to like they have to then tell the senior
editors what they're working on. But the senior editors, who are maybe more generalists, don't really know like why like some sort of like you know, one yen carry is important and that a really important role within any organization is essentially the team that can translate between the generalist team and the specialist team. Absolutely, and so that's an interesting observation and the sort of engineering world of like, okay, these are tools that are in sometimes translation tools.
So we talked I agree completely, by the way, but we talked about vibe coding and Joe has this application that I don't think you're looking to monetize.
No, it's I'm just trying to make it for the good of the world, right, Okay, so when did that become a crown? I'm not a lotizing it, But.
Like this opens up massive questions for software as a service, right versas because if everyone can write their own software, you can replicate anything that's out there that is currently charging money. What's going to happen to software?
I think software is pretty screwed. A lot of it, at least not all of it. You know, you still it depends on whether you call that cloud provider software or not. You know, you still need to run this stuff somewhere. And I think there's there's certain kind of software that you know, you just don't really want to be in the business of writing, you know, I as someone who's tried to build a project management system, I'd really rather I don't think anybody should be in that business.
But I do think fundamentally, I mean, we see this every day inside enterprises. The sort of build versus buy pendulum has just swung. And you know, I mean I used to run a SaaS company and we sold to enterprises, and you know, for a long time that I think that made a lot of sense, right because like, hey, it just didn't make sense to try to build this thing on your own. And so but the price of that was, you know, won the price, right, like, and
it got to be more and more expensive. The other price was that you were paying for a lot of stuff you didn't need, right, because the whole job of building SaaS is you need to generalize problems, and so you build things that are going to work for everybody.
And that means either you.
Have to sort of adapt or you have to build this sort of very configurable software. And I think, and what I see you just you know firsthand, is that inside these organizations you can now solve very specific problems that are highly valuable and not only can you solve them better than generic software, but you can actually, in a lot of ways do it for less money because you're trying to tackle less stuff. You didn't need the sixteen other features. You bought it for the one that
you really really cared about. And so I think that part of it, you know, I don't like there's I definitely think there are pieces of the software industry that are gonna, you know, come out the other side. You're gonna nobody wants to deal with payroll, right like you know, somebody you're still gonna buy some payroll software and you're
still gonna have that. But you know, I do think there are a lot of pieces where the software existed essentially as a kind of wrapper around a database, and now you're just gonna, you know, with just the database, you can do that. And then you know, the other piece I'd say here is it's this is not this is a kind of confluence of circumstances where it's not just the coding, it's also the fact that you have ai to do a whole bunch of work. So you know, if we pick on CRM for a second, right, like
you know, Salesforce dot Com Salesforce dot com. We can you know, you look at what the interface of that is. And essentially it has existed to get salespeople to take unstructured data, which is sales meetings, and turn it into structured data that so can be stored into database. And now you have AI, and AI is very capable of taking unstructured data directly from the source, so you have people recording meetings, and then it can structure it into
any data that you want. This is one of the very first sort of mind blowing moments I had was that I could give it a Jason interface. I could describe exactly what I wanted the data structure to be, and it would give me back that information in that data structure. And we've just basically been having a bunch of humans do that work for a very long time, whether it's in CRM or project management or any of
these other places. And the ability to just kind of get rid of that whole thing, I think it really does bring into question the value of a lot of these software companies.
Well, so we have seen like a lot of software sucks. They look like melting ice cubes right now. Maybe they So what is it I want to talk. I mean, this is like you know, our listeners who are investors, there's a pretty high stakes question of like what residual value there is. But talk a little bit more about Salesforce. Maybe there would be a time to learn what sale what it actually does as it's massively being disrupted. Now
we get around to learning what Salesforce is. But I know it's like many things there are apps that people built onto Salesforce. But this sounds like we're hitting on when I think probably one of the crucial questions for
like the future of the software industry. So talk a little bit more about like the current approach and what people are buying when they buy a package or subscribe to a service from Salesforce, and then what the unlocked opportunity is from having AI like live in the same world as all your files.
Yeah, so I think if we take CRM as the general category, know the biggest players there are.
That's customer relationship.
Customer relationship management, that's like what you know, Salesforce does it, SVP does it, HubSpot does it for the mid market. You know, when I think about that product and I
think about the way we've used it inside enterprise sales organizations. Essentially, you know, it's a database of companies, it's a database of contacts, it's a database of deals you have in the pipeline, and it's a way to track all those deals you guys hit on something before that I think is really it, which is like inside companies, there is a huge group of people and who exist to answer the question from management of what is the status of
something right? And you know that can be sales management, it can be product management, it doesn't matter, right, it could be within a newsroom. Somebody wants to know what the status is and somebody else exists to go figure out what the answer to that question is. And so fundamentally, I think those CRM tools are bought first and foremost to answer what is the status right? What's my pipeline
look like? And to answer what your pipeline looks like, you need a bunch of salespeople putting deals in and those deals are associated with contacts and companies and they say when is that deal going to close? And essentially you were asking the sales people to make the updates in the system to do that and just very tactically, I mean, you know, I run a company. Now we talk to a lot of we have a lot of
sales calls. We record those calls and they get transcribed, and the AI then looks through them and makes decisions about where this deal should be in the process. And it's much better than having somebody try to go update it, because those people never updated anywhere. The secret of all of this enterprise software is that nobody was using it the way that anybody wanted to anyway. And so you know, I think that that is sort of you know, a
lot of what's happening there. Again, it's sort of some of it's the coding, some of it's just the core capabilities. And then you know, you still need databases, right, so it's like, you know, you look at what data bricks and stuff like. You know, I think those folks are still sort of genuinely sitting in a pretty good place where you know, all software has to sit on side on top of some database that you can sort of
read and write to. But you know, I think some of those categories that were specifically focused on kind of like human input. Now, of course, you know, Salesforce has a whole AI thing and they're saying, hey, you shouldn't have humans inputting in salesforce. You know, at sales is just one small piece. They have a whole customer support thing, which obviously also has an interesting implication where you know you're doing support with AI agents and so some of
it comes back to seats. I mean, you know, it gets to be fairly complicated, but I do think I think the fundamental underlying thing is anybody who buys software that is, you know, uh SaaS, you're always buying for a subset of the functionality that's nobody is using one hundred percent of the functionality of SaaS, and so there's always a trade off that's happening there where you know you're spending more money than you need to because you're
not using all of these pieces. And so you know, if you can more narrowly focus that, that is where you could say, hey, we could solve this kind of more narrow problem. And not only can we solve it more narrowly, we can solve it way more effectively because you know, the trick with AI is that the more specific you are with it, the better the output is. Right, So it's like, if you know, if outside of coding, if you just ask chet gpz to write you a story.
It's going to write you a very very median story, right, sort of exactly the median. But if you work with it and you you know, then you're going to get it. The more of your own expertise, you imbuing it, the further up above the media, and it's going to be and it's going to be you know. Of course that also means it's less where the line is between what's AI and what's not AI is going to continue to get lorier, Joe, how.
Much does Claude Code actually cost?
Do you know?
Well?
I paid for the two two hundred dollars a month version, but like high roller, yeah, I know, but uh, you know, I think it's you can get it with the pro version of like or whatever. The sub version of that blow twenty dollars. But I hit a limit fairly quickly, and I was like, I didn't have my website up so like, and then I bought the five. Then I paid five dollars for the extra compute, and I was like, this is dumb.
I think, yeah, okay, so we're going out to.
Two nice dinners right month, that's not you know, when I think about that way, it doesn't see that big.
Of a deal, it's worth it to you. Yeah, okay, so I think we can all agree this is like a valuable service that claud Code is providing. But we touched on this in the intro. It seems like the models just keep replicating themselves really really quickly. So anything that claud Code can do, I would expect another model will come in in like a month maybe less and
do the exact same thing. What does that mean for the actual like valuations of these companies and the models, Like, how are they going to monetize it when it seems so difficult to actually differentiate yourself, especially for like a substantial portion of time.
Yeah, well, so again here I think we have to distinguish between claud Code and the Claude model. So in claud Code's case, if you're using you know, the latest version, you're using Opus four point five, which is the model. Opus four point five has a price of something in the dollar fifty to two dollars for a million input tokens and whatever it is on the output, which is like roughly the going rate for cutting edge modles. Gemini three pro is the same price open a Chachipt five
point two is. They're all the same price. So the first thing is is you have to differentiate between those. And so I think a big part of what Anthropic is trying to do is they're trying to lock people into claud Code. In fact, there's just some controversy amongst some nerds where open Code, which is a competitor to claud Code, used to let you use your Claude Max
two hundred dollars. So the trick with the Claude Max plan is if you're just buying those that number of tokens, it would cost you significantly more than two hundred dollars. It is a super super discounted plan. So like you, you are probably you have the access, I have the access to use I would guess in the thousand or two thousand dollars of tokens for my two hundred dollars a month. So it's a very very heavily subsidized plan. And open Code, which is an open source version of
Claude code, a sort of competitor. They had found a way that they would let you use your Claude Max plan with open code, and Anthropic last week shut that down, and some open Code people got very upset because they said, like this is not what you're supposed to do or I'm not sure exactly what they said. I never felt like I got a particularly good argument out of it.
But you know, I do think part of what they're trying to get at because is that, you know, at the very top, models like these are all amazing, like the Google Opening and Anthropic, their best models are all on par with each other. I mean, I would move them around a little bit. I still think Opus four point five is the best model out there, but you know, I mean that might change tomorrow. Like and that's where something like cloud code is really interesting because it's a
product that is very it's just theirs. It's a piece of software. It's not an AI model, and so it's sort of it's less able to be disrupted. Now again, I think if if somebody else wanted to copy that exactly,
they could. Codex has one, Gemini has one. I just think they take a very different tact with it where it's much less and so, you know, I think what they're trying to do is get developers like me to feel very comfortable inside that so that when we go open I still open Codex or tried Gemini, or I was playing with open code the other day, and it just doesn't feel familiar in the same way that you know, if you're trying to move somebody from a PC to a MAC, it doesn't feel familiar. Right.
They want to own like the ecosystem.
The environment, that environment or what a world. Noah, thank you so much for coming on odd Laws. I was like dying to do an episode about this topic.
Thanks for having me.
Way, I don't have AI psychosis. I have a claud complex.
Why is everyone making that joke?
Wait, which joke?
The psychosis joke?
I think you were going to be proud of me for saying Claude complex.
Oh oh that is very good.
I do one pun finally for Tracy And why was it for making that joke?
Well, I was a joke.
I was handing your sirt. I finally make a pun, and you just jump right over.
Everyone keeps saying that claud code is AI psychosis for smart people, right, Like, how did that beck come a thing? Yeah?
All right, but there's a good pun.
Also very pro coded. I find you think so all of AI is pro coded.
Uh, this is true. We should talk more about this, you know, we should have David Shorn he's been doing a lot of polling about various demographics and how they feel about AI. We should need some interesting that. Yeah, we should do that anyway. No, thank you so much for coming on offline.
Thanks for having me.
Well that was fun, Tracy. I really like I It's obvious to anyone who's been within five minutes five feet of me for the last two weeks. I'm like totally addicted and gone down, I know, gone down the rabbit hole and stuff. But like I for the first time, unironically, I'm like, Okay, this is transformative technology beyond being very impressive technology.
Right, So I've been coming to a conclusion, which is that you know, AI can be both under hyped and overvalued simultaneously. Like and I feel like that's kind of where we are at the moment.
That were you're making your stock call.
Yeah no, but seriously, like it it's a big deal. It's going to change the way we work, But is it monetizable? Can you differentiate the actual models? The better the technology gets, like, the easier it is to just do what everyone else is doing. And also like the compute gets cheaper and cheaper. So I just don't know how you monetize this.
Well, so that's very interesting his point, which is that it's the tokens are heavily subsidized still, and so that if you're paying and actually using that two hundred dollars max program and you actually use it to the limit, Claude is going to lose money on this and then the prices keep dropping. And I know, like claud code is okay, they're attempting to create something that resembles a traditional software ecosystem that you feel as a user that
you're locked into. But so far, in my various like since November twenty twenty two when I started playing with AI, it hasn't felt like anyone has established lock in with anything. And it's very it's very movable, and I suspect even though I have this file now on my desktop that has a file called claud md that gives instructions, et cetera, I'm certain that if I open this file with Codex or Googles, I could probably just pick it up the same.
Yeah. I also think there's a fundamental issue with the lock in strategy because when you're talking about technology on the Internet, it just feels very against the grain to try to lock people into anything, and we've seen various projects over the years, and it's a lot harder than it looks.
Yeah, I mean, I guess I would say it's a lot harder than it looks. But then we also know the flip side, which is that tons of people are locked into software that they hate, right, Yeah, people are Oh I hate people. How many times have you Oh, I hate Outlook right? Or I hate Microsoft teams and I hate this and I spend money on it every month and my organization can't move off of it, or we can't migrate off of it. So I do think
that cuts both ways. I do think he offered the best explanation I've heard of why the AI coding models are a threat to a lot of pretty big software businesses, especially especially the point about how the user never uses all of the features that they actually that the software got built for, and therefore maybe the build versus by calculation really starts to shift when they can just design that one feature very quickly.
I totally agree on the software side, it seems like an existential threat, but just like the locked in ecosystem of a particular model. I know he said it's not actually a model, but that seems like a bigger issue to me. I don't know. I guess we'll see.
We're going to see and I don't know. I kind of think we're going to see quickly.
Yeah, that's again, that's the only certainty is like stuff is happening.
What is happening now?
Yeah?
Okay, shall we leave it there?
Let's leave it there.
This has been another episode of the Odd Lots podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway.
And I'm Jolly Isn't all. You can follow me at the Stalwart. Follow our guest Noah Brier, He's at Hey, It's Noah. Follow our producers Kerman, Rodriguez at Carman Arman, Dashill, ben Att at Dashbot, and Kilbrooks at Kilbrooks. And for more odd Laws content, go to Bloomberg dot com slash odd Lots with a daily newsletter and all of our episodes, and you can chat about all of these topics twenty four to seven in our discord Discord dot gg slash odd Lots.
And if you enjoy odd Lots, if you like it when we talk about advances in AI, then please leave us a positive review on your favorite podcast platform and remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free. All you need to do is find the Bloomberg channel on Apple Podcast and follow the instructions there. Thanks for listening.