Coté (00:00.637)
All right. Now last week we covered that one of the most well, let me follow up from last week. First of all, Brandon, did you order the tape measures?

Brandon (00:08.936)
I did not. I did not. I've let us I've let us down on that, but I'll I'll work. I don't know. I don't know if we have the budget for that. We you we have the sticker budget, team measure budget, whole different department.

Coté (00:18.033)
Oof. Yeah, especially when you get those you get those Fat Max ones. Those I think they charge double for that and the the yellow color. Well, we were talking about how it's important to have I forget which one of you came up with it, but you want to have a dad box where you've got all the stuff that people always want scissors, tape, tape measures as mentioned. I think it's a few other things. And one of the things we mentioned was I think Brandon, you had a a phrase for them. I would call them nail clippers. You just said clippers, which which I think is fine.

Matt Ray (00:19.982)
Go to conferences.

Brandon (00:31.388)
Mm-hmm.

Coté (00:49.095)
Of course. and I was just today I had I had, you know, what you might call or what you would call a hangnail. Just all of a sudden it's like, why does my nail keep catching on my shirt? or whatever. And, you know, then then you do the I think they I think they call it worrying. You mess around with it for a little bit. And and then eventually I was like, I gotta cut this hangnail off. This is this is no fun. So so what I wanna ask you to, and then I'll reveal what I do is when you have I assume you you know, you're living

Maybe not as soft a life as I am, you know, so maybe it's a little rough, at least not softer. I assume you get hangnails. Now, do you when you get one, one, do you cut it off or you just let it ride? Or two, do you only cut off the one nail or are you like, now we do all? Like are you buying do you buy the new tire or do you balance out and you buy all the new all new tires at once?

Brandon (01:21.341)
Mm.

Brandon (01:43.689)
this is man, this is a deep question here. I think I think I just this is why people come to the podcast. I I guess I kind of like ride it out till it's very uncomfortable. Then maybe I take some action with some type of I guess clipper of some kind. And then no, I don't I don't I don't even know if I understand the last part of the question. Do I then do the full do I do like all my nails or something?

Matt Ray (01:48.63)
This is what people tune in for.

Coté (02:00.988)
Mm-hmm.

Matt Ray (02:07.113)
Is is

Coté (02:07.298)
Do do all your fingernails. Right, right.

Brandon (02:09.392)
Yeah, I guess once I out yeah, I guess I guess I clip them. But yeah, I guess it's like if I have them and I I do yeah, I do a quick pass, you know. I don't know. I think are the are those the right answers? I don't Matt, what are your answers?

Coté (02:13.315)
All of them.

Matt Ray (02:22.126)
Well, if you're not doing it my way, it's not the right answer. I'm I'm a I'm a you know, clip the offender. yeah s sometimes it might get pulled off, which is always stupid, but I do it. yeah, I know, I know. I d I but but no, I I I'm not a you know everything everything has to be retrimmed, because it it's usually over on the side and and my

Coté (02:36.582)
Do

Coté (02:49.126)
So you only you only cut the one the one nail.

Matt Ray (02:52.29)
I only cut the one. I I my my nail regime is every weekend I just cut my nails so like I don't have to think like did I do it or not or when was the last time I did it? It's just like it's part it's part of the the mental like to do list or yeah.

Coté (03:07.258)
Hmm, interesting. Interesting. Well, you know, I I'm I'm like you, Matt. I just cut the one nail. And then I have to say, that becomes as almost as irritating as when when when I had the hang nail. It's it's annoying. And then even worse, I don't know if you have this experience. I went to open something. Maybe this is it. I went to open something and I forgot that my nail was short, and so I ended up hurting myself trying to pry open this olive oil bottle. yeah. Yeah. Well

Matt Ray (03:30.978)
How short are you cutting it? That maybe that's the problem. See, that's that's what people tune in for, Kote, hygiene lessons.

Brandon (03:38.856)
Well Kote, have you decided maybe like also all the things we're talking about, have you thought about outsourcing? Go the full manicure.

Coté (03:38.864)
I guess

Coté (03:47.931)
You know, I feel like they probably they they do that here in Amsterdam, but it's not like back in the US where like every street corner has a Verizon shop and a and a nail shop. Like you gotta go you gotta range a little bit further to to find these things.

Matt Ray (04:02.721)
They I thought they were first world. Come on.

Coté (04:06.742)
Maybe there's an ordinance against it.

Brandon (04:10.61)
This episode is brought sorry. This episode is sponsored by Cygnadot. That's S-I-G-N-A-D-O-T. Coding agents are shipping 10 times more code, but verifying it is the new bottleneck. Every change ends up waiting in line for one shared staging environment. And duplicating the whole stack for every change doesn't scale either. It's never just the services, it's the databases, the cues, every dependency. Cygnadot takes a different approach. Lightweight, ephemeral environments that only spin up the services you're testing and share the rest.

Developers and their agents get a high fidelity environment in seconds and run real end-to-end tests on every change before it merges. Try it for free at signadot.com. That's S-I-G-N-A-D O-T dot com. We talked with SignaDot.co back on episode 581. Worth a listen if you missed it. And of course, we thank SignaDot for sponsoring our show.

Coté (05:02.736)
Well, there was now it's on the topic of AI, which thankfully exists. Otherwise I wouldn't have quote unquote read this hundred page PDF about AI usage. Or I don't know, maybe I would have. The back in the old days, who knows how long I spent with a PDF scrolling through it. I bet it is it is an a a study. So there's probably like forty pages of appendices and citations in there that you could easily skip over. I wonder if there's like academics who are just like,

I'm reading every character of every citation. And they just they they just they j they just get their grad students to come in and like, you know, fan them while they read it and like, nice use of a square bracket there. Like how they did that. Odd choice for the the three letters from their names. Hmm. Don't know about that. but anyhow, it was a as I understand it from from Claude understanding it, the the the people over there at Google.

Matt Ray (05:36.619)
Yes. Yes there are.

Brandon (05:37.54)
Coté (06:03.362)
they looked at, and this part I think is I spent a lot of time this morning looking over this. I think it's important, but they looked at the conversations that are basically in the consumer tier of Gemini. So not if you used API stuff or not if you use Google Workspace. And also, and and spend some time on this too. Also, if you you know how it's the AI overview it shows you in a search, it doesn't count that conversation. It counts the conversation when you click on it to go deeper into it.

And then also some paid Gemini use. But anyways, the whole point of that is unlike an anthropic study that was out a couple of weeks ago, I I don't know when it was, but it doesn't really, I don't think, represent like broad enterprise use because it's it's a lot of not paid for use. Now that said, it is interesting to see what what people are using it for. And it looks like you know, top of the list is still kind of like entertainment, leisure.

hanging out, doing things, asking questions about stuff, better search, so so to speak, which I don't know. Yes, that that seems accurate. Like well in my recommendation, if I remember, I'll go over something that I think was much better than searching that I used not Gemini, but Claude for. And then you get into you know, education stuff, having people, I guess, study. People are only studying with it, right? They're not, they're not actually cheating or do anything.

Matt Ray (07:05.345)
Yeah.

Coté (07:29.998)
And then the the the some a little under fourteen percent was like for work related stuff that people were doing. Again, this is like at the freeish consumer tier, so maybe they don't want to use it for work. But you know, what I think I think one interesting thing that they noted that will probably reflect what at many of us do is they're actually were saying that there's a lot of broadly speaking mechanics, whether they're car mechanics, house mechanics, I don't know about biomechanics.

Who take pictures of something and are like, how do I fix that? And you know, that that that seems pretty perfect. It's it it essentially it's like kind of searching through stuff. And then there was also a little bit of like, help me fill out some paperwork. So I feel like entertainment, start looking things up and just like handling all the bullshit forms and paperwork you have to do, which sounds like perfect. That might advance GDP by a few percentage points in in the long run.

Matt Ray (08:03.281)
Brandon (08:30.312)
So I don't know, but I guess I kind of come back to a little bit of I guess it's I don't know. Well, let me ask you this way, it's presented, or maybe the way you presented it, was there sort of like, well, like that's all it does? Cause I think one thing that was in here, it basically it it sort of does the glue work, right? You know, the drafting, the looking up stuff, the connective tasks and stuff. and I don't I guess it just depends on like what your incoming perspective was. I guess if your incoming perspective is

We're all going to be replaced by agents and the robots do everything. Then this is sort of not meeting that that that goal. But I think if I flip it around, if you're just like, yeah, this actually is you know, kind of we what we talked about all the for years, developer toil, you know, it's just sort of like what's to call it, like white collar work or actually just toil of anyone, just everybody's personal toil. And it's like the fact that it can take so much of this off your plate, like I look at that as like, yeah, that's great.

I mean that seems like a a a huge fantastic outcome, but is

Coté (09:30.589)
I I I think I I think you're you're there's there's a there's a intro fun grouping there you're you're kinda pointing at one, there's like let's let's let's broadly say automation. We'll kind of throw out the searching and research because I think that one's pretty straightforward or not. But y you know, you're just like automating stuff in your personal life or automating stuff in your work life. And I think probably if you were to bring in like obviously if you were to bring in anthropic, you'd get all the clawed code stuff.

Matt Ray (09:31.553)
I

Coté (09:58.951)
But then also the paid tier of stuff, you would probably get more automation of work stuff. Like one thing notably, I I mean, I assume that like Palantir and all those people aren't using like these tiers of Gemini to like shoot missiles and stuff and and help ice out. And and so so so for example, you have like from from the you know helping people identify

Matt Ray (10:13.066)
No. No, it's still this this is the steel your alpha tier.

Brandon (10:14.194)
think we can I think we'd be confident they're not using the consumer chair.

Coté (10:26.086)
Targets and people and stuff, that use of AI, that's not in here. and then like, you know, I don't know, programming, all this other like like any sort of regulated enterprise. Like I'm pretty sure if you have like HIPAA or GDPR requirements, you're probably not using these types of Gemini thing. But the whole point being that we would intuitively think if you looked in those areas, there's also just a lot of automation and things going on.

Matt Ray (10:43.532)
I hope not.

Brandon (10:50.738)
Right. But I you know, in the the work slice of it. So I guess if you break down the let's call it the twenty percent. No, I guess it's of the thirteen point five percent in conversations that were work related. So we break that into it. It looks like it breaks into something like forty percent is sort of like kind of drafts something up, right? But you but with the the the belief that like someone's still gonna edit it. I think that's fantastic. I I mean, I do that all the time. Thirty percent is sort of like, you know.

Coté (11:07.281)
Yeah.

Brandon (11:16.154)
Explain this to me, educate me, teach me about this, which was probably replacing like a Google search, right? It's like you would do a Google search and maybe you would hope to find an authoritative site, but now AI just sort of just tells, you know, gives you the summary, which again, I think that's a lot better. And the fact that you can ask follow-up questions, always like that. And then only ten percent of that thirteen point five percent, so we're getting real small there, is something like do everything, right? And and maybe that's a little bit around people don't like

Coté (11:39.803)
Yeah,

Brandon (11:45.223)
I don't know. I mean, especially if the consumer tier probably can't do everything because you haven't there isn't like the automation or the tools hooked up to actually do stuff. So it's probably unrealistic that the consumer side would do that. But I kind of look at that and say to myself, like, the I don't know. Again, like I think these are like again, it just kind of comes back to perspective. I think those are huge wins. I'm like, yeah, I love the fact that they can like draft things for me and like help me, you know, sound

Coté (11:56.13)
Mm, that's a good point.

Brandon (12:08.1)
Smarter or like clean up my English. Like, I mean, I just I do that all the time. I don't feel bad about it at all. And I think, I think, you know, Kote, I think you you did this. I saw some other people do it. Like, I'm gonna put you in kind of the the writer category, Kote. You know, it's like I saw another person do something else that was like a that like takes pride in like really thinking about writing, but they they spent a lot of time like helping, or if you will, teaching Claude to like write like they write, right? And it's like

Matt Ray (12:10.38)
one of the one

Coté (12:32.272)
Yeah.

Brandon (12:33.584)
Yeah, that's good. Like I I think that's also really good. I don't think that's cheating at all. So so all these things I look at it like, yeah, I mean, they're not replacing like all these jobs or all these people, but like they're making us all a lot more productive. So I like it. Matt, you must have some thoughts.

Matt Ray (12:51.664)
well what's interesting is you know, Corey Doctor's la latest, you know, most recent-ish newsletter was kind of a defense of his use of AI. And basically he used the use case that you were just talking about as a writer

He uses it to kind of review stuff. And yeah, of course he went out of his way to tell you that he uses, you know, only a hundred percent free range organic AI. so you know, it's it's it's local LLM on his laptop, you know, no no network access, but whatever. But he he pointed out like the key though to using it is he knows what he's doing as a writer and

Brandon (13:23.655)
Right.

Matt Ray (13:33.964)
he has judgment a and and can can can deal with like the feedback it gives. He can tell like, you know, that no, that's just wrong. You know, that's not useful. You know, quit telling me this. You know, you're not helpful. So he's he's like, it's it's better autocomplete. It's better grammar checker. It's better spell check. but

What w what when when when I was looking at the notes for this, you know, the the second bullet point, you know, we we said, you know, the first use case is you know, consumer leisure socializing is first. And then the second one it said beating education. And I was like, Yeah, that's right, people use it to cheat on on schoolwork. And and to me that is like the most worrisome.

Part of all of this is is that offloading of of mental fitness. because you we we shared a link in and Slack about

a professor who had you know some hidden text and white w you know white text on white background and a take home exam and you know it said like bring up Madagascar and something like eighty percent, ninety percent of his students submitted responses that mentioned Madagascar, which had no place in the essay. Right? So not only did they like you know offload it to an AI, questionable, they didn't review the content. Or they didn't know what they were talking about anyway. So they just looked like they

prove that they didn't know what they were doing. And and to me that's the worst part.

Brandon (15:04.104)
All right. So I love this. Hold on. I want I wanna pause you there. It's like, okay, I've seen like this story, not this specific story. The quote unquote kids cheating with AI story has been now been written five million times. Okay. So my question is like, I guess I'm I'm gonna zag here on this one. At what point do we turn this back to the teachers and the professors and say, Listen, fundamentally you have to you're going to have to give exams and teach in a a slightly different way. So the fact that this professor

Matt Ray (15:16.608)
I know, I know.

Brandon (15:33.821)
Went out of his way to like in white text hide the prompt that said, you know, mentioned Madden's gas car and then failed 30 of the 32 students. Like my thing would my my question to him would be like, you probably knew this was going on, right? And so it's and again, like, and let's also like not let's not like let's I don't want to fall into the trap of like kids today. It's like, listen, if AI was probably available 20 years ago, all the Gen X kids would have done it, like.

Matt Ray (15:49.364)
Yeah. Well what what can they do? They they

Brandon (16:02.78)
Before Cliff Notes, right? People would have read Cliff Notes.

Matt Ray (16:04.748)
But we we had that twenty years ago. We we we copied from the encyclopedia and rewrote it.

Brandon (16:07.952)
Right. There was plagiarism 20 years ago. Before that, there was something else. Before so it's I I just kind of look at it and also too, like knowing the the class you're at. Like, listen, I wish every student, you know, as a as a parent of a of a kid in high school and both of you as well, like I wish nothing more that all the kids would go to school in search of just the the joy of knowledge and they would read and just wanna take every opportunity to learn for the sake of learning. So far, not seen that in the in the current high schools. I don't think that was in the high schools, but I was so it's like at some point, like the

The teacher needs to adjust and say, like, well, listen, you we're gonna, I'm actually giving the old school blue book exam, right? Everyone will sit down and like, and like you'll just have to adjust the way you're teaching. Or, like, I even think for like a high school English class that I'm watching is like, why not like literally in class be like, okay, first everybody, every and this is back to like style, right? First, everybody open up your Chromebook and just write in raw text the essay. Everyone do it together, right? And then, hey, why don't you run this prompt on it?

Coté (16:42.756)
Mm.

Brandon (17:05.364)
What would this prompt do? It would give you some guidance, be like, here are some mistakes, right? And let every kid go through that cycle, literally in the classroom, and like learn how to do it. And then when it comes time for the exam, be like, I'm giving everyone a blue book, right? Everyone's gonna sit here and write it together. And so, like, let's take responsibil. I don't know. I just look at it like at some point, the education system has to like take responsibility for like, yeah, like if it's this easy, you need to kind of rethink how you're testing and grading people versus being like

Ooh, the boogeyman. It's just like the calculator thing. It's like, my God, the kids can't add anymore. Guess what they did? They just give you a test where you don't have a calculator. I mean, that's like pr it's pretty basic to me. So I don't know. So that's just my my zagging on this. Go ahead, tell I'm wrong, Matt.

Matt Ray (17:49.117)
You're you're you're not wrong, but it also just shows like people don't even bother to cover their tracks. They're not checking the answers. They they it's not like, I was using this to supplement my education. I'm using this to replace my education. I didn't read the output.

Brandon (18:02.472)
But like this is just snor, like kids, like there will always be a group of kids that want to cheat. Like you have to, but I'm just saying this is like if it was this easy, this is the whole thing. It's like it just, it's too tempting, right? I'm saying, like, it's just like saying, like, it's just when people give a c let's say a class, like a, you know, just your typical high school class of of things, and you say, okay, everyone put the calculator in front of you. Okay. And tell everyone, no one can use the calculator. The teacher leaves the room.

Matt Ray (18:07.935)
thirty of thirty two.

Matt Ray (18:15.487)
Mm-hmm.

Brandon (18:32.412)
There's nothing but addition on the thing. And it's just like, I don't want you to use that calculator. Don't do it. And it's like you leave it with all these kids. And it's I mean, some of these kids are just gonna do it. So it's like, that's so what do you do? You take away the temptation. Everyone put all your books away. Everyone put your stuff away. This is a proctor exam. It's like, so I don't know why suddenly like if you could cheat on the SAT, if it was really simple to cheat on SAT, people would cheat on SAT. It's hard. So people don't. So I don't so it's

Matt Ray (18:46.944)
yeah. Yeah. No

Matt Ray (18:56.309)
They they yes, yes. I'm not I'm not disagreeing, but it's also like how how little effort was put into like the covering your tracks. So they didn't bother to read the output. That's

Brandon (19:08.52)
But I think it's like by definition, it's like if you've already decided you're gonna cheat and just get the good degree, you're not gonna do it. Like you I'm just saying, like it's like, of course they're not going to. It's just like if someone gave you the answers to the test beforehand, you'd be like, the answer is A, B, C, D, and you just wrote it in on the the scan tron, right? And it's like, okay, like I got the right answers. You know, and so again, like and then how would people detect this before? They'd reverse the questions or something and they'd be like, Everybody put in the same thing. And I and this is how they do it when they're like looking at a standardized test. So it's like

Coté (19:26.02)
Well I think we've we've

Brandon (19:36.873)
I don't know. So it just seems like lazy to me. It's like, yeah, there's a tool out there. Kids are going to cheat. You got to adjust how you teach, right? Like, and again, like I would go back to this professor and be like, you're so smart. You put a prompt in there and you caught all your your kids cheating, versus like preemptively tell them, like, either don't do that, right? If I'm giving you, I'm I'm trusting you all to write this essay and I will look at it and I will fail you if you don't, or just make the whole thing moot and be like, guess what? Here's the blue books, everyone spend an hour.

Write your essay, I'll grade them and and see what it is. So it's I I don't know, it's just it's just lazy to me. It's just like lazy from the like the adults are in charge. Like we're the fucking in charge. Like, you know, don't like just let the students do whatever they want. Of course they're gonna like do the easy thing.

Coté (20:20.06)
I I think I think you found the one place where return to office applies and and that is that is in the schoolroom. We've gotta have the people return to office and and and just and blue books. That's what we're gonna have a blue book education. The

Brandon (20:34.194)
But they do to be fair, man. They do have like online proctor. I've seen some of this now, like the online proctor exam, where you have like take the camera, show your room. And like so, so again, like again, if that's if that's what you want to do. But yeah, I do think yeah, probably for you know, teaching a bunch of high school students or I shouldn't high school. I don't want to pick on teaching a bunch of kids that maybe aren't there because they love this, maybe it's a required class, maybe don't love this class. Like you probably do have to put in some safeguards to be like, okay. Now maybe at a graduate level.

Matt Ray (20:42.028)
yeah. Yeah, yeah, yeah.

Matt Ray (20:54.443)
I think they were college. Yeah.

But they're paying for it. Yeah.

Brandon (21:03.44)
Right, where everyone's there getting their doctorate or something. Like you do trust them. Be like, I don't know why you're here. Like, why are you spending your time at this low-paying, incredibly difficult task and you're cheating on your PhD? Like, you should just go leave and get a job. So so there's like levels to this. I mean. Well that's what I say, but there's levels, right? Let's say, you know.

Matt Ray (21:15.861)
They they they barely take classes in the PhD. Yeah. So so don't worry about that part.

Coté (21:19.59)
I I I I I think I think you're you're right. There there's the ultimate is like if as you get higher and higher degree lever levels in the humanities where there is absolutely no economic payoff to doing it, just like I don't really care. You could have AI write it. I don't care. There does doesn't even matter. I don't know why you're like you're saying, I don't know why you're here.

Brandon (21:34.704)
Yeah. Like what do you wa I mean, why are you wasting your time? Like who like great. You got you you you did a PhD of all fake research. Fantastic. Like what did you I mean, that's really gonna help you. Good luck.

Coté (21:46.589)
You know, I I I I do as I I like your idea though of like you each get your own little agent to help educate you. And at some point you've got to stand on your own and prove it for yourself. Like I think that that would be like you know, adjacent to what you're talking about. It would be fantastic to the the same the same like little help we get doing stuff that we just went over in the the survey there. It'd be cool if if it was sanctioned and

If the same way you would educate a kid how to type, you could educate them how to like use AI stuff to help themselves.

Brandon (22:19.644)
Well, on the writing front, I mean, this is where I'd love. I hopefully some English teachers will like chime in and I'd love to hear them either email me and tell me what what I'm wrong about. But I just think on the writing front, it's like, you know, the idea that like because the number one thing I hear from people that do like teach like writing in English, just like the hardest thing for is like they don't have time to grade all the papers. If everyone writes a five page paper and you got a hundred papers, that's a lot to grade, right? And so they don't have time. So it's like, like to me, it's like, yeah, well, this is like actually I think AI is pretty good.

At like at least like looking through a thing, a paper and being like, here's the rubric and scoring it. And it's like, well, and because then I think the problem is like you know, teachers don't assign long writing because they can't grade it all. And it's like, well, why don't you just turn that around? Assign long write writing, right? Have AI help you grade it. You can sit there and you know, grade it along with it, make sure it's it's in line with what you want. And then people would write more. And if people write more, they usually get better, right? I mean, that's not like it's you know, it's not that complicated, you know.

But instead, this professor's like, I put a hidden prompt in there and I found everybody that was cheating. It's like, okay, great, great use of time. Fantastic. Fail them all. You know what? Fail them all, kick them out of school, expel them, and let them never come back. It's like, great job that you've really done a good job educating. Mission accomplished, sir.

Coté (23:31.836)
It was like like A AI destroyed jobs, just not the way you think.

Brandon (23:36.38)
That's right.

Matt Ray (23:37.066)
Well, I think it's destroyed education. That's that's the problem. Is is people don't want to take responsibility for for learning the content, you know, or or you know, processing it. And yeah, I mean the professor is is catching them in the act, but it's like it's just watering down the entire field. And yeah, I I I still maintain like, you know

Use the AI, don't get caught. And if you're not getting caught, you're reviewing the content, which means you probably learned how to edit your material, which means you read through it and you paid attention. These people did none of the above. Copy, paste, go.

Coté (24:18.832)
Yeah. You should at least have some review in there. Well, meanwhile, it seems like the companies who are helping fund the destruction of education are are doing quite well. The and and you know, I I think I I had I had to read through the the Google earn I didn't read through the Google earnings, the coverage of the Google earnings a few times. Brandon, you always pay attention to this stuff. It sounds like Google as a whole or alphabet or whatever, they made a shit ton of money, like eighty percent more profit than.

Brandon (24:19.036)
Like it.

Matt Ray (24:29.918)
Ha ha.

Brandon (24:46.215)
Mm-hmm.

Coté (24:48.59)
last year or something, which you look at that and you're like, I don't I don't know how that's possible. I certainly haven't been Googling more stuff, but it sounds like that's because they have a bunch of equity for anthropic and SpaceX. And so it's just just their investments went up quite a bit.

Matt Ray (25:01.096)
I I I think that was the I that wasn't alphabet, that was Google Cloud. Yeah.

Coté (25:08.378)
okay. Was that what I could that's what I thought I read three times and got confused about. But whi which wh

Brandon (25:13.98)
Well, I think where it actually falls on the balance sheet is probably less important than it's really two things. It's the investment in Anthropic is not surprisingly, has gone up significantly. So they're making tons of, you know, money just as an investor. And then, you know, on the Google Cloud side, it's like not surprisingly, they have incredible utilization of, you know, their AI infrastructure, which while we don't know for sure, but I think we can look in there and be like,

Matt Ray (25:14.14)
Yeah, yeah.

Brandon (25:40.221)
Well, because your anthropic is, you know, using you as one of your their providers. So probably most of the utilization of that comes from anthropic. And if not most, a giant share of it. So as an investor, you're like, or just a Google, you know, I don't know, you know, Google employee or whatever. It's like you kind of think to yourself like making tons of money. Everything's great. I guess long term, I guess the you know, the open question we keep coming back to is sort of like, well, all right.

It's good, but like at the same time, you know, you wish that was Gemini, you know, and you wish all that was, you know, basically them using your own model. so I don't know, it's kind of a I don't know, it's one of those, yeah, I don't know, it's golden handcuffs or whatever. It's like, I don't know, it's a weird problem to have. It's like we're making tons of money. We're just could also be funding our competitor at some point, but at the same time, it's like we've invested in them. So maybe we've at least protected ourselves a little bit. so I don't know. I mean, Google's in just a weird spot there.

Coté (26:36.888)
It's it seems like it y i the in the not the worst case scenario, which you were just outlining that you're like funding your your eventual death. But then the kind of middle case is you have like you know, as everyone knows, I don't know about chips, but you know, you've got like a TSMC thing where it's like we're kind of the only people who make stuff except those wherever they are people. And so like even though you're like making chips for a whole bunch of other companies that go and sell them.

Because you're the only one of the only people who can actually do that, your position is pretty secure. And so similarly so sim similarly, like if Google is one of the only people who can build out and run like huge data centers that you would need for a massive AI company, then somehow the massive AI company would have to get all of that money, which maybe they wouldn't, right? Like not every company that like d has makes a bunch of money off chips can like actually make the chips. They want to centralize it down to something. So

Matt Ray (27:11.752)
NVIDIA?

Coté (27:35.504)
I don't know. I I I imagine Google doesn't like thinking of itself as like an MSP or something, but maybe, you know, at least they have a good business doing that.

Brandon (27:41.981)
I do like, you know, I would say this though. I mean, I guess it's the TPU, right? So that in theory it's it's not the NVIDIA stuff. So they so it's a great case, and of course they use it for themselves, the you know, the obvious Gemini. So it's a great case study, and like, hey, we do have a chip that is really good at running AI workloads and inference, and this is a very viable way to do it. So that's good. I mean, that's sort of proofing concept. And of course they're gonna keep

know, they have a lot of money. They have a long game to play. So the TPUs keep getting better. And maybe over time that gets to be more parody. And then of course that also all this money, you know, it's not like anything is over. I mean, they can still still make Gemini better and they can still increase it and maybe figure it out. So it's not, you know, it's one of these things where it's like, I don't think it's like like it's clearly gonna be foregone conclusion that anthropic wins, but it's just sort of like at the moment, you know, the anthropic workloads are the ones that are likely winning, right? A are being used.

The most. So it is kind of like going forward, it's kind of an open question, right? And then, of course, like you know, this is the Ben Thompson thing. I mean, at some point, this all comes to a head, right? It's like, you know, you know, the circular because it's not just Google, it's like or Alphabet. It's like Amazon's invested in anthropic. I think Google's invested in anthropic, Microsoft and Anthropic have a relationship. And then, of course, Anthropic is buying the chips and simultaneously competing. So at some point, there's gonna be this break where.

Anthropic and OpenAI will have to go their own way and say, like, we're building our own data centers and we're building our own chips and we're not going to, you know, pay for pay you to run these anymore. Right. And then that's and then that's the moment where, you know, somebody, you know, Google and Google is in the best position. Like if you think, you know, they're in a dilemma, think about, you know, another news story this week is Amazon shut down their AGI lab. So they basically are no longer building a frontier model that probably none of you ever used. So they're out of that game.

Matt Ray (29:13.733)
but

Brandon (29:31.322)
And then, you know, Microsoft has never been able to like build a frontier model that actually competed. So I mean, if you're them, you know, you you kind of have nothing, right? You just have all you have is access to the open weight models and running the infrastructure. And then of course, not surprisingly, they're suddenly very in favor of open weight models, right? Well, of course you are because like you don't have any, like you have nothing else to serve. So what are you gonna do?

Matt Ray (29:55.934)
Course that's that's also like the same week that, you know, potentially a a Chinese model is is you know, leaped ahead to the to the forefront. at the same time, you know, they're abandoning their frontier models and and becoming MSPs, as you say. I I think it's kind of interesting, but you you you said that you know, open AI and and Anthropic might have to buy their own data centers. that's part of the the whole problem though is like

They're not buying their data centers, they're building financial devices to, you know, have something by the data centers if someone else hold the debt while they get to rent them on the promise of being profitable in the future. It feels shaky.

Brandon (30:40.924)
Yeah, well, this all I mean, at some point this all has to come back to like I think we talked about a while back, like final demand. At some point, someone's got to be profitable. And it's like, you know, I don't know, you know, it's like anthropic. I thought Ben Thompson had a good thing, but I said something around like, you know, as you move down in the models, if you go from like, I guess, Fable and you go down to like what high haiku or one of the others, it gets significantly cheaper. Right. So the question, I guess, is like, what are their margins? Like,

Where do they have margin? Are they making a lot of money on the cheaper models? And will this just and this would be the big the bull case would be like as these models, you know, get more optimized, more distilled, better, their margins go up and up and up, and the usage goes up and up and up, and therefore they are tr cr truly a profitable business. Or the bear case is like, no, these models get distilled by the Chinese or someone else.

They become even cheaper to run and they can never get to a model that has margin. And then then we get into this situation, which I think is like the weirdest chain if you follow it. It's like first OpenAI and Anthropic built all these models by scraping the internet, probably taking a lot of liberal use of copyright. Everyone gets mad at them. So they make their models. Then the Chinese, violating their quote terms of service, distill their models down to produce new models.

Right, that are even cheaper. Of course, they're then mad. Then those models come back to the US where the hyperscalers use those to host and provide them. And then that's where the money is actually made. So it's like so it's kind of like everyone's like like somehow all of this has been like right, like everything has kind of been stolen. But at the end it's just like, well, we as consumers, we get good models. We do just get we, you know.

Matt Ray (32:19.593)
Stolen fru stolen fruit of the poison tree all the way down.

Coté (32:22.628)
That's true.

Coté (32:28.186)
It's it's like it's it's it's like when your dog runs around and eats like cat poop and then a few mo moments later throws up or poops. It's just like this continuous like cycle of of things. But yes, that is like I think I think you would call it rich or whatever, where it's just like really you're complaining about someone using your models to distill their models after you used humanity's models to distill your models. Like it really is just I don't know. It's like it's it's yeah.

Brandon (32:36.198)
Yeah.

Matt Ray (32:54.803)
Fair use, man.

Brandon (32:55.496)
Well it's also funny too because they're like they could like there's no one to complain to. They can complain to like the Chinese government who's like we don't we don't view copyright though, or we don't view terms of services are they I guess I would say this like we view terms of service much like you view copyright more optional, you know? And it's like so there's nothing, you know, like what are you gonna do? Like you're not gonna prevent them from doing it. So I think it's just like you just have to accept it if you're you have to build a business with the constraint of like, yes, people are gonna distill our models. But that kind of does bring me something I wanted to ask you guys this week is like

Matt Ray (33:11.473)
It's cardrails.

Brandon (33:25.152)
I have found that my own personal use is like because I switch the models in Claude because I'm cheap, right? To prevent my usage going over. But I find and I've tried some of the other like codecs and others. It's not and it's not that I don't like those things. It's just like I'm pretty attached to Claude code, the harness, right? Because I have the my skills and I it all works the way I want. And I'm actually okay. Like what I don't want is I don't want to leave the harness. I'm okay switching the model, right? Although it me, you know.

It's not super easy. It can be done. Like you can switch to like an open weight model for Claude. It's like it kind of works. But but this is kind of brings me back to like, no, I think the harness is like, I guess the harness is the thing I actually care about the most and I don't really want to switch from. So I don't know if like going forward if the value is like maybe more there than the model. I don't know. What do you think, Matt?

Matt Ray (34:20.217)
when I mean th I I I think the harness you know I I think the harness is definitely the strength. And when when I looked at the you know the the comments on like GCP's usage and I thought well they're

probably very efficient with how they leverage all their different models because they're not just saying like you know YOLO it on on the the the best Gemini model for everything it's like no they they know very very clearly which models to apply across their infrastructure where to use which ones and and how to outsource that to their consumers and

Brandon (34:54.834)
Right.

Matt Ray (35:02.141)
you know, when when you you let a consumer, you know, a customer pick, they always just like, you know, give me the best. And then they, you know, promptly use it to like spell check an email. And and so

Matt Ray (35:19.484)
I think what what we

probably need to be doing is is invested in owning our own harnesses and and building out that logic of like, well, I have a tool, I have a tool chain that understands how to route between different models, how to have some memory, how to build all that out. And that way I can switch out the LLMs underneath because at the end of the day, like I can't tell you which LLMs I'm using. I'm switching constantly between them and I'm just using the interface in front of it. And if I'm going to

to go all agentic on things. I need to I need to invest in my my interface, my my harness rather than the underlying tools. I didn't post it in in in

and Slack, but Wiz had a interesting article about their their latest agentic framework for vulnerability, you know, for for finding vulnerabilities for pen testing. And it beats out most of the frontier models. it's currently like the leaderboard and and it what it is is a harness that's a stack of different LLMs specialized in all their different things and a tool chain that double checks the work as it goes through. And so like

At no point do they're like, it's all, you know, Fable five dot six or whatever. They're like, no, it know, it doesn't matter because we're gonna keep switching out and tuning this engine to to go faster and faster all the time. So don't worry about the Frontier LLMs, just worry about the harness.

Brandon (36:50.834)
But I think the thing there, I I think I just wanna before we go on there, before you go, Cotes, is like I think the the key thing that I've been thinking so much about the harness though is like inside the harness is the prompt, right? And I bet in the case of Wiz, the real value is there. They've written excellent prompts f by pen testers in like and this is where I go back to like, you know, I I like your idea about you know, you know, being mod you know, kind of harness agnostic, but you can't really like they don't publish the prompt. Like so you

Coté (36:50.854)
Yeah, I I I I

Matt Ray (37:08.317)
yeah, absolutely.

Matt Ray (37:19.014)
No. L L agnosti. Yeah. Invest in your harness. Invest in your harness. Yeah.

Brandon (37:20.358)
Yeah, I know L but it's like, but we're but that well, that's the whole thing. The butt but it's the prompts inside the harness that they're not like like Claude doesn't this openly show you like here's the prompt I'm using, just like Wiz isn't gonna say. Here's the prompt I'm using for the pen test. And it's like, and I think that's like it's almost like a personality or something. It's like, but that's the thing that I've found that like I think I'm really I really like. It's like, this is pr these system prompts I kind of get now.

And I don't wanna lose those. I don't wanna go to like a different personality. But go ahead, Kote, interrupted you.

Coté (37:50.961)
Yeah, I I mean I I think I think what what you're both saying is yeah, like how how the the software of how you use all these models clearly is like what becomes important. Like and and then the it brings back, remember prompt engineering from two years ago? That was gonna be the whole thing. Just whatever you type into this box, that's what's key. And yeah, I'm I mean I think like I I I I I for some re

Brandon (38:04.583)
Uh-huh.

Coté (38:17.668)
I went and tried to use like goose hooked up to a local LLM and then I used LL LM studio and you know, I moved to a different harness from what I'm used to and you can just feel all the stuff that's different or just purely missing, right? And you can of course build up and put all that stuff in then I have a lot of coworkers who are of the you know, they're developers, so they have to build everything themselves. So they they go f they go build out their own like harness setup and they use

open this and open that and things like that and they're building things out. But it is like all the all the stuff that you wrap around it is what becomes important long term, which is kind of confusing commercially now because I don't I don't know of like a lot of vendors that sell that necessarily. I I don't know. Maybe maybe they just don't sell it directly. Maybe I mean I mentioned them earlier. I don't really know what Palantir does, but maybe they're just a big harness around AI models.

Matt Ray (39:16.667)
Mm-hmm.

It's like Salesforce for for government agencies.

Coté (39:21.208)
Right, right. Yeah. Even better. And and so just like, you know, how it's it's similar that I guess like, you know, a database on its own doesn't do anything. You need the stuff you need the stuff that you have wrapped all all around it. Now, continuing our joke of of dogs eating cat poop, like I have been thinking when it when I was wanted to make those like using goose and LM Studio better, like you could probably set up some agent to like use Claude Code for a week and be like reverse engineer this for me.

And like pull out all the stuff and and build up a harness. No, it would probably be a little weird, but like, you know, that seems like the logical thing that that harness people would do is just like, let's bust this thing open and go look at it. And isn't like isn't I mean all the all that cloud code stuff is written in JavaScript, right? So it must be like in plain text and decodable. Like

Brandon (39:49.639)
Mm-hmm.

Matt Ray (39:51.74)
Yeah, absolutely. That's distillation.

Matt Ray (40:09.702)
Yeah. Well like TypeScript. Yep.

Brandon (40:11.282)
Well they did. And there was a a while back they like accidentally shipped all the harness stuff. So there's a point in time where you could go read it and see what they did. So so you have to do it. But I I think this just kind of gets to like I I guess it's kind of back to like where the value is because it's like, yeah, you can do all that or you can just use cloud code. And then you can just switch to like the cheaper like it's like the version of cloud code is just like, Well, you know, I I think Matt's right. It's like most of the time everyone opts for the most exp the best model. But then if you do it for a while, you're like, especially if you start to like understand it's like, Well, actually I d I can you I actually want to go the other way.

Coté (40:17.393)
Yeah, yes.

Brandon (40:40.402)
For all my proofreading, I want to use the cheapest model. And in fact, I would like Claude to just automatically like be like, Brandon's asking how to spell a word again. Let me give it like super, you know, simple haiku slash whatever and like charge him like, you know, a fraction of a cent. and so that's where I think it's like the the that there may be more defensible mode, right? If Claude just keeps the price cheap enough, I don't want to, you know, I don't want to leave. Like I'm happy. Right. But if if it gets to the point and it's like, well, it's just so expensive and I just can't afford it, then I'm like,

Coté (40:43.536)
Yeah, yeah.

Matt Ray (41:08.808)
How are they gonna get to one point four trillion at with that model?

Brandon (41:13.308)
Well, but I think I don't know, but I like I guess it's one of those things. It's like there is you know, there's a time and place for everything. It's like when you do something important, I'd like the fact I can flip back up, you know, the to the expensive model. And I just think, and I guess it's just back to margin. It's not like I'm not willing to pay anything. If you I I'm it's fine. Like as long as I'm getting my work done, I'm fine with a cheap model. And like you can like charge the margin. Like, just don't make it crazy expensive. Like, you know, you can only use Fable for so long or you have to be rich, right? You're like, okay, I'm out.

I'm not smart enough to use this anymore. My prop my questions I'm asking or things I'm doing are not worthy or not fable worthy. I'm gonna be sonnet level, right? Or I'm gonna and then you know what you said before too, Kote, you can do a lot of this stuff. You can get a lot more out of these models just with you writing better prompts, right? If you write like kind of back to Matt's example with Wiz, it's like if you give me a pen tester and I spending a lot of time with a pen tester and writing the prompts, absolutely those prompts are gonna do better with an inferior model.

than a great model with like, you know, no prompt from an expert. So

Coté (42:14.184)
So so to to cap this off, we we touched on this, we we talk about this every now and then, but with with Amazon seemingly shutting down their lab here, you know and and other people, like our are what's our theory on how many major like models are g that are gonna exist? Or is it is it gonna be three? It seems like it seems like, you know, and you don't have to say anything, Matt, but it seems like Google is always just gonna be there doing it, like because they probably use a ton of it themselves.

I guess you always have Facebook there, which I'm always like, I don't understand what's going on here. But then you've got open AI and Anthropic. And like, are we always just are we gonna have right, right, right, right. Yes, yes. And so

Brandon (42:52.55)
OpenAI, Anthropic, Meta, probably, Google, of course, and then the Chinese labs. I think those are your and then, you know.

Matt Ray (42:59.653)
Well and and the the Chinese labs are at least as many as all the American ones. I mean I you know, th that that's doubling it at least.

Brandon (43:04.21)
Sure. But like I think you just kind of like put them in a whole category, right? So

Coté (43:08.988)
But there's not th so so what what do we say? Like maybe there'll be like five or less? That's that's

Brandon (43:13.928)
There probably is. I mean, it's it's given how much money it takes, it probab I think it's more like on the or I mean the Chinese labs are hard like figure But like yeah, I guess you could say there's gonna be two three of those companies. So like I would say Yeah. But I don't think I don't know. It doesn't feel like it doesn't feel like they're I was gonna say they they feel like they're they've hit the escape hatch and they're going with being a a a great they're great at building the data centers. That seems like the way that they're going to

Matt Ray (43:15.526)
No ten.

Coté (43:16.016)
Ten? All right. All right.

Coté (43:24.633)
and then they and then there's SpaceX. They've got AI too, right? We could have had them in there.

Matt Ray (43:29.863)
I I I think they're they're getting outsourced. Yeah.

Coté (43:40.252)
Mm.

Brandon (43:42.448)
Of course, you know, Meta to be fair, Meta did the same thing. Meta has agreed to lease a lot of their capacity to anthropics. So that one is probably like Meta's maybe like third or fourth, but we should like we put like a little flag there, be like, hmm, we should check back in. Not not totally sure they're committed, right? You know.

Coté (43:53.861)
All right. All right. I like our we should we'll in in in ten years we'll we'll have to revisit this. We'll see was there five, three, or ten? What or or more? What what numbers are we got we got Matt Ray for ten. I think I think I think you and I, Brandon, we're like maybe five. I'm gonna go for three. There's just gonna be three.

Brandon (44:06.728)
Mm-hmm.

Brandon (44:11.888)
Yeah. I think well that would be like that would be Jeffrey Moore, right? We'd like, you know, one of his bucks we'd be like it usually ends up with three, like both both most big markets condense down to three and then a bunch of tiny little, you know, specialized models. So

Coté (44:24.356)
We folk.

Matt Ray (44:25.359)
I I I think we don't care about the L LMs in the future.

Coté (44:28.907)
that's good. Yeah.

Brandon (44:30.386)
But to that end, before before we go, I did want to get your your take. I spent the better part of a week trying like doing the full anthropic training, like going through and trying to build out an agent and like you know, kind of like I was like, What a what is a forward deployed engineer? So I'm like, I'm gonna take the I'm gonna take the training. I'm gonna try to like be I'm gonna try to like act like one at least for a week. And so having done all of this stuff.

Matt Ray (44:31.407)
I think it

Coté (44:39.15)
nice.

Matt Ray (44:40.688)
Yeah, yeah.

Matt Ray (44:44.293)
I need that.

Coté (44:48.156)
Like I'm gonna be one.

Brandon (44:55.022)
and again, of course, like you know, you you I just feel like I'm just rediscovering the most basic thing that of course somebody you made a video on. It's like, you know, I actually having done all of it manually, I'm like, really? You know what's really the way that most of the value is? It's just the skills. Like most of like when you hand code an agent, create a prompt, manage the conversation, build some tools to like go out and do it after you've done all of that work.

It's like, you know, I could have just written a skill and used Claude code to mostly just do this. And it would have taken me, I don't like 10% of the time. So I was like, my working theory was just sort of like, I think most of this productivity, kind of back to where we started around like all these things people are trying to do. It's like just writing a skill or using the skill builder prompt and in Claude, I think gets you like ninety percent of the way to most of these productivity gains. So I don't know, you know, Kate, I know you've used a lot of skills.

I like do you find that? Do you ever find yourself like I really wish I should build an agent instead of just write a skill?

Coté (45:54.447)
Yeah, you know, I I sort of don't even know what building an agent means. And in well, to answer your question that way, I guess I don't r I don't think I ever build an agent. Now, I I don't I also don't necessarily build skills. I have Claude build skills for me. And there's there's there's maybe two two things that that goes on there. One, there are there's there's like

Brandon (46:12.935)
Right.

Coté (46:19.96)
markdown skills that I have that are pure markdown, and then there's markdown skills that have a bunch of Python scripts that I didn't write.

Brandon (46:26.162)
Wh which is a tool, which become is the new way to do skills and you just put all in the same folder. Yeah. Very common.

Coté (46:29.084)
Right, right. And and so it's I think I think mm probably I'm gonna say sixty percent of the skills that I have have Python scripts with them. I I don't know. It's hard it's hard to tell. But yeah, I don't maybe I need to take the course, Brandon, because I don't really know. If someone were like, Go build an agent, I would to answer your question, I would be like, Okay, I'm gonna go to Claude and tell it to do this for me. I I I'm gonna type that.

Brandon (46:55.132)
Well, the reason I was thinking about this, I don't know if anyone else remembers this, probably not. But remember there was the big Google keynote and I think we talked about in the show, they were like doing like they were designing a marathon course, right? And they were right, and they were going through and they showed how you would build like, okay, we have like a like a marathon course agent, and then we have a marathon course like like simulation runner, right? And so they were showing all these agents and the reason so to demystify because I don't know.

Coté (47:08.318)
that was one of your favorites.

Coté (47:20.453)
Yeah, yeah, yeah.

Brandon (47:24.198)
Maybe everyone listening to this knows this, but I get it now. It's like writing an agent basically just means like you get an anthropic key and then you're gonna use some programming language, but then you're taking on the responsibility to basically get every message, right? And then append all the responses and then add the new messages and go back and forth. Like you're taking on that, that responsibility. Okay, because you're doing all of that work, all the work that Claude Code would do, you're doing all of that. Okay.

Coté (47:46.542)
Mm, okay.

Like all the orchestration of of different things.

Brandon (47:51.769)
All that orchestration stuff. And then, of course, you can work in it you know, your own adverse observability or other things that you want. And then, of course, like if you're in a programmatic environment, if a human was outside the loop, this is kind of how you have to do it, right? Because you're not there in front of it. But if human in the loop, right, and this is, I guess, I didn't, I just dawned on me like why that marathon or the marathon running thing was like so crazy to me. Is like, no, you would just do this with a skill, right? You would just like now if someone said if

Coté (48:05.658)
Yeah, yeah.

Brandon (48:21.118)
that presenter had said, like, hey, we're gonna build a skill on how to create a marathon course, right? And I'm gonna sit down and in their case, a codex and like run the skill. And then you would do some prompt. You'd be like, I looked at the map. it turns out I don't want them to go this way, change the course this way, do this. And then you would run a similar, it's like that makes sense to me. Like, you would be like, okay. And then you could put the skill in GitHub and you could share it with people, and people could kind of use it. But this idea that like,

I'm creating an agent, then you have to deploy the agent, of course, right? And then you have to deploy your own chat interface to the agent because you've taken it outside of the harness, right? So it's like, why would I do like I get it? Let me just let me preempt people emailing. I get it if you're doing something completely programmatic, right? That you're just like, I need to do this, and there's no human in the loop. But for almost everybody else, I'm like, just write the skill and use and then all that stuff, all the stuff I said, all that minutia, you just use your.

Coté (48:50.586)
Mm-hmm, mm-hmm.

Brandon (49:17.276)
Harness of choice, in my case, Cloud Code or Codex, and you just let it do all that stuff because it's already figured out how to like manage the conversations and like you know help you with the prompts. And it's like so, so I was kind of like, and again, there's a whole video, funny enough, from Anthropic where they basically these presenters say, Don't build agents, build skills. So if you want to see intelligent people give you that argument, you know, they do it in 15 minutes. And I was like,

I probably should have watched this before I spent all this time trying to build an agent. Or maybe it was it was it was good exercise to do, but I was like, but I guess having built an agent, I was like, no, this seems wrong. This seems wrong in most cases. Don't build agents. But love for someone to tell me what they're doing and if like I'm just like off base.

Coté (49:58.727)
Well so so an important follow up question, Brandon. Are you gonna take the test to get your certification?

Brandon (50:02.984)
I wanted to. So I was like, I'm ready. I'm going to do it. Cause I was like, I want to hold myself accountable. And I'm not going to cheat. I wasn't going to cheat. I was like, I'm going to hold myself accountable. It was it's proctored. I was like, I'm ready. But like the only way to take the test is you have to like be at a company that's like a partner with anthropic. And it's like, and so I was like, I don't, I don't have, I don't, I don't think we have we need seven more co-hosts, which we're not adding to this podcast. We need to be 10 plus company and then.

Coté (50:08.774)
Put it on LinkedIn.

Matt Ray (50:12.472)
Isn't isn't the exam proctored anyway?

Coté (50:21.708)
That doesn't make sense.

Matt Ray (50:30.64)
Well, maybe they c maybe they could be virtual co hosts.

Brandon (50:32.932)
Yeah, I was like, I don't know, because I I wanted also too, like it was funny because taking the anthropic training, because like they changed so fast. Like one of the things in the training, this little if you if anyone wants to do it, what I learned was in the training, sometimes the they the model they used in the training had been deprecated. So I was like, I can't use Sonnet four. So so what am I gonna use? I'm gonna use Sonnet five. Like it's all gonna work. It's just AI. It's like, no, it turns out they changed all the API parameters between these models. And I was just like,

Coté (50:50.534)
Mm.

Coté (51:00.368)
That's funny. Architect this.

Brandon (51:01.308)
Which is then funny. So then I'm take then I am back to Matt. I'm like kind of cheating. I'm like, none of this is working. Cause I cause I did the thing where like, and I'm sure Matt, you've done and Kote as It's like when I'm trying to learn something, like I don't just watch the person type it out. Like I type out all the code myself, right? That way I'm forcing myself to do it. And then of course I'm like convinced I'd like to type something wrong. And then I give, you know, after 30, 40 minutes, I finally put it into Claude. So I am cheating there, Matt, to your point. And then Claude's like, turns out that's been deprecated. So I was like, so a little thing if

The training says like sonnet four and don't just put sonnet five in there and expect it to work because there's going to be some gotcha. So yeah, if anyone from anthropics maybe, maybe anthropics like weeding us out. They're like, Yeah, sucker. That that was, you know, that that was that was like the the professor embedding the white script. It's like, yeah, I wanted to make sure you did typed in the work and actually found the errors. So I did. I found the errors, but a little heads up, that's right, but a little heads up, just like they could say something like, Don't use the like.

Matt Ray (51:51.43)
But Brandon, you learned. You learned.

Brandon (51:58.39)
And what I figured out was like if I deprecated, if I'd used a a model of the same number but a lower model, that worked okay. But if I went above, so

Coté (52:04.666)
I like this. You you it turns out you were an FDE for the FDE course.

Brandon (52:08.656)
Yeah, I was like, but of course it's so so ironic to I was like, you what they should do? They should have Claude go through and take all the training and see if it actually worked. You know, so it's like they they should see like what happens when they try to use Sonnet four and it's deprecated. So it's it's a good experience, but it left me thinking to myself, there's not, I don't know, building the agents in most cases, just build the skill and you'll be much, much happier, is my my opinion.

Matt Ray (52:15.45)
Mm-hmm, mm-hmm.

Coté (52:32.944)
Well, y you know, we we we went o I think we talked a lot about AI stuff today, nothing as always about SNMP and MIBs, which always makes me sad. But the y you get a lot of benefits in automation like we were talking about, but there's one thing I know people don't want automated, Brandon, and that's bureaucracy. They they enjoy it. Do we have any this episode?

Brandon (52:49.554)
It's right. If only. A couple of things. I sent stickers to a few people. One to Andy in Alabama. He's going to go to We Are Developers. So he's got stickers to represent. that's the conference coming up this fall. Also sent excuse me, sent stickers to, I don't know how to say this. Maybe Kote Ido is it? Vil Vilness in Lithuania. And he's going to be attending that DevOps day in, I don't even know how to say it, the city that starts with the V. How do you say that city, Kote?

Coté (53:19.11)
I don't know, I've been saying it on the podcast for a while, saying Vilnews, but I think that's a French way of saying it instead of Lithuanian.

Brandon (53:22.374)
That's probably wrong, but that's what he's not gonna be there, but he's close by. So he's disappointed he won't be there. But he wrote in to say, after listening for a hundred to two hundred weeks and thinking about asking for stickers to place on my bicycle luggage or laptop, I finally decided to request some. So there you go. That's exactly what everyone should do. And he wrote in some nice things about all the things that are going on in that city. Actually, very tech centric, like Austin. So very cool. So listen, if you would like a sticker and if you've been listening for don't wait a hundred to two hundred weeks. You know, I've got them right now. You know, I know there's lots of people like,

we'll never send me stickers. I don't have time. Yes, we have time. Just send me your postal address at stickers at software defined talk dot com. Happy to send you a sticker anywhere in the world.

Coté (54:01.316)
Well, coincidentally, in our conference section, I'm gonna be speaking in in Riga, Latvia, which is close, theoretically, at AI Connect, which which will be exciting. that's August twenty second. and I think I've been there sometime recently. Anyways.

It's it's it's a similar situation that it sounds like is going on in Lithuania there, where there's there's a whole lot of there's an interesting tech community in Riga. Like you go there and there's a lot of consultancy people there. It seems like a lot of hiring abroad that happens. A g interesting set of people there. So lots of lots of tech people there. We'll see what's going on with the the the AI conference. I finally this will be a a renewal, a renewal, a an update to my playing D D with AI stuff.

Enterprise edition, since I've actually done a bunch of enterprise y stuff since then. But that's at AI Connect in in Riga on August 22nd next month. There's also DevOps Days Graz, which is September 4th to 5th. The Cloud Foundry summit, which is two days this year. That's why it's a summit instead of a day. on September 21st and 22nd in Heidelberg, Germany. I'll be speaking there, of course. And also my company, Tonzu.

we'll be having like a a dinner for people and lunch stuff as well. If if you you would know if we would invite you or not, you should ask if if you're gonna be there. there's also DevOps Days Rockies coming up September 22nd and 23rd. And then we are Developers North America in San Jose, September 23rd to 25th. We've got a discount code, DevPod26, and maybe some free tickets left if you go look. There's also a discount code for DevOps Days Rockies if you want.

There's several other DevOps days in September. There's one in Dallas, one in Vilnew, and those are the end of September. Then there's also DevOps Days Istanbul, where I'll be giving a keynote October October 24th, and the one I know everyone's waiting for, the VMware User Group, Orlando, October 20th to 22nd. And then I'll also be giving a keynote about goblins and AI at Cloud Native Denmark, November 19th. And that's what we have so far. Now

Coté (56:18.488)
If if you know of other conferences, you should contact us about contact us about and we'll we'll we'll look into those. Now, speaking of things that we'll look into, in your recommendation area, what are you going to be looking into, Brandon?

Brandon (56:31.12)
Well, I think I've recommended everything about this the show in the in the books as well. So I'm just gonna do my obligatory silo season three is out on Apple TV. It's very good. It's based on the book wool or the short stories. So if you haven't been watching, got two, you got yeah, you have season one and season two already sitting there. Those are really good, so you can catch up. And there's that it's gonna conclude with season four. So we know there's an ending in sight. I really enjoyed the show. Looking for something to watch, check out Silo season three.

Coté (57:00.156)
How about yourself, Matt Ray? Recommendation.

Matt Ray (57:02.405)
my my recommendation this week is a book I'm I'm rereading, Hyperion by by Dan Simmons. If you haven't read it, I yeah, I'm not spoiling it by any means. It's it's basically a f futuristic version of the Canterbury Tales where you've got, you know, a a number of

of travelers all on a on the same journey and how they got there and their backstories and everything. I think it won like the Hugo Nebula, you know, all the the big awards for science fiction. it's not a new book, you know, it's thirty years old or so. But as I'm rereading it, you know, it it's

It's it jumps out at me how well it would adapt to like a an Apple TV series like Silo, where you know, it's way too convoluted a story for a single movie. and so I you know did a little little background research and apparently it's been sitting in production hell for for like two decades.

Coté (57:57.66)
Mm.

Matt Ray (58:11.275)
So so don't hold your breath, but it's it's a it's it's a great book. It's it's one of my top science fiction recommendations.

Brandon (58:18.184)
I feel like we gotta get on the phone to the new Apple CEO, be like, hey, this is what we pay you for. Like you make a lot of money on all the stuff that and we make a lot of money on these cables. And in exchange, you have to go make all these sci fi books in a good series. That's the deal. That's how that's how it is. And yeah, 'cause I think they're making new new romancer. Was it what's the one? Yeah, that's coming out. So it's like, listen, you owe us, you know, you know.

Matt Ray (58:25.934)
Yeah.

Matt Ray (58:33.165)
Yeah, yeah. apparently Bradley Cooper's had it for fifteen plus years.

Nermancer, yeah.

Brandon (58:43.302)
These these lightning cables, they were a rip off. You owe us some good TV in between. I don't expect you to make money on these series. That's right.

Matt Ray (58:43.493)
just

Matt Ray (58:48.397)
Yeah, keep keep it away from Amazon and I'm happy. Yeah.

Coté (58:53.23)
Well well I got excited there 'cause I thought it was the Hyborian age, about Con Conan, but that's the that's that's different.

Matt Ray (59:00.399)
That you know, that that's that's easy T V to make too, yeah.

Coté (59:02.916)
Off off by a letter. And otherwise it w would have would have been something. Well, my recommendation is I'm gonna call it cooking with Claude. And every now and then, you know, I ask actually a lot, I I ask the AIs for a recipe. Cause then you don't have to read about like someone's uncle and their story about it. And it just gets down to it. And you know, I've noticed over over the over the past

Matt Ray (59:21.669)
Code is like I I I want no humanity in my cookbooks. Thank you very much.

Coté (59:28.816)
That's right. Recipes are not made of people. And and you know, I've noticed it used to, you know, all the AIs, they would just spit out text, but Claude has started doing this thing where it gives you like, I think they call it an artifact. There's this little app in the chat thing that one will like list all of the recipe, but then you can click start and it opens up this little this little card, they used to call it. And it has like each of the steps of the recipe. And it even has things like tonight, one of the steps was to boil an egg.

And you can set a timer, which then reaches out to your OS and sets a timer on your phone or whatever. And so, you know, they're doing some interesting little app building there. But of course, it's also important that the food taste good. and I think I wanted some recommendations for like, you know, it's pretty, it's it's hot here. And of course, we can go over it again. No air conditioning here. but so when it's hot, it actually is hot. And I wanted to have like some cool summer dishes. So

Brandon (01:00:11.857)
Mm-hmm.

Coté (01:00:25.572)
It recommended having gazpacho, which is good. We had that. And then it recommended a potato salad with mackerel on top and I was like, No, no no mackerel, not

Matt Ray (01:00:36.164)
And then it was like, You can make your cheese pizza extra sticky with glue

Coté (01:00:41.052)
Exactly. And then and then it had a had a recommendation for cold sesame noodles. Now there was a flaw in all of this, which is it required several hours to cool down things. However, the resulting recipes were good. And it's also nice, of course, because you can go back and forth about about doing things and ask more questions. But I think that little app they have in there is is is kind of interesting. I don't know. I'll I'll read through the session that I had to make sure there's nothing embarrassing in is it in it, and maybe I'll share a link to it in in the show notes here. You can see.

Gazpacho by Claude. and

Matt Ray (01:01:14.638)
Well, be be careful. It might already be getting indexed by search engines, apparently.

Coté (01:01:17.584)
That's true. that reminds me though, speaking not being careful, but the other cool thing that it did is is you can say, Can you add make a list for me of things to buy at the grocery store? Of course it can do that, but that would be fine. But then it can actually call out to the reminders app and then add that to your groceries list, the which auto sorts it, and then you have it all there, just ready to go.

Brandon (01:01:39.996)
I think Kota, you gotta go one more in your skill, one more toll. Like have it look connect to your what is what's your grocery store? The yeah, like have it automatically, you know, order it and have a a delivery. Come to the that would be the final step.

Coté (01:01:46.372)
The the plus.

Coté (01:01:50.869)
love it. Yeah. Yeah. Yeah.

Matt Ray (01:01:53.356)
Yeah. Just just be careful. Don't don't don't let it order through DoorDash.

Coté (01:01:58.311)
That's true. And and and if and if a bunch of Elmer's glue shows up at your doorstep, check the recipe.

Brandon (01:02:04.21)
You know if it broke in the hugging face, you've gone too far. You know if if if that that happens, you're like, What I th Yeah, I just th it was people thought it was gonna be a paperclip that took down society. Turns out it was this soup made by Kote.

Coté (01:02:06.716)
I just wanted Gazpacho.

Matt Ray (01:02:09.75)
It's like it's like hugging face is delivering gazpacho for me now.

Coté (01:02:13.548)
Yeah.

Coté (01:02:17.21)
That's true. That yeah, yeah. There there's like Kote needs gazpacho. So I'm gonna reorient you know, Hugging Face as a company to make and deliver gazpacho in the in Amsterdam. That's yeah. Well, speaking of good ideas, what you should do, listener, if you've listened this far, is join the the Slack channel that we have at software defined talk.com there. Or you can put a slash five eight three, just look up in the header and join Slack.

Brandon (01:02:29.65)
I like it. That sounds like a good idea for you.

Coté (01:02:45.372)
There's a whole lot of conversation things that we have there that's interesting. Also, if you want the show notes for this episode, you can go to softwaredefined talk.com slash five eight three. You can find links to all those conferences and discount codes I mentioned, the links to relevant stuff we talked about, and many things that we we didn't talk about, and probably a link to Claude's gazpacho recipe. Old old Uncle Claude. This is how we used to cook gazpacho on those hot Amsterdam days.

And boy, it was good. With that, we'll see everyone next time. Bye bye.

Brandon (01:03:23.282)
So every week we come on here and we talk about how hot it is here in the northern hemisphere. But I think this week at the end, we should turn our attention to Matt. Matt, for those that are watching along in the video, will you reveal your hands to us? Show us what you're you're wearing during the recording of this. For those that cannot see, Matt is wearing gloves with cutout fingers and thumb. It's like I what what's going on? Matt, give us the backstory. What what what's going on in your home to require this?

Matt Ray (01:03:50.062)
well, you know, it's it's July, so clearly it's winter in s the southern hemisphere and my my office my home office is in kind of the the basement, if you will, and so it's cold. it's always colder than everywhere else and we don't have centralized heating in here, so I think I don't know. it's

like forty eight degrees outside Fahrenheit, you know, for for for those of you. And so yeah, I I'm I've got a space heater under my desk. I've got a blanket trapping the space heater. I've got fingerless gloves. I have a dehumidifier. it's full on winter in here.

Coté (01:04:24.22)
Mm-hmm.

Brandon (01:04:31.656)
I'll just say as long as we do this podcast together, I will never accept the fact when you say it's cold in July. Cause it's of course, you know, where where I'm at, it's right now at late afternoon. It's roughly ninety five degrees, cooling off from over a hundred. So it's just like the fact that you could be somewhere anywhere in July and be cold is like I my brain just cannot comprehend that that happens.

Matt Ray (01:04:45.497)
Ha ha.

Matt Ray (01:04:54.5)
Brandon, I'm wearing long johns.

Brandon (01:04:56.654)
my gosh. That's even my gosh, that's crazy. So well listen, hopefully mm-hmm.

Coté (01:05:00.336)
Well, lo let me let me ask you that what what what do you have a dehumidifier in winter for? I'm sure it makes sense. I I just don't I don't know it.

Matt Ray (01:05:09.596)
the the cold damp like everything just is kinda damp in here. I mean I'm also like near the coast, so you know, humidity's always like seventy percent plus and so I I I just yeah, like to

Coté (01:05:12.796)
Mm.

Coté (01:05:22.8)
I see. I see. I guess I'm not used to that growing up in Texas. It's just like it's it's cold and then and then dry. Hmm. Yeah, I wonder if it's like that here. Huh? Dry cold.

Matt Ray (01:05:28.938)
No, no, it's very different. Yeah, yeah.

Brandon (01:05:34.364)
different kind of heat. It's a different kind of cold. It's a dry cold. No, no one says that either. No one really says that. I don't think that's a word.

Matt Ray (01:05:39.268)
No, a moist cold is is bone it gets into your bones.

Coté (01:05:43.623)
Mm, mm.

Brandon (01:05:44.154)
All right. Well, everyone wish Matt well as he suffers through winter. And I would say roughly eighty percent of the w listeners to this podcast cannot relate to to Matt's dilemma at this point. But you know, there's probably there's probably a g there's a good, strong and s Australian contingent down there that's with you. So

Matt Ray (01:06:00.996)
Don't forget about South America, South Africa. Yeah.

Brandon (01:06:02.576)
Yeah. Yeah. Michael Neal's probably, you know, he's probably like, yeah, it's cold. You know, he's probably the one person we can count on to actually sympathetic 'cause South America although I did see they got two hundred inches of snow in Chile recently. And that was another thing that was like hard to like see. You're like, huh, it could wow, it could snow a lot in the winter in Chile. Didn't really think of that. So all right. Well, thanks for everyone that tuned in and got to see Matt's gloves. If you if you did, that was a special treat. Hopefully you'll watch us again.

Matt Ray (01:06:17.132)
It's a big country.

Matt Ray (01:06:21.796)
Exactly.

Brandon (01:06:32.556)
For those that watch against watch us live, I really appreciate it. And those that watch us after the show on the YouTube, we appreciate that as well. And with that, we'll talk to you next time.

Coté (01:06:40.47)
On the YouTube. Bye.