Dan Shipper on Claude 4, Every's $2M 'SIF seed' round, and why Anthropic dominates coding

everybody thinking that he's trying to kill Google and then goes for the jugular on on Apple. It's very interesting. I think Sam's going after everyone. Yeah, I think he wants it all. And why not just everything in energy, in servers, in chips? Well, we have Dan Shipper. Welcome to the stream. How you doing, Dan?

How's it going? Excited to be here. Fancy you guys. Great to great to have you on. I've been wanting you to do this for a while. Uh and you've had a busy week. You have some some personal news uh at every you're also at you're at Anthropics event right now. Is that correct? I'm at uh Code with Claude. Yes.

Anthropics event where they just dropped uh Claude Opus 4. This is amazing because we didn't get a chance to talk about Anthropic or Claude 4 on the show yet and we have a perfect guest to talk about it. I am your correspondent out here reporting live.

Can you kick us off with just a little bit of introduction on you for uh any of the fans who are watching who might not be familiar and then we'll go into what the events uh they'd have to be living under a cluster. They'd have to be living under a data center, but some people do. Uh my name is Dan Shipper.

I'm the co-founder and CEO of Every uh at Every publish ideas and applications at the frontier of AI. Um so we have a daily newsletter. We write long form essays about what's going on in AI and then we also create software products. We bundle all together and sell it to the audience. Very cool.

Um yeah, and if uh Jordy was describing you as like a Claude stand. You've been a fan of Claude for a long time. We've talked to some people in AI and they've said like you can just walk into a party and immediately clock like, "Oh, that's a chat GBT user. Oh, that's a Claude user. " Uh, is that true?

Are you beating the allegations? Uh, uh, I would say, um, it's interesting. I would say I used to be. I think Claude like 35 sonnet was the first model where you're like, "Oh my god, this thing like actually gets me. " Um, but uh, since 03 came out, I would say 95% of my AI usage is in CHP.

Um, and also with their memory feature, like memory is just incredible. It's very sage. Um, I do think it's tough to have a friend with no working memory. It's like, you know, yeah, I mean, uh, like concretize this a little bit more for me.

Uh, my experience with AI and LLMs is that I'm using them as as information retrieval, agentic reporting, deep research a ton. I never just sit there and have a conversation. So if it's if it's answering a little tursly, if it's dropping bullet points on me, I'm kind of fine with that.

Uh but when I talk to some people about the differences between clawed and open AI, a lot of it is in the subtleties of the language. And I just feel like as a consumer, maybe I care about that less, but do you think that's important? Does it matter to you?

Am I missing out on something by not I'm just like a very sensitive, emotional guy, and so I want it to be nice to me, you know? Um, but I also think um to some degree even for uh like more business like stratified business use cases, it's actually really helpful for it to have a little bit of that sensitivity.

For example, something I use it for a lot is if I'm having a management problem.

Um, we report almost everything that we do at every in Granola and it's really easy to take your transcript of a meeting and be like what happened here like why did why was there a fight here like or you know an employees having a having a problem like how do I help you know that kind of thing that makesense and it sort of has a little bit of that radar that can that can really um help you move through a lot of interpersonal situations that come up often.

Okay. So that's interesting. Well, before we just go and and just uh into a much longer discussion and and talk about Anthropics news, why don't you talk about there's a couple things this week. You had your own fund raise and it was a unique structure. So, why don't we kind of start there?

Yeah, we announced this week that uh we raised a little bit more money. We've we previously had raised about 700K in 2020. Uh and then we raised uh just now we announced uh a $2 million raise uh led by Reed Hoffman. Um, but we did it in a weird structure that I'm like sort of tongue and cheek calling a SIF seed round.

Um, basically like Reed and Starting Line DC, who's another one of the investors, um, they committed up to two to two million, but we can pull it down when we need it. So, we have not pulled down all of it. Um, and I think that's a really nice structure for us because we're a media company.

Uh, we we also have a lot of software attached, but I want to maintain our ability to be weird and just like have a VO. It's creative playground where we make cool stuff. And uh so this gives us enough money to experiment but not so much that we're like locked into a particular growth path. Yeah.

Are you thinking that I mean like we saw this with OpenAI they were a nonprofit and then they wound up creating like the most dominant consumer app of all time. They needed to spin that out. Like the the parallel here is that you are building software tools but you also are doing media.

If one of those software tools takes off uh all of a sudden does the media become a marketing engine for that tool? Does the tool spin out? Are you thinking about it like incubation where there's a separate cap table, maybe investors ride along? Like what were those discussions like and how are you thinking about that?

Well, didn't you guys already spin out? Oh, yeah, we did. We spun out Lex um which my every co-founder Nathan now runs the CEO of that and uh we raised a seed round from True Ventures for Lex. Oh, cool. Um almost all the incubations that we have are actually their own separate LLC's.

I think like in general I [ __ ] love having a writing business. Um, and I want to keep the main thing the main thing. And I think that we can build a really incredible universe of uh of apps and other offerings around the writing business.

I do think sometimes um we will uh hopefully like run into opportunities that seem really massive and not going to be like just really big independent businesses on their own.

And yeah, we have um if that happens, we have the ability to go and actually like further spin them out, you know, turn them into C corps, have them go raise money, all that kind of stuff. Um, so we had the optionality to do both. Cool. Awesome. All right. Well, I mean, there's a bunch of stuff to talk about.

Why don't we Why don't we uh as our official uh enthropic uh correspondent uh which which is the title that we're giving you, obviously not not related to the company. I think we are going to have some folks from Anthropic on uh next week, which I'm excited about. Yeah.

So, the the the headline in Wired is Anthropic's new model excels at reasoning and planning and has Pokemon skills to prove it. uh Claude 4 Opus and Claude Sonnet 4 can remember over long periods of time. A capability that's helpful at Pokemon and other tasks that require an ability to stay on track.

So, uh what was most exciting? Where where should we start first with understanding the progress at Anthropic? So, first of all, you've got to read the every headline and the subhead. Um every time one of these models drops, we do a vibe check.

So basically we get access to these models hands-on um before they come out and then we use them hands-on for the daily tasks that we're doing to build software and and make writing and then we give you basically hands-on review of like how every part of the model works.

Um what I found is uh it's it's a piece of coding. It really runs autonomously for a long time on like complex PRs that were not like not possible with 37s on it and are probably a little bit beyond the reach of like Gemini 2. 5 and 03 in particular in cloud code. It's [ __ ] awesome.

Um, and so I do think that that's a it's a sort of game changer if you're if you're doing any kind of development work. um for writing and editing. I real quick uh go for it.

How do you what do you look at the what what are the kind of like market implications of clawed code in terms of the broader developer tooling uh ecosystem right you know we've seen you know cursors developing their own models windsurf joining openai there's a lot happening but you know do you think that clawed code is something that could have billions of of of arr independently or or or what's your your read on on how it's going to evolve?

That's a good question. I do think that the the landscape is starting to consolidate a little bit and with each with each model advance um the extra intelligence matters less and less.

So I think you'll see people um you know instead of jumping every time there's a new model released instead of jumping and and turning their entire stack from from an old model to the new one I think you'll see people start to stick a little bit more in the ecosystems that they've chosen.

Um it's we're entering a little bit of a different era of that of the race and I think it'll be a little bit more difficult for Anthropic to be like a totally independent lab and go up against not only OpenAI but also Google assuming Apple at some point gets their [ __ ] together, Microsoft, whatever.

Um but I I really think that they have a very very strong play with developers here. Um they've 3 35 Sonnet and 37 Sonnet are like superb coding models. 37 Sonnet is a little bit overeager. So it's um uh you know probably not as good as Gemini 2.

5 Pro, but it's still like really up there with the best coding models and um I think cloud code is is one of the best coding experiences like AI first coding experiences.

It's also just very different from like a cursor or a windsource where those are basically text editors with AI on the side and cloud code is a command line interface that's like you're just directly interfacing with the uh with the AI and that's it's it's just a different experience.

It's intended to be an assistant an agent rather than like a chat. Yeah. Uh do we have any idea of where these new models uh Claude 4 specifically sit relative to the OpenAI models?

is this because I remember 45 came out and they didn't give it the GPT5 name, but it did seem like it was trained on an order of magnitude more compute. And so this is kind of where the the pre-training wall discussion came from. But is it fair to think about Claude 4 as a GPT4.

5 class model in terms of kind of the scope and and the size of the model or do they give any do they give any indication as to the vectors that they're pulling on? We've seen you know Facebook with uh Meta Lama behemoth is going for this massive context window these trillions of token parameters.

Uh there's a whole bunch of uh there's a whole bunch of different threads that the different foundation model companies are pulling on. What seems to be their underlying motivation or or what do they seem focused on in terms of optimizing towards?

So it's a little hard to uh to tell at least for me because I didn't get access to any kind of model card before launch and I've honestly just been running around like just at the event and stuff. So if there are actual uh numbers and stuff like definitely go look them up. I I don't I don't know for sure.

On the blog post, they mostly just focused on the uh software engineering suite bench and the different uh the different benchmarks around graduate level reasoning, agentic tool use, multilingual Q&A, visual reasoning, and the stats are impressive, but not a huge jump here.

There's a jump on agentic coding, but Claude 3 Claude sonnet 3. 7 actually outperforms in visual reasoning by like a half a percent. not much.

But um it's one of these interesting things where it feels like we're increasingly in the era of tradeoffs around models and we might be seeing more fragmentation around a really great we talked to some founders who are building super specified LLMs just for JSON decoding or just for translation or just for profanity filtering and then they run they inference on a on a consumer grade GPU because it's so narrowly defined still using the transformer architecture but just much more narrow.

And so I'm wondering if if that's the future that we're going to see at Anthropic is they're they're they're dominating in code. They're very popular with developers and they're going to go after a few other areas, but they're probably going to stay out of a few other battle battlegrounds.

I do think that's an astute point. Sometimes you do see um tradeoffs and so like 37 Sonnet, for example, was a better programmer than 35 Sonnet, but it was like way more optimistic. Um interesting. And uh but but to to answer your your previous question like going back to the difference between QG4.

5 and Opus uh Opus 4 and Sonnet 4. Um they're a it's a little bit apples to oranges because 45 is just uh a base model. Uh I mean it's it's a it's instruction tuned but it's uh it's not a reasoning model. And um opus is uh is both.

they can go back and forth between um uh being a base model and uh and being a reasoning model. Um and that is like this sort of hybrid thing that I think all of the model providers are starting to move into. Um but I think Sonnet has Anthropic has done it first and um it's really good.

Uh it makes a big difference when um it both can like do the kind of you know chain of thought reasoning that's like good for map proofs and coding and it also has a little bit of a vibe. of you know like a GP45 which is just really good at writing and really good at like creative thinking. Cool.

How do you uh how do you kind of navigate and figure out what's real when everybody in AI is deeply conflicted in different ways whether they're sort of like secretly an adviser to this lab or you know happen to invest in the series B of this one or you know everybody's sort of talking their book.

I mean, obviously I just got advisory shares in OpenAI. I just got 1%. That's it. I'm not conflicted. Just like a point. Yeah, just you got a point. Careful, John. People are going to believe that.

Every time we joke, every time we joke, every time we joke around, every time we joke around, somebody takes it 100% as fact. Um, no, but I obviously the immediate uh, you know, way to figure out what's real and what's not is just to use the new products and models yourself.

But um I'm curious how you navigate it as somebody who's like trying to provide you know really really you know precise coverage of of everything in real time. So first of all if anyone wants to throw me some anthropic or open AI stock like I'm open. The answer are open.

Um but I think um yeah I think that's a really good question. Um I think I I honestly think most of the benchmarks are kind of [ __ ] and you can change the benchmarks and we see that with Llama where like it look good on the benchmarks but it's just not a good model. Um that's why we do vibe checks.

Um, so when we launch new models, we're using them hands-on for the tasks that we do every day.

And I think that's why it's really valuable to have um, every the media company and every the incubator startup studio in the same organization because I'm literally just going and hanging out with Kieran who runs Corora, which is our AI email assistant, or Danny, who runs Spyro, which is our AI content automation product.

Um, and we're using it to ship features or to make better writing or to do all the things that we do every day. And so I can get a pretty good sense of just like the vibe or the flavor of a model from using it myself.

And I think that's going to be the best um I think that's going to be the best way to tell if it's any good. What was your reaction to Google IO? What was the single thing that that stood out to you from the So I I was I was at Build.

I did a I had a I did a really fun interview with Microsoft CTO Kevin Scott and then flew right from Build to um to uh uh Code with Claude.

Alex Duffy uh is another one of our writers who needs AI training for us and he was at Google IO and so he would have a better response as to like what is the one thing but I do know he was like he was like kind of in tears in our discord and was like this is the future.

Um and I was and so you know I think that they've got something cool going on. I think he's really excited about they're pushing forward on a lot of different parts of the ecosystem all at once and we're moving super fast and um he seemed really psyched about it. Awesome.

What about the products that Anthropic hasn't launched?

We talked about specification uh the the specialization a little bit and I remember so uh the big the big AI product that really grabbed me from Google IO was V3 and obviously Google has an immense advantage with the YouTube data set there and the results have just been fantastic and I've been fighting tooth and nail to get more V3 credits because it's so much fun.

um and Anthropic, it feels like they've been potentially behind the ball on image generation. Is that just not a focus for them? Is that a do you think that's because of safety concerns uh since obviously that that runs very deep in the culture at Anthropic?

Or do you think it's just uh more of a business case that uh the the coding market is much more lucrative than the image generation market? And so just focus more wood behind fewer arrows. It's a good question.

I honestly don't know the history of Anthropic's relationship with coding model with image models other than that they just don't have one. Um and I don't know why. I do think it seems like they're really going hard after the kind of agentic coding market and I think it's a really valuable thing that they can.

Um I'll also say just generally like Mike Kger is their chief product officer and like he knows the ship. Um, and so I think they'll probably have some He knows the thing or two about images. Yeah. Yeah. Um, so I assume they'll have some really cool stuff on the more like consumer product side coming soon.

I think we're getting a Zoom air. I'm uh I'm searching 03 right now for anthropics offering. I don't think that they have any uh any models. I don't think they uh uh Dan down. Dan down. Bring it back. I think Do you think there's a widespread cyber attack? Probably because Axe is really struggling.

They know who we talked to earlier this morning. They know. They do. They know. They figured it out. We have a We have a secret guest that will be released later. Um, they're on to us. But 03 is cooking. I mean, it's I don't know if it's writing code yet, but it's definitely it's it's written. Look at this.

It's like I just asked it. I just asked it. Has Enthropic shelved any any models and it's like searching the web everywhere.

But it is it is a tougher it is a tougher question for 03 because I'm asking for the absence of evidence, not not I'm asking for it to find the absence of a model, not the not the presence of a model.

And so it's very easy to search does does Claude have a coding model and it can just search anthropic code and find the first result and return that to me. Um but it's very difficult to uh to to find a null and say okay I searched the entire web and I couldn't find anything.

Therefore anthropic does not have a video model. Um but yes uh I I I believe anthropic scrapped maybe it was a texttospech engine something like 11 Labs competitor which you got.

So, I've been I I went to try to ask Claude if if why doesn't Claude have an image generation model and the signup flow is couldn't be more opposite of OpenAI. I I've just been I I created a new account with my TBPN email and it took me there was I mean Nikita would probably have a heart attack.

I mean it was like 20 individual steps of approving this and and approving that. Uh but it says anthropic hasn't publicly released a dedicated image generation model, strategic focus, technical and safety considerations, resource allocation, market positioning. So kind of everything that Dan said.

Um just uh it's in, you know, it does make sense in in a way to just say, hey, this isn't strategically critical for us. We're going to try to, you know, focus on what we're really good at, which is coding. That's great. Um, well, we have TJ coming in for the second time. I'm very excited for that.

Yeah, did we fully lose Dan? I think we did. Let's just move on to some some some news. Let's talk about um the uh I mean there's plenty of timeline we can go. Dan, if you're listening, thank you for coming on. Enjoyed the conversation and uh always welcome to this is the future.

We have AI that can oneshot build you Zoom, buddy. But it won't work. But Zoom is still going to go down. internet will still fail constantly. Anyway, uh I do think it's fun that uh that he says his phone overheated. He was just going so hard on on on on you know corresponding. Wow. Yeah.

That's that's honestly as a technologist the only time you should put down your devices on a daily basis is if they're overheating. Like you need to be going that hard podcasting so hard. You need to be live streaming video, uh, doing spreadsheets, coding on your phone, doing everything.

Thank you for leaving it all on the field. On the on the field. You're an absolute dog. We will see you again soon.

Let's I think it's fun that they're doing uh that Anthropic is doing Pokemon because Pokemon's such this iconic game and it's so uh and it and it lends itself to like you can watch the full stream of Claude play Pokemon and it is this interesting challenge that it takes a long time.

They need to think about the different reward functions and all the different steps that you need to take. I was playing Pokemon on the Chromatic Palmer Lucky's new Game Boy.

And uh it's actually really hard because at a certain point you get to a place where you just kind of have to go farm XP and like level up your Pokemon evenly before you can go beat the boss. Like like you can't just you can't just constantly be going forward.

You actually have to have to plan out your your attack a little bit. It's it's not the simplest game. So impressive that they're able to play it for what 24 hours at a time. Um anyway, let's go through some timeline. We'll we'll get to Elon Musk later.

There's a there's a there's a piece we got to put in the truth truth zone, but it's going to take more than seven minutes. So let's uh talk about OpenAI. They committed to a giant UAE data center in global expansion. This is Stargate UAE. We'd love to see it.

Uh they're partnering with G42 to build a 1 gigawatt AI data center in Abu Dhabi. It's the first large scale project outside of the US. shake just him. He's on a tear. I cannot wait to have the opportunity to have I want him on the show. him on the show. I want him on the show so badly.

Uh 242 will fund the construction. OpenAI and Oracle will operate the data center. Soft Bank, Nvidia and Cisco are also partners in the project. The Kuretsu of AI is coming together. Um and so uh another 100 billion for Nvidia Jensen.

So uh the UAE is trying to become uh one of the biggest funders of AI companies and infrastructure and build their own AI factory. The first 200 megawatt chunk of the data center is due to be completed by the end of the thing that's interesting here.

They're having to like match the investment that they do American companies. It seems like a good deal. I'm I'm I'm very bullish on it. They're doing deals. That's great. Uh, Strava raised at a valuation north of two billion. Let's give it up. This is fascinating. I have to ask, what are they running from?

What are they running from? What are they running from? Maybe they're running from they they could very well be running from Zack Pogb, who is in some ways coming for them. They're running from not having generational wealth by getting liquidity at a multi-billion dollar valuation. So, congrats to the founders.

from Yeah, it took me a second. They raised some debt. Let's Let's give it up for leverage. They also acquired the Breakaway, a cycle cycling training app, marking its second acquisition in two months. They have more than 150 million users and it is approaching 500 million in annual recurring revenue. That's good.

Let's revenue. Let's give a moment of silent for the VCs that passed on the seed in the A because it seems, you know, yeah, like running. I I get a lot of people do it, but how big can this thing be? Easy critique. Easy. How big can this be? 500 million in annual recurring revenue. They showed you. Got them.

Turns out a lot of people are running from Strava lets users whom it calls athletes record in quotes. Wall Street Journal putting their term in the truth zone. Hey, runners are athletes in some context. Yes, they're running down a grid iron maybe. uh share activities with friends and across dozens of sports.

I've been, by the way, I've been avoiding we had Rob Mower, co-founder of Hubman Lab. It's just more more. I always He even texted me. I I botch his name every time. Terrible. Uh Rob uh I've been avoiding a run with Rob. Oh, yeah. For years now. For years. We live uh close by. He's a good runner. He's going to smoke it.

He's an absolutely insane athlete. Oh. And uh I just always end up busy. How's he uh how's he on the how's he on the bench press? Is he repping two plates? Because maybe you could give him a run for his money over there. Little big dog. Show him how it's done. Good. Potentially do that.

Huberman bench three plates and they're just going to I can see Rob just secretly repping three plates. Just demolishing us. Anyway, uh let's go through some timeline. Uh Jenny says glasses won't win. Earpiece with a tiny camera will. I don't know about that. What do you think?

I think I just would like a I would like a camerast I'd love to put a nostril camera in that just sort of mounts to my nose and just comes out right here. No one's talking about nasal computing. Nasal computing nasal computing interface could be the future for for guys like us that don't want to chip.

Yeah, the nasal passageway is a great place to put a computer. I mean, there's probably so many different applications. You could have little fans in there that accelerate the air flow in so you're getting more oxygen in your blood. Brian Johnson loves it. Hyperbaric chamber for your body. Yeah. Whoop on the wrist.

But we typically have watches on our wrist. Yes. Yes. And so Whoop on the nose. Whoop in the nose or a ring in the nose. Or what about an or I know where you're going with those. What are those? What are those? Those septum piercings. An aura. Septum piercing. Oh my god. Oh, that's insane. AI. AI generated right now.

The the nasal nasal aura ring. That's insane. Uh Anna says, uh, "We're in the or Anna says, "We're in the good old days in the beginning of the movie with the overly saturated colors and the smiles and laughter. Your future is being written by about six people competing to well their ideal civilization into existence.

" Record freeze frame. Record scratch. Hi, Anna. Here. You're probably wondering how I got here. Yeah, it does feel like it's a it's a monumental time. It's a it's an amazing time to be a technology podcast. Jensen Elon Satia Zuck. Yeah.

Uh Accelerate Harter says, "This blog post is styled like a cross between a wedding announcement and a memorial web page. It is a very funny photo. It is black and white and the font choice is is something that we haven't seen from OpenAI or Apple. " They're boys, though. They're just guys being dudes.