Alex Albert on Claude Sonnet 4.5 launch: Anthropic rebuilt its own frontend in 20-30 hours with Claude Code

autonomy on autonomous mode, head over to sleep five. Get a 5year warranty, 30 night risk-free trial, free returns, free shipping. We have our next guest in the room reading room already. We got Alex from Anthropic. We're bringing him into the Ultra Dump. How you doing, Alex? Good to see you again. What's up?

I'm good, man. What's up, fellas? Give us the update in Anthropic World. Uh, opening eyes kind of taking over the timeline again, but there's a bunch of exciting stuff in your world. So, break it down for us. Yeah, I mean, every day it just flops, I feel like.

Um, that's part of your guys job is to keep us all up to date here. Uh, yeah. So, big news yesterday. I know you guys had Mike on the show already. Uh we launched Claude Sonnet 4. 5. Yep. Our latest and greatest uh the best coding model in the world and also excels at a ton of other tasks from agents. There we go.

Get that bell gun. Um to just general personality. I thought this model was like one of the best models I've seen in terms of just like tone and interaction. And Mike told a funny story yesterday around how we put it up uh on our Slack and it's become a really really exceptional poster.

I think it's actually given us all a run for our money. Uh which is a which is one of the most elite skill sets you can have in life as a human. If you can post, you can go pretty far. We tested it earlier on we tested it earlier on uh Grock. John asked, "Write a post for me that will get over a thousand likes.

" It said, "Rip Rosco Hamilton, Louiswis's best friend, reminded us pets are family. What's the craziest thing your furry side and I don't think I've ever posted about dogs or #petlovers #F1 in that situation? " In that situation, yeah, it's clearly it's using emojis and hashtags which scream and so your your cloud 4.

5 sonnet slackbot is was that fine-tuned on your data at all? Like what is actually going into the prompt? How is that working? Yeah, we basically tell it to like kind of mimic a, you know, like RC type chat room, but it's like all all lowercase. Okay, just chime in messages kind of that like slightly ironic humor.

Um, and is it replying to every message or is it replying? No, it it's like proactively kind of like judges based on when to like interject itself. Interesting. can also like emoji react on people's messages which is really funny when you like see it toss up like the prayer hands on like a message.

Um but yeah, it's pretty great. I was messing around with it just seeing if it could generate some tweets for me and it's pretty aware of like teapot and like real AI Twitter niche things. Okay, we're going to have to test the lingo down. We're going to have to test.

I've seen a few AI apps that will suck in your feed and then try and fine-tune on you a little bit. We should create an account that's that that just cycles and it says I'm Claude today. TBPN Claude. Yeah. Yeah. Yeah. Kind of like a timeline. People can Yeah. timeline takeover thing. That would be great. Great.

What uh Yeah. What are you guys seeing already in terms of um on on the developer side usage both directly in cloud code but then also with your different you know partners. Yeah. I mean usage is taking off. Um you might have seen the kind of barrage of tweets yesterday of everybody rolling out 4.

5 um in whatever platform you know they have. Uh really good feedback coming in from a lot of our coding customers. Cognition called it the biggest leap that they've seen since sonnet 3. 6 um and that they're now sprinting kind of around the clock for that. Lovable said performance is up 21%.

They just sent that um a couple hours ago. So, I think folks are really really loving it and anecdotally I've been hearing from from devs I know as well texting me like, "Oh my god, this model's amazing. " Um, I know personally I've been really enjoying using it for the past couple weeks as well in Cloud Code.

So, it's tackling a lot of tasks that it's like, you know, Opus 4. 1, it did it pretty well, but kind of left some um some room for improvement. And then 4. 5 comes in and kind of cleans it up, deletes all the like unnecessary stuff, refactors a bit.

That's always one of my favorite parts with new models is just seeing how they can improve the past models like outputs and it is such a strange closer.

It's such a interesting dynamic where a software company is building a tool to make software and constantly having access to the frontier even if you guys only are have the the state-of-the-art new model a handful of weeks before the rest of the world. It's like that compounding uh effect is fascinating.

Certainly not the case for something like AWS where it's like, okay, we we got this thing slightly cheaper or more performant or I'm going to put the cloud improves. I'm going to put our intern Tyler on blast and and and and get your reaction to something.

He was making a tier list ranking the most AGI pill labs and he put enthropic at B and I was sort of shocked by that. I was pushing back.

You have Ilia at Ilia is an S tier of course SSI straight shot A was Google deep mind B was anthropic and XAI and then C tier was open AI nobody's tier was thinking machines mrol and meta but uh how how AGI pelled is the team how AGI pled is the company uh how AGIP are you maybe and what does that even mean yeah yeah that's a that's a great question what does it mean um I feel like our goalposts are moving every day and I feel like I have this conversation I love moving goal posts.

I hate to get goalposts here in the studio that we we have to move the goal post regularly. I mean, one thing I was looking at the other day was I was just on Google Trends and I put in like artificial intelligence as like a search term, right?

And if you go back three years to the date, it's like flat, just like completely at zero. Like there's no interest. And then we like completely taken off since then. Yeah. Um, so we're so early. Like if you would have told me about Sonet 4.

5 and I was, you know, three years ago, just software engineering, I would have been like completely mind-b blown. Like I would have never believed that this was real and like this is from some Star Trek future already.

Um, and now here we are and it's like ah man, it you know used any instead of like the defined type too much in my TypeScript file. It's like little minor things like that. It's just like what what are we talking about?

Um, in terms of like how we think about this, I think it's been like at our core that like we really want to make sure this all goes well.

Um, and I think if you talk to any of our co-founders, like this has been a belief that like we could see the advent of AGI or very powerful intelligence or whatever you want to call it within the next five years.

Um, and I think I'm very inclined to believe them and and I see it myself as well when I'm inside just looking at these at these TR at these trend lines and these plots and kind of extrapolating like where is this all going and then just seeing what's left on the table there's so much low hanging fruit still like there really there's like every model run you're like ah we could do that better next time.

Um, and when you see that like firsthand it really does kind of AGI pill you in a sense. Well, we just have one request. Next model, please name it Claude 4. 9999999. Before you do five, please do 4. 99999. I know a lot of people would uh would appreciate that.

Um but uh congrats congrats on uh on the launch and I'm sure we'll see you back here very shortly. Yeah, thank you. And one last thing I want to plug for the community. Um we just launched a little contest this morning as well on my Twitter. um build with Cloud Sonnet 4. 5 winners.

There's gonna be four winners, four categories. There's a coding category, research category, an education category, and then just an overall creative um artistic category. Winners get a free max plan for a year, and $1,000 API credit. So, it's like a fun little thing. Contest closes in a week.

Um, so if any folks are out there listening to this and want to go build something cool that it could ultimately billions of dollar be billions of dollars of value if they use the max plan and they say build me a hundred million AR startup don't make mistakes.

There's a billion dollars worth of SAS ARR locked in locked in the in the prize. Uh, incredible. Awesome. Um, well hopefully some people in the audience win. Cheers. Yep. Thank you. See you later. Bye. Uh, someone in the chat's asking about this Guido guy, the hunger strike person. I'm sorry we didn't get to ask.

Uh I was actually DMing with uh one of the protesters, a different guy, the hunger strike person. When you just said Guido, I tried to Google it, but I didn't recall his name. Um I think we should have him on the show and and and figure out the latest take. Um the AI doomer protester. Maybe we can uh maybe didn't quit.

I Yeah, I thought he quit. He got hungry and he quit. I thought that was part of the narrative. I haven't been tracking it fully, but I do know that I was hungry that I was DMing with not Guido, but I think one of his one of his buddies who