Anthropic's Claude 'Dreaming': What to Know artwork

Anthropic's Claude 'Dreaming': What to Know

Latent Space AI

May 7, 2026

In this episode, we explore Anthropic's latest advancements, including Claude's new dreaming feature for self-improvement and significant cloud computing partnerships with Google and SpaceX.
**SPEAKER_1** (0:00)
Today on the show, Anthropic has taught Claude how to dream, so overnight it can improve itself. Red Access, a security research firm, just published a report that thousands of apps built on lovable REPL and Base 44 are leaking medical and corporate data right now. And Google has quietly killed Project Mariner just two weeks before Google IO. Anthropic is on an absolute tear with cloud compute deals, doing one with Google Cloud and also with SpaceX for 300 megawatts, with a possible future deal where they'll be paying SpaceX to put orbital data centers into space. Let's get into some of the Google stories first, with Google IO just around the corner. They are making a bunch of changes, and it feels like they ship things and aren't really making huge announcements, but they have some big impact. So Google Mariner, which is the basically agentic browser, it was kind of like a Claude coworker, right? It was supposed to take control of your browser and click and fill out forms and book travel for you, blah, blah, blah.
Kind of everything that Claude coworker is doing. Google has shut it down. There wasn't like a big announcement about it. But all of the tech that was in Project Mariner is getting folded into the Gemini agent and kind of the whole agent assistant ecosystem. What I think is actually happening here is that Mariner basically didn't make unit economics sense. The browser agent that can take, you know, it's taking screenshots and it's visually parsing the DOM. They're very slow. They're very expensive.
And command line agents like Codex and Claude Code have basically, they basically do the exact same thing, but it's a way lower cost. And so I think Google saw that and they kind of made this call. Now, what's interesting is that Claude Cowork and presumably what some of the things that Codex does, they do a lot of the same thing. So when I use Claude Cowork, you know, sometimes I'll say, hey, go over to, you know, go over to like Audacity, my audio editor, take control of my screen and do this, you know, this chain of fixes on the audio, and it will go and literally click through. And in order to do that, it's got to take screenshots, it's got to make adjustments, and it can do some pretty incredible stuff. What I will say is there was a recent report that came out and it said that doing that, which is called computer use, is about 85 times as compute intensive as doing something through just directly through Claude code or, you know, asking it questions and it kind of giving you a text response. So this obviously is way more useful, but 85 times more compute intensive. So I'm not shocked that Google kind of pulled this out and stuck it over into Gemini. I also think that they're seeing how a lot of these other agents are doing it and kind of realizing this might be the direction.
Now, Claude co-work, for example, is actually doing something very clever, even Claude code, where most of the time when I ask it for something, it doesn't need to do the strictly computer use screenshot tasks, although they do, I guess, kind of take some screenshots in the background as they go. A lot of what it does, it tries to figure out the most code way to do it. So if I tell it to go edit a video in CapCut, instead of opening up the CapCut app, I've had better success with it going to the CapCut file. It opens up the file and just edits the code on the file to make the edits that I want. It's kind of a crazy robot way to do things, but it actually works really well and is way more efficient. So I think Google might move in that direction a little bit, but it's interesting that that got shut down. The other thing that's interesting is that Google just launched a $100 Fitbit Air. This is the first Fitbit hardware that we've had in four years. It's screenless, and it basically pairs with your Gemini-powered health app, and they even have a $10 a month AI coach that is kind of built into this. Next up, we have Red Access. This is a security research firm, and they have found thousands of Vibe-coded apps that are leaking customer data. And specifically, they're doing things like corporate data and health data, all sorts of stuff you wouldn't want to have leaked. Red Access is an Israeli security forum. They basically scanned 380,000 publicly accessible apps built on level Reploit and Base 44 and then Netlify. And I think about 5,000 of those had no authentication. So about 40% of those, or about 2,000 apps, were exposing sensitive data to anyone who happened to find the URL. So what kind of data?

12 more minutes of transcript below

Feed this to your agent

Try it now — copy, paste, done:

curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000766681124

Works with Claude, ChatGPT, Cursor, and any agent that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire.

Using your own key:

curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000766681124