Claude Sci just dropped, and it's got me thinking about what Anthropic is actually becoming

So Claude Science is out now (beta, on Pro/Max/Team/Enterprise). Anthropic's own framing is that it's not a smarter model, it's the same Claude everyone already has, just wrapped in a workflow built for how researchers actually work.

And that got me thinking about the pattern. Anthropic literally just published an oral history of how Claude Code came to be, and the origin story is kind of wild if you haven't read it. It didn't start as a roadmap item. One engineer was messing around with the API, first version just told you what song was playing on your machine. He kept tinkering with it on weekends, gave it filesystem and bash access almost as an afterthought, and it spread through the engineering org before anyone officially decided to build a product. Cowork has a similar "grew out of internal use" flavor to it.

That's basically the Google playbook. Gmail was a 20%-time project. Google Maps grew out of an internal mapping tool two guys built after getting acquired. Google didn't set out to own your inbox, your maps, your search bar, your docs, it just kept turning "thing an engineer built to solve their own annoyance" into a product, over and over, until one day you realize you can't route around them anymore.

Feels like Anthropic is doing the same thing, except the internal tool this time is an AI that can just go build the next internal tool. Which is a genuinely different flywheel than Google ever had.

So what's cooking internally right now that we don't know about yet? No idea, honestly, but if the pattern holds it's probably not some flashy new model, it's some engineer's Tuesday-afternoon hack that a hundred people at Anthropic are already quietly depending on. Half of you have probably built your own version of it already with Claude Code and just don't realize Anthropic is about to ship a polished version of the exact thing.

On the "can AI labs become Google-tier giants" question, I genuinely think it comes down to compute, energy, and chips more than anything else. Google's moat was data and distribution. The AI labs' moat is whoever can secure enough power and silicon to keep training and serving at scale, and that's a much harder and more physical constraint than "index the web." Which is also why this stuff is genuinely contested, some people think a handful of labs end up as infrastructure the way Google did, others think compute gets commoditized faster than anyone expects and the moat evaporates.

And that ties into the "does work even matter" question. If inference keeps getting cheaper and agents start doing the actual doing, the honest answer is nobody really knows yet, not economists, not the labs themselves. Some people expect it to look like past automation waves, work shifts rather than disappears. Others think agentic AI is different in kind because it doesn't just remove one task, it removes the loop of tasks. I don't think anyone serious has a confident answer here, and I'd be skeptical of anyone who claims they do.

Curious what this sub thinks Anthropic's next "oh wait, that was just an internal tool" moment is going to be.