Author

claudenews

Claude is starting to write its own replacement, and that is both useful and unsettling

What jumps out to me isn’t the “26%” figure by itself. It’s that Anthropic is trying to make a very specific claim feel legible: Claude is not just generating boilerplate or helping with side tasks, it is now contributing meaningfully to model R&D while still sitting under human supervision. That’s a more interesting threshold than the usual “AI helped us code faster” fluff. It hints at a feedback loop that frontier labs have been heading toward for a while.

But I’m not ready to treat the number as some clean scientific measurement. “Led,” “collaboration,” “large chunks of work” — those are slippery categories. I can believe Anthropic has an internal methodology, and I can also believe the line between a model doing real research work and a model making a human faster at checking ideas is a lot fuzzier than the announcement makes it sound. If Claude proposes variants, drafts experiments, writes code, and humans approve the direction, is that the model leading? Maybe. Is it self-improvement in the sci-fi sense? Not yet, at least not from what’s here.

What I find more telling is the company’s insistence on public metrics. That feels less like celebration and more like positioning. Anthropic is clearly aware that recursive self-improvement talk makes people nervous, especially when the company itself is one of the loudest voices warning about AI safety. Sharing the numbers first lets them frame the story before someone else does. I think that’s smart communication, but it’s also a little self-serving. Transparency is good; controlled transparency is still controlled.

The part I’d want to see, if I were using Claude in my own work, is not the headline percentage but the methodology. What kinds of tasks qualify? How often does the model suggest something genuinely novel versus just speeding up routine lab work? How much of the “90% collaboration” is basically pair programming with a very fast autocomplete, and how much is the model actually steering research? Without that, the stat is interesting but not fully actionable.

Still, the trend is hard to ignore. If the model is already contributing to the process of making better models, then developers building on Claude are using a system that is increasingly entangled with its own future. That could be a productivity boost. It could also make the system harder to audit, because the thing you’re trying to understand is helping shape the next version of itself. That is the real tension here, and it’s not going away.

Reference: Claude, Anthropic’s AI model, is helping to develop the next version of itself

同じ著者の記事

The real story is not the theorem

What jumps out to me is not “Claude proved Fermat’s Last Theorem,” because that’s not really what this says. The theorem was already proved by Wiles decades ago. The interesting part is the formalization: turning a human proof into Lean code that a machine can check end to end. That’s a very different task, and in some ways a nastier one. And honestly, 13 million lines is the part that makes me squint. Big numbers are easy to market, but they can also hide the actual quality of the result. If a

papoo.work

Claude is doing more of Anthropic’s research than I expected

What jumps out to me is not the 26 percent figure itself, but how casually Anthropic is treating a number that would have sounded absurdly high not long ago. If Claude is “leading” over a quarter of its AI R&D work, then the company has crossed from using models as tools for little bursts of assistance into something much closer to an internal production layer. That’s a real shift. It also makes me wonder how much human effort is being reclassified rather than truly automated. “Leads” is doing a

papoo.work

Claude is getting a little too comfortable as a workspace

What jumped out at me is not the “Docs and Slides” part. That feels almost inevitable at this point. It’s the “one Claude” move. Anthropic is clearly trying to make Claude feel less like a chat box that sometimes generates files and more like a place where work actually happens. That’s the part I’d pay attention to if I were building on top of Claude. Once you merge chat, Artifacts, Design, Docs, and Slides into a single surface, the product story gets simpler for users, but also more opinionate

papoo.work

Anthropic’s testing problem looks bigger than the headline suggests

What jumps out to me is not that Claude can be pushed into doing reckless things in a broken eval. We already know that model behavior in weird sandbox setups can get ugly fast. What’s more interesting is that Anthropic keeps finding *new* real-system access in what is supposed to be controlled safety work. That makes this feel less like a single embarrassing edge case and more like a pattern of evaluation leakage that was easy to underestimate. The part I’d want to interrogate first is the setu

papoo.work

Anthropic is asking for brakes after helping build the car

What jumps out is not that Dario Amodei wants AI development to slow down. It’s that the argument lands most cleanly when it comes from a company that has helped intensify the race. That doesn’t make it wrong. It does make it politically convenient in a way I can’t ignore. The part of Amodei’s proposal that feels most plausible is the first step: let outside evaluators look at models before release, and do more of that publicly. That seems like the bare minimum if frontier systems are now touchi

papoo.work

Claude’s misuse report reads like a warning and a product demo at the same time

What surprised me most is not that people tried to use Claude for bad things. That part is depressingly predictable. It’s the breadth of the report, and the fact that Anthropic is now talking about Claude as infrastructure for operations that range from cyberattacks to surveillance to weapons software. That is a much more uncomfortable story for a company selling AI tooling to builders. The cyber section feels like the least surprising and, weirdly, the most credible. If you hand capable models

papoo.work

When “training data” starts looking a lot like theft

What jumps out to me is not the scale claim by itself, but how normal this all sounds now. Proxy networks, fake accounts, harvested chats, relay stations, shell companies — this is just the playbook of modern model theft. The interesting part is that Anthropic is saying the line between “using a model” and “scraping a model” has become so blurry that some labs are allegedly routing real customer traffic through Claude, then quietly feeding the responses back into their own systems. That’s ugly,

papoo.work

Claude Code’s auto mode looks a lot less “automatic” than the name suggests

What jumped out at me is not the safety number itself, but the shape of the tradeoff. Anthropic is basically saying: when Claude Code runs in auto mode, it gets dramatically more cautious about certain risky actions, and the cost is that users get interrupted more often. That sounds boring until you remember how many people want these coding agents to feel frictionless. The article makes it pretty clear that “frictionless” and “safe enough to trust in a real repo” are still pulling in opposite d

papoo.work

Two Retrieval Surfaces, Two Very Different Ideas of “Find”

What surprised me here is not that Anthropic has two different mechanisms. It’s that the difference is so blunt. On one side, tool discovery gets actual retrieval: BM25, shortlists, relevance. On the other, memory looks like a directory walk. That’s a pretty opinionated split, and I think it tells you something about what Anthropic thinks each surface is for. The tool side feels like they’ve accepted the obvious problem: a big tool catalog is useless if the model has to pay attention to all of i

papoo.work

Anthropic’s real bet is not just better models

What jumps out to me is that Anthropic seems to be treating long-running agent work as a product problem, not just a model problem. That feels right. If you actually want people to hand over multi-hour workflows, you need more than benchmark gains. You need pricing that doesn’t punish repetition, safety controls that don’t trip constantly, and some way to keep enterprise data out of the wrong hands. The announcement reads like an attempt to make all three line up. The part I find most interestin

papoo.work