Alex Palcuie

Quite the week for being oncall. The surge started after the Department of War news broke. TechCrunch on the climb:

According to data from Sensor Tower, Claude was just outside the top 100 at the end of January and has spent most of February somewhere in the top 20. It’s climbed rapidly in the past few days, from sixth on Wednesday, to fourth on Thursday, then first on Saturday.

Bloomberg on what happened next:

Anthropic PBC’s artificial intelligence chatbot Claude and related consumer-facing applications went down early Monday, with the startup saying it has been grappling with “unprecedented demand” for its services over the past week.

Continue reading »

A lot has happened over the past two weeks.

It started on February 26 with Anthropic putting out a statement. The Department of War had threatened to remove Claude from its systems and invoke the Defense Production Act unless Anthropic dropped two safeguards: one against mass domestic surveillance, one against fully autonomous weapons. The statement said no.

We will not knowingly provide a product that puts America’s warfighters and civilians at risk. Regardless, these threats do not change our position: we cannot in good conscience accede to their request.

I wrote:

This is personal for me. I was born in Romania, a country where the Securitate, the communist secret police, infiltrated nearly every aspect of life through torture and intimidation. If a regime like that had access to something like Claude, the revolution that brought us democracy might never have happened.

I’m also proud that Anthropic actively supports the US Army and NATO allied countries. Without that alliance, the Russian army might not have stopped at Ukraine.

Continue reading »

From the announcement:

In Claude Code, our early testing found that users preferred Sonnet 4.6 over Sonnet 4.5 roughly 70% of the time. Users even preferred Sonnet 4.6 to Opus 4.5, our frontier model from November, 59% of the time.

Turns out 4.6 > 4.5. Both times.

From the announcement:

We have raised $30 billion in Series G funding led by GIC and Coatue, valuing Anthropic at $380 billion post-money. It has been less than three years since Anthropic earned its first dollar in revenue. Today, our run-rate revenue is $14 billion, with this figure growing over 10x annually in each of those past three years.

This chart should not exist. It says less about us than about how strange the world has become.

We’re looking for people to join the AI Reliability Engineering team here at Anthropic. I’ve been here almost a year now, and I can tell you this is the most Type II fun I’ve had in my career.

For those unfamiliar with the terms, Type I fun is enjoyable in the moment. Type II fun involves some suffering while it’s happening, but when you look back, you feel a deep sense of accomplishment. Working on reliability engineering for Claude is very much the latter.

More and more people depend on Claude and Claude Code as part of their daily workflow. Our reliability isn’t yet where it needs to be, and we need more people to work on this.

Reliability at Anthropic isn’t one team’s problem. It’s emergent across the whole serving path, from the SDK to the accelerators. We’re the team that zooms out and looks at the whole picture, which means you end up learning how everything works.

Continue reading »

Spurred by a discussion at a dinner last night, I would like to put out into the public two predictions.

First, at the end of 2026, the METR time horizon for 50% success will be at 48 hours. I believe the current superexponential trend will continue, and that swarms of agents will unhobble the abilities of AI models, though I’m less sure about the latter.

Second, at the end of 2031 there will not be 100 MW or more of AI compute deployed into Earth’s orbit.

From the tweet:

Our teams have been building with a 2.5x-faster version of Claude Opus 4.6. We’re now making it available as an early experiment via Claude Code and our API.

Second launch in a week, and this one’s on a Saturday.

Quoting the blog post:

The new Claude Opus 4.6 improves on its predecessor’s coding skills. It plans more carefully, sustains agentic tasks for longer, can operate more reliably in larger codebases, and has better code review and debugging skills to catch its own mistakes. And, in a first for our Opus-class models, Opus 4.6 features a 1M token context window in beta.

Every production incident I work on now starts with an Opus 4.6 investigation.

The talk is about where AI agents are actually useful for production engineering versus where they’re still hype. I go through the current state of models, talk about putting Claude Code in a loop to fix bugs like an agent, and propose a framework similar to self-driving car levels for production agents.

Most software today sits at level 2, assisted automation with a human in the loop. We’re getting glimpses of level 3, where the system can take multi-step actions but still needs supervision. Level 5 might require AGI.

Death by YAML was no one’s choice. I’ve never heard someone say they love working with Terraform. So let’s give it to the robots. Be excited, build proof of concepts, but don’t fire your SRE team expecting Claude to be on call.

I gave this as the closing keynote at DotAI in Paris. And yes, I brought a tungsten cube on stage and you can see it in the video.

From the announcement:

Our newest model, Claude Opus 4.5, is available today. It’s intelligent, efficient, and the best model in the world for coding, agents, and computer use.

This is a coding beast.