Anthropic (7 blogmarks)

Laurie Voss wrote No, Claude's watermark doesn't make the writing worse which is probably the most accessible starting point. If visuals help, Laurie pointed to this interactive watermarking post.

The idea and research behind the watermarking solution being proposed by Anthropic is SynthID Text which huggingface introduces here with code examples for getting started with it.

This post from Anthropic does a good job of explaining the whole thing in a nutshell in this paragraph. The core gist being that while probabilities or odds of what is generated are unchanged, this doesn't impact the model's weights, but rather "randomness decider" uses an internally-known key rather than an RNG.

Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it. When watermarking is used, choices are still made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick. That is, the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.

Using /clear and @-mentioning files are the essential things I do from this list -- I run multiple Claude Code sessions everyday while I work and I never hit usage limits.

The other thing that makes a big difference is planning out my work and handing off well-scoped, verifiable feature descriptions to CC. This allows it to do targeted work and quickly check that it met its objective without a bunch of churn and re-work.

Something else I do fairly often is ask CC to build a phased (sometimes "Red/Green TDD") plan of how to implement the feature and to /grill-me with any questions. This obviously has some up-front token cost, but I usually find that we address ambiguities, contradictions, and missing requirements that way. The phased plan, especially with the TDD, allows CC to proceed in a more methodical and direct way to the finished feature. I think that general leads to lower token usage overall.

Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout—for economies, public safety, and national security—could be severe. Project Glasswing is an urgent attempt to put these capabilities to work for defensive purposes.

Anthropic released the results of a recent study they conducted where they warn of the impact that LLM agent use for software development tasks can have on learning and skill development. This impacts everyone, but is crucial for earlier-career developers who haven't developed as many of these skills without LLMs in the picture.

None of this means we shouldn't be using LLMs for software development tasks, but rather we have to be intentional about how we use them.

Importantly, using AI assistance didn’t guarantee a lower score. How someone used AI influenced how much information they retained. The participants who showed stronger mastery used AI assistance not just to produce code but to build comprehension while doing so—whether by asking follow-up questions, requesting explanations, or posing conceptual questions while coding independently.

And here is another excerpt from the end of the article:

Our study can be viewed as a small piece of evidence toward the value of intentional skill development with AI tools. Cognitive effort—and even getting painfully stuck—is likely important for fostering mastery. This is also a lesson that applies to how individuals choose to work with AI, and which tools they use.

Around March 14th, 2025:

"I think we will be there in three to six months, where AI is writing 90% of the code. And then, in 12 months, we may be in a world where AI is writing essentially all of the code," Amodei said at a Council of Foreign Relations event on Monday.

Amodei said software developers would still have a role to play in the near term. This is because humans will have to feed the AI models with design features and conditions, he said.

Via this Armin Ronacher article, via this David Crespo Bluesky post.

I'm giving Claude Code a try. It is a terminal-based LLM agent that can iterate on software tasks with a human in the loop to prompt a task description, confirm or abort individual actions, and guide the process.

Claude Code does usage-based metering. As I'm currently looking at the "Buy Credits" page, there is an initial billing limit of $100 per month:

All new accounts have a monthly limit of $100 credits/month. This limit increases with usage over time.

After purchasing credits, I'm presented with a hero section large font that says:

Build something great

Once logged in to Claude Code in the terminal, I am first shown the following security notes:

Security notes:

 1. Claude Code is currently in research preview
    This beta version may have limitations or unexpected behaviors.
    Run /bug at any time to report issues.

 2. Claude can make mistakes
    You should always review Claude's responses, especially when
    running code.

 3. Due to prompt injection risks, only use it with code you trust
    For more details see:
    https://docs.anthropic.com/s/claude-code-security

An AI coding tool that I use directly from the terminal?! 👀

Claude Code is available as a limited research preview, and enables developers to delegate substantial engineering tasks to Claude directly from their terminal.

"thinking tokens"? Does that mean the input and output tokens that are used as part of intermediate, step-by-step "reasoning"?

In both standard and extended thinking modes, Claude 3.7 Sonnet has the same price as its predecessors: $3 per million input tokens and $15 per million output tokens—which includes thinking tokens.

The Claude Code Overview shows how to get started installing and using Claude Code in the terminal.