johnjwang

There exist methods to detect what kind of watermarking tool that someone is using, and most of the big tools have specific signatures that you can look for.

For Anthropic, it’s highly likely that the watermark is a SynthID type mark similar to the one that Sean is talking about (I actually ran the analysis here https://johnjwang.com/post/2026/08/12/how-claude-watermarkin...). When we get confirmation of whether all models actually are watermarked, I think we’ll be even more confident.

Of course it’s always possible that Anthropic has come up with a proprietary scheme, but I think it’s definitely harder to implement.

I think the game will be a cat and mouse game similar to LinkedIn and other websites trying to block scrapers: each iteration makes it harder for someone to figure out the watermarking scheme, but likely not impossible

I ran some tests from the watermarking literature: https://johnjwang.com/post/2026/08/12/how-claude-watermarkin...

Short version: it's likely a SynthID style statistical watermarking scheme that Anthropic is using. There's likely no hidden unicode and no large stable green/red scheme.

Longer version: I found some really statistically significant outcomes when using Gloaguen et al.’s (https://www.sri.inf.ethz.ch/blog/probingsynthid) detection algorithm for SynthID, though unfortunately it seems like the models might be pre-biased to pass that test already. The difficult thing is that it's hard to conclusively know which models definitely already have watermarking enabled and if there are any models that don't have watermarking, which means that being able to get a highly positive statistical result is much harder.

You can take a look at the full writeup, but many of the clues from Anthropic's blog post, plus negative results on many of the simpler tests, point to them using such a scheme. Of course, it's possible that they came up with a novel new scheme/algorithm, but amongst the available literature, the SynthID family of models is probably the best bet.

Fascinating! Hadn't heard of speech-core before, but will check it out if we are thinking about adding voice (though right now are focused on some more core-platform specific things like allowing for multiple changesets in a single session and adding memory across runs)

(Author here) I stand by this comment and I think it’s really important for engineers to recognize that everyone has different places where they gain and lose energy.

My team and I have been extremely lucky in hiring Joe, our excellent head of engineering, and an extremely strong set of engineering managers. Not to mention incredibly strong product and user experience management.

I think it’s pretty obvious that my approach wouldn’t work if I didn’t have this bench of talented managers, but because I do it affords me the luxury to spend time doing things that I love and which are also valuable to the company.

In general, I wrote this article because I think that the classical approach to engineering management isn’t the only path you need to take, and a lot depends on the team you work with (thankfully we have a team that complements each other really well).

(Author here): I hear what you’re saying, though I’ve never “crowed about regularly checking code in on Saturdays and Sundays” and I think that’s a false characterization of my article.

Do I love to code? For sure. Is it something I do on the weekends? Generally yes because it’s something incredibly fun for me, and it gives me a lot of energy. Now, is it an expectation I have of my team? No, it’s not because I want a sustainable pace for the team and I recognize not everyone has the same relationship with work or coding as I do.

And on the “circumventing process” bit — what I shared wasn’t an example of blowing past legal/security review recklessly. It was a case where I, as someone with full context, could quickly build something safe and unblock a customer, going through our normal code review and deploy process. I don’t expect anyone (myself included), to have any exceptions to this.

Very true, and I promise we didn’t solve P vs NP over lunch :). We just got much better at not exploring useless parts of the tree

From the API standpoint, it makes a lot of sense for us to be able to provide different types of providers. And we've found also that different models/providers are better at different types of tasks. For example, the Gemini models have really great latency, which are good for specific types of tasks that are very latency sensitive, but we've found reasoning to be quite strong with OpenAI/Anthropic.

To be clear, we still hire engineers who are early in their careers (and we've found them to be some of the best folks on our team).

All the same principles apply as before: smart, driven, high ownership engineers make a huge difference to a company's success, and I find that the trend is even stronger now than before because of all the tools that these early career engineers have access to. Many of the folks we've hired have been able to spin up on our codebase much faster than in the past.

We're mainly helping them develop taste for what good code / good practices look like.

In my experience, it still does take quite a bit of time (minutes) to run a task on these agentic LLMs (especially with the latest reasoning models), and in Cursor / Cline / other code editor versions of AI, it's enough time for you to get distracted, lose context, and start working on another task.

So the benefit is really that during this "down" time, you can do multiple useful things in parallel. Previously, our engineers were waiting on the Cursor agent to finish, but the parallelization means you're explicitly turning your brain off of one task and moving on to a different task.

Some engineers on my team at Assembled and I have been a part of the alpha test of Codex, and I'll say it's been quite impressive.

We’ve long used local agents like Cursor and Claude Code, so we didn’t expect too much. But Codex shines in a few areas:

Parallel task execution: You can batch dozens of small edits (refactors, tests, boilerplate) and run them concurrently without context juggling. It's super nice to run a bunch of tasks at the same time (something that's really hard to do in Cursor, Cline, etc.)

It kind of feels like a junior engineer on steroids, you just need to point it at a file or function, specify the change, and it scaffolds out most of a PR. You still need to do a lot of work to get it production ready, but it's as if you have an infinite number of junior engineers at your disposal now all working on different things.

Model quality is good, but hard to say it's that much better than other models. In side-by-side tests with Cursor + Gemini 2.5-pro, naming, style and logic are relatively indistinguishable, so quality meets our bar but doesn’t yet exceed it.