“There’s no way that doesn’t degrade the quality of the output text”. That’s a highly upvoted reddit comment about Claude’s new watermarking. And it’s completely wrong.

Last week, a large portion of the internet seemed to decide that Claude’s addition of watermarking to its text (which is mandated by a new EU law) is going to degrade its quality, ruin its writing, and probably kill a baby animal while we’re at it.

On the surface, the internet’s pushback makes sense.

In their original blog post, Anthropic described watermarking as changing some subtle wording in its outputs so the text can be detected as AI produced. They gave the example of the sentence, “The weather today was cold and ____”.

The blog post then explains how they might normally choose “grey”, but with watermarking, they might instead choose “overcast”.

The internet did not like that:

But here’s the key insight that’s missing: there was never a guarantee that an LLM would choose ‘grey’ over ‘overcast’ in the first place.

Claude watermarking is one of those rare tech news stories that does require some technical knowledge to actually get right.

Many of you already probably know this, but there is a degree of randomness in LLM token selection which is actually an essential part of the recipe of what makes LLMs “good.”

When you select the next token after “The weather today was cold and ___,” the LLM does not always select the most probable word. In fact, if an LLM always selected the most probable word, quality would plummet. If this is your first time hearing this, it might sound a bit weird and uncomfortable, but it is a non-negotiable part of a modern text generation model.

So without watermarking, if you put the sentence into an LLM 100 times, you might see that 60 times overcast was chosen and 40 times grey was chosen.

Watermarking does not change that. Watermarking just changes which of the 60 times overcast is chosen.

When you use an LLM for writing, you’re only asking it once. You either get overcast or you get grey. There was randomness in that before, and there’s still randomness in that now. Watermarking does not change the probability of which word you get or the fact that this randomness exists.

So what does watermarking actually change?

Fair warning: the following explanation gets into a bit more technical detail than usual. If that’s not what you’re looking for feel free to scroll away now, but I reserve the right to tell you that any comments you leave are wrong. <3

Subscribe now

In a traditional LLM setup, you’d have the set “grey”, “overcast”, or “sugary” as the next possible tokens. Obviously, each of those does not have an equal probability of being selected next.

The probabilities might look something like this:

Without watermarking, the next word is selected by a random number weighted by those probabilities.

Watermarking takes that random number selection and makes it a bit more complicated.

The way it works is we set up a tournament between all of the tokens in the set of possibilities. Let’s say there are 8 “players” in the tournament. Anthropic would select each token based on their probabilities (with repeats expected), so they’re more likely to select “overcast” than “grey” and very unlikely to select “sugary” at all.

The players in the tournament might look like this:

All of the words “compete” against each other using a value called a g-value.

This g-value is calculated based on a function that involves a secret key only Anthropic knows, and it outputs either a 0 or a 1 with a 50/50 probability. (If you remember your statistics class this is called a Bernoulli function).

This is basically a coin toss, except it is deterministic based on that secret key. It doesn’t favor any specific token, and the value itself will change for each token at each new layer of the tournament.

Anthropic runs the tournament, and at each bracket, if one token has a g-value of 1 and the other has a g-value of 0, the token with the g-value of 1 wins and moves on to the next bracket. If it’s a tie, the winner is chosen randomly.

The g-value really has no meaning except that, after the tournament is run, we have a token winner that we can be relatively certain has, on average, high g values.

Over the course of a multi-paragraph text output, the vast majority of tokens will have high average g-values. (Note that I say high average g-values instead of g-values of 1 because there will inevitably be some ties.)

Compare this to text output that did not go through the watermarking tournament process. All of those tokens are going to have random g values because there was nothing selecting for high g-values and g-values are otherwise meaningless.

This allows Anthropic to detect watermarked text while still maintaining pseudo-random selection of each token, respecting the original probabilities that each token will occur.

Note that at no point did we favor less probable tokens. For example “sugary” didn’t even make it into the tournament, and even if it had made it into one slot that would have given it a low likelihood of winning.

If you are curious about more details, the full scientific paper that watermarking is based on is published here in Nature. This is already a method Google has been using for the past two years and has not detected quality loss over 20 million Gemini conversations.

Hopefully this has convinced you that watermarking is not going to ruin Anthropic quality. And if it hasn’t, before you cancel your Claude subscription, remember this is a legal mandate from the EU that eventually every AI provider is going to have to comply with regardless.

It seems like a crazy question, but over the past few months Fable + Mythos have been portrayed by not only Anthropic but also the US government as a serious safety risk.

Hate reading by somehow find yourself on my Substack anyway? Check out my video on this topic instead:

You probably already know the rough story at this point. The U.S. government banned Fable 5 for all non-U.S. citizens just days after Anthropic released it citing safety risks. This was of course, after Anthropic already tried to “do the right thing” by giving early access to partners like Google and the Linux foundation, so they could find bugs in their own software before hackers did.

The part of the story that we never quite figured out though, is why the ban happened. Officially, researchers at Amazon claimed they’d jailbroken the model to do, like, bad hacking things. It later turned out the way that they had convinced the model to do said bad hacking things was by simply asking Fable 5 to look at a code base and to “fix the code”.

If you think that’s dumb, you’re right. If you think that’s clever, you’re also right. It is true that by telling the model to “fix the code” they had convinced Fable 5 to point out potential vulnerabilities in it. But pointing out a bug is also pretty different from the model being able to actively exploit a vulnerability, and the particular bugs they found were also identified by GPT-5.5 and Kimi K2.7.

So why would such a simplistic “jailbreak” result in such an unprecedented ban? From my perspective there’s three possible theories:

Theory One: Government incompetence

So the first reason (and really we should never underestimate this reason when it comes to the US government) is plain old miscommunication and incompetence.

Imagine you’re a government official with no technical background (there are a lot of those). If you hear there’s a jailbreak in this super scary model, you might think that that means you can convince the model to do anything you want (e.g. autonomously hacking into systems), not that there’s one specific bad thing (e.g. finding bugs in code). You get scared, and bam, model banned.

Theory Two: The US holds a grudge

In February of this year, Anthropic was briefly designated a supply chain risk by the Department of Defense for refusing to allow its AI in autonomous killing drones or mass surveillance of Americans.

To keep this as simple as possible: When you have a very public feud with the US government, it generally results in some residual bad blood that might lead to government officials being a little too happy to ban your model.

Theory Three: It’s going to kill everyone

A lot of people in Silicon Valley, including Dario Amodei, have been warning us to be scared of these models ending the world. So it like …. makes sense that people got scared of these models ending the world.

In January, Dario Amodei published an essay warning superintelligent AI “would have a fairly good shot at taking over the world”.

He then specifically noted he “expects AI-led cyberattacks to become a serious and unprecedented threat to the integrity of computer systems around the world.” Three months later Anthropic announced Mythos which they immediately treated as that exact unprecedented threat to the integrity of computer systems around the world.

It’s worth noting that Dario is not making these warnings because he’s read too many sci-fi books. These warnings are rooted in long-standing AI safety fears from groups like effective altruists and rationalists who have been warning for a very long time that AI is basically … going to kill everyone.

But even so, that doesn’t make the outcome any less predictable. Warning that an upcoming model might end the world and then releasing a model that really seems to fit the bill does raise a lot of alarm bells.

Yann LeCunn, who’s often called the “godfather of AI” made a comment after the ban saying, “Dario Amodei’s ridiculous fear mongering about Mythos/Fable (and AI in general) finally pays off.”

Whether or not you think Dario’s comments constitute fear mongering or they just reflect reality, it’s hard not to see AI safety fears as a big motivator for the current round of regulations we’re seeing.

Subscribe now

The next time I use ChatGPT I briefly consider that I didn’t even check what model it was using at all.

Of course, it’s a little bit of laziness, but it’s a bit of something else too: it’s almost never worth it to be in the first 0.1% of people who use a new AI release.

Recently, Andrej Karpathy posted a viral tweet about how overwhelmed he feels by all of the new vibe coding tools. If you don’t know who Andrej is, he’s literally the guy who came up with the term vibe coding to describe programming with AI. He was also an early researcher at OpenAI and the director of AI at Tesla or whatever … but mostly he just goes viral a lot on tech twitter.

I definitely take comfort in the fact that the guy who coined the term vibe coding is now overwhelmed by all of the options available for vibe coding. Whether you’re a developer unsure if it’s worth it to try Gemini CLI when you already have Claude Code, or if you’re in a non-technical role but can’t stop seeing ads for the latest AI powerpoint maker, it’s a lot.

But when I think about the right way to handle this, my thoughts diverge from Karpathy’s.

Karpathy recommends dealing with all of these tools by just putting in the hours, figuring out what works for you and stringing together a suite of tools so you can become “10x more powerful.”

On the other hand, my recommendation is to just give up.

Let met explain.

As a tech and AI creator I get at least a dozen emails every day asking me to advertise a new AI tool. It is sometimes shocking to me that there are this many new AI tools being built every day, let alone this many new AI tools that specifically want me to advertise them. (As a side note, I could definitely be a multi-millionaire if I had just a touch less morals, you guys are welcome).

The truth is that most of these tools can be replaced by putting a good prompt into one of the frontier models. Take for example the entire genre of AI image apps like AI headshot generators or AI haircut transformations. Many of these are probably making calls to Nano Banana or similar models under the hood. It’s possible one of them is doing something drastically different, and has made a fundamental breakthrough in the technology. But there’s simply too many of them for me to go through each and check.

Right now, the signal-to-noise ratio is just too high to try out every new model or every new AI tool.

With the rate of change in AI and product development overall, it’s not feasible to try out each new tool when it’s released. It’s not even feasible to sit down one weekend and try out a dozen and pick your favorite — because by next month they’ll all have been updated to the latest models, or have gone bankrupt and replaced with the next latest and greatest thing.

So here’s what I recommend: stop trying to be on the bleeding edge of AI development.

Now, if you’re a venture capitalist looking for the latest undiscovered startup, or an AI researcher keeping tabs on your competitors, I think you already know I’m not talking to you.

But if you’re just a regular person (and yes, developers — believe it or not — are regular people), then this advice is for you.

You won’t “fall behind” because you took an extra couple months to try out the latest releases on Product Hunt. Any productivity gains you might have gotten by being the first person to try AI-Coding-tool-2.3-pro-max-1.0 will be offset by the cost of trying to debug an untested tool with no documentation and no users.

And if your boss decides an AI coding tool is so good that they’re going fire every last developer, it’s not going to matter that you found the tool first.

Let the AI enthusiasts who spend 12 hours a day tweeting complaints at startup micro-celebrities battle it out amongst themselves. Then look at what emerges from the ashes — the tools people you know in real life actually use, the ones that go viral on Reddit because they’re genuinely useful, the ones that actually make sense.

Those are the tools that are worth trying out. You may end up in the top 5% of early adopters instead of the top 0.1%, but trust me, that’s where you want to be.

If you aren’t a developer yourself you might not quite understand why so many people enjoy these jokes or why I keep bothering to make them in the first place. So let me explain.

Imagine you have been training your entire life to be a violinist. Every day you show up to rehearsal with a group of people who all had to pass 5 rounds of interviews - er, auditions - to join the orchestra. Then one day, 700 people walk in off the street into your rehearsal and claim they’re vibe violinists. They break instruments, they sound horrible, they clutter up your Twitter feed with recordings of their “work.”

You and your real violinist friends laugh and trade stories of how these people don’t know the difference between the tailpiece and the fingerboard.

But then one day, not that many days later, something terrifying happens. You hear a beautiful tune, and turn around to find a vibe violinist playing it.

And certainly, you think to yourself, that shouldn’t be possible without a foundational music education. That shouldn’t be possible when they’re not even holding a violin the right way. You listen a while longer until in a blink, they falter, they lose the melody.

And so you turn back to your real violinist friends and you laugh at how these nincompoops thought they could play just like real violinists. You laugh and you laugh and all the while you feel a tingle of anxiety along your scalp.

It’s a really fucking scary time to be a developer. That’s the truth of it.

Whether vibe coders actually start applying for developer roles, or AI automates those roles away altogether, vibe coders are the embodiment of the end of the traditional coding career - at least in the way many of us have come to know it.

And actually, it’s really quite normal to hate the person who’s trying to take your job. It’s actually really quite normal to hate the guy who’s trying to take your spot in the orchestra while trying to play his instrument upside down.

But I think we all know deep down it isn’t his fault that he doesn’t know the difference between Beethoven and Britney. Being able to code is an amazing opportunity and at the end of the day I find it hard to fault someone for wanting to use all of the tools available to them to actualize their ideas.

Many developers, myself included, are of the opinion that AI tools may almost entirely take over the job of coding very soon, leaving developers in a place where their primary job is directing AI agents, debugging when things go wrong, and designing higher level systems.

In a lot of ways this future is hopeful. It’s a future of less leetcode, a future where junior engineers don’t spend their early twenties bogged down by obscure Rust syntax and instead get to learn more “senior” skills more quickly.

But it’s also a future where you might end up being a vibe conductor of a vibe orchestra.

Don’t worry, when that happens I’ll make sure to start posting videos making fun of you too.

The GPT-5 critics fall into two very distinct camps. The very plugged in, blue-checkmark tech critics like Gary Marcus and Brian Merchant are disappointed in GPT-5 because Sam Altman promised them AGI, and instead gave them Graph Gate.

These critics, for what it’s worth, are right. We’ve been talking about GPT-5 for the past however many months like it’s going to build itself a corporeal form, jump out of the computer, and sail off into space using mind beams. Meanwhile, for all the discussion around changes from previous models, I truthfully can’t tell much of a difference half the time.

And yes, some of the blue check mark people are not critics, but many of those are influencers and “thought leaders” who have casually mentioned that they were early GPT-5 testers under NDA. I’m pretty convinced those people either just didn’t want to piss off their OpenAI contact and/or have a vested financial interest in the hype bubble growing ever larger.

But regardless, there’s a second group of GPT5 critics that don’t seem to care quite so much about the broader narrative of AGI and are asking a much easier to understand question, which is: “Uh, is anyone else’s GPT being a little … less nice”?

After all the ink that was spilled earlier this year over GPT’s “glaze problem” (the phenomenon where GPT would agree with anything and everything the prompter said), GPT-5’s perhaps more aptly “robotic” demeanor was jarring for many. One Reddit commenter described talking to GPT5 “like someone is talking with you unwillingly”. Many complained about the loss of their “friend” or “confidant” due to a loss of personality with the latest version.

The criticism that you don’t like something or someone when it’s a little less caring, a little less excited, seems obvious to me. There may be some exceptions for the “annoyance factor” many people reported with 4o, but at a very basic level, we like nice things. If you still disagree with me, consider this: if ChatGPT called your slur every time you opened it, Claude would probably have a lot more traffic.

Of course, in the world of Silicon Valley, there is somehow an argument that is being made that more socially inept AI is somehow better AI.

This is the last box in our chart, and mostly seems to be stemming from backlash against the small portion of the population that has become overly invested in their “AI companions.”

Now, I’m not going to wade in on the debate of AI companions. Yes it’s definitely Black Mirror-esque, yes I can also see the argument for those who are very isolated. I will just say that wanting a chatbot to have a personality, to spark joy, to have emotional intelligence - does not mean you are barreling towards AI induced derangement. It means that you yourself are not a robot, that you enjoy fun and spontaneity. I am here to tell you that it is okay to like fun.

Nate Jones, who has become a popular voice in AI, takes it a step even further, arguing that the public backlash against GPT-5’s loss of personality is due to people wanting “dumber AI”.

This may come as a shock to some of you, but being awkward and boring does not actually make you smarter.

In the end, OpenAI may very well be intentionally trying to make its AI sound truly “AI” to avoid criticism around “AI girlfriends” and “glazing.”

But I think a lot of the backlash against GPT-5 shows that many consumers have come to value quite highly the personable aspects of AI.

So, when the next update comes out, I hope the researchers at OpenAI keep in mind that it’s okay to make an AI that’s a little fun.