The more apt word is steganography, rather than watermark. Basically subtly adjust the weights of the model so that some subtle patterns appear. Think of how AI text prefers certain words and phrases that ordinary humans don’t use as often, like “delve,” but presumably much more subtle.
And no, as Anthropic has already said, this watermark may not survive editing/formatting.
Claude Now Watermarks Your Text | Vanja Petreski - https://vanja.io/claude-invisible-watermark/
Chinese being more token efficient is a myth, and seems to stem from the superficial fact that characters are only visually more space efficient.
The fact that each Chinese character takes up 3 bytes (as opposed to 1 byte of English), words in Chinese typically require compounds of several characters, and that tokenizers have a limited vocabulary limited to mostly English means that Chinese is actually token inefficient.
No, Chinese Is Not More Token-Efficient Than English for LLMs | markhuang.ai - https://markhuang.ai/blog/chinese-token-myth
Nemotron Ultra is a open source and already gives DeepSeek R1 performance (admittedly not that good anymore). NVIDIA has open sourced the entire training process, including raw datasets and synthetic data generation. The raw data is mostly curated web crawl data.
Already happening for years.
Amazon mechanical turk, Outlier.ai, scale ai, are all gig platforms that pay people to produce real human content for LLM training.
At first they were paying people to actually produce content, like worked solutions to maths problems or translations, then they pivoted more to rating LLM output.
And as with every enshittification cycle, wages and work rapidly dried up, so people responded by asking LLMs for the answer just to meet deadlines and get by.
Instead of searching for existing gifs, Google will now pipe your search terms to Gemini so it will generate a gif for you.
Same energy
The industrial revolution began in Britain around 1760, but living standards for most people did not meaningfully improve until the late 19th century, they even fell in the first few decades.
That’s over an entire century, or at least four to five generations for meaningful improvement.
Sadly from personal experience, no, the general public does not know that coal power plants work by burning coal to boil water, and are horrified at how trashy the concept is.
High-key a great way to persuade them to support renewables. Associate coal with the image of people in underdeveloped countries burning fossil fuels in their own homes to exploit their own preconceived biases.
It was too good to be true. Microsoft had to implement additional 5 hour and week usage limits to cope with demand.
The problem was, for the entire history of copilot, there was literally no way to even check these usage limits. All usage counted towards these limits, even those with a 0x multiplier that didn’t consume premium requests.
It was so bad that many people couldn’t even use the premium requests they paid for the month. The only around it was to switch to auto model routing, which would tend to route to lower quality models.
Copilot was uniquely awful at this, because up to until literally days before the switch to usage based billing there was no way for people to track token usage, despite repeated calls from the community.
Microsoft only added a billing “projection” feature on the admin page that was meant to download a spreadsheet (which straight up didn’t work for most people) less than a week before the new billing structure.
Khan Academy has been on the enshitification journey for a while now.
Sal Khan in 2023 proudly proclaimed that AI could revolutionise EdTech with highly personalised tutoring, and made Khan Academy go all in on AI with their Khanmigo AI assistant tutor. Part of this ted talk explicitly has the AI pretend to be historical figures for assistance in history.
Three years later, Sal Khan admitted what all teachers already knew: that the AI tutor failed because most students didn’t even both with it. Digging through forums reveals the similar suite of reasons: factual inaccuracies, inability to understand basic arithmetic, and opaque pricing based on “usage.”
Didn’t stop him from partnering with big tech to offer $10K bachelor degrees though.
Sal Khan: How AI could save (not destroy) education | TED Talk - https://www.ted.com/talks/sal_khan_how_ai_could_save_not_destroy_education
Why Sal Khan is rethinking how AI will change schools - Chalkbeat - https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/
This CEO has teamed up with Google, Microsoft, and McKinsey to build an AI degree that could rival Harvard—and it will cost only $10,000 to attend | Fortune - https://fortune.com/2026/04/15/sal-khan-ceo-khan-academy-google-microsoft-ted-ets-higher-education-institute-bachelors-applied-ai-gen-z-college-upskilling/
Heartbreaking: man diagnosed with UnitedHealthcare
Ironically it’s because they paid very expensive artists.
The windows 8 logo was done by Paula Scher of Pentagram, one of the most exclusive and expensive graphic design firms.
Meeting with Microsoft early in the development process, Pentagram asked: “Your name is Windows. Why are you a flag?”
They really tried to hype an L as a W
Because people are too poor to purchase the extra output, assuming that the AI has actually lead to better productivity.
When there isn’t enough aggregate demand to absorb extra output from productivity gains, firms will opt to increase profits by reduce their costs by cutting back on spending instead.
Actually, it seems that even killing the hostage is fair game.
Microsoft Copilot Tells User Suicide Is an Option - https://futurism.com/the-byte/microsoft-copilot-user-suicide
Millennials caught the tail end of physical 3rd spaces, but experienced the golden age of digital third spaces. The internet until the mid 2000s was way more decentralised and chaotic.
Instead of anonymous chatrooms and forums as digital 3rd spaces, Zoomers used less accessible proprietary options such as Discord or Twitch instead.
Usual suspects
100k/year in San Francisco is definitely low income.
Official government figures using an exemplary family of four (two adults two children) already classified a household income of $117,400 as low income back in 2018. By 2025, that threshold has risen to $159,000.
San Francisco: Where a six-figure salary is ‘low income’ - https://www.bbc.com/news/world-us-canada-44725026
Places Where a Six-Figure Salary is Considered ‘Low Income’ - https://internewscast.com/news/places-where-a-six-figure-salary-is-considered-low-income/