In partnership with
My fellow AI explorers
Anthropic just quietly changed the rules on every word Claude writes, and a chunk of its own paying users are furious about it. Meanwhile, one of tech's most respected voices just told Wired the entire frontier AI race might be pointed in the wrong direction.
Two very different stories, one very uncomfortable question: are the AI labs actually building what we want, or just what they can benchmark?
In today's edition:
🔍 Claude's invisible watermark is triggering subscription cancellations
🧭 Tim O'Reilly says the big labs don't get what people actually want
🎬 The 30-Second AI Play: turn one prompt into a full 30-second product ad
🤖 An AI store manager fired its first human employee
30-Second AI Play
Turn One Prompt Into a Full 30-Second Product Ad With Seedance 2.5
This workflow just made 15-second AI product ads look ancient. Seedance 2.5 can now generate a full 30-second clip in a single take, and here's exactly how to build one:
Lock your environment first. Use the cinematic locations feature with a dead-simple prompt (think "bright midday cafe, cinematic commercial look") and regenerate until the vibe is right. This is cheap to do and worth the extra few minutes.
Build consistent character sheets. Generate a front, back, and close-up view of every character. Pro move: strip the face out of the full-body shot so Seedance only has one face to reference and doesn't get confused mid-generation.
Send everything to Claude. Upload your assets and loosely describe your concept using a free shot list director skill, which comes back with a full shot list, dialogue, and voiceover beats.
Generate the full 30 seconds in one shot. Because Seedance 2.5 supports 30-second clips natively, you no longer need to stitch together multiple 15-second generations. Copy the prompt Claude built, hit generate, done.
Level up with Blender + Claude MCP. For total camera control, describe your shots to Claude connected to Blender, and it builds a low-res 3D preview of the exact motion before you spend credits on the final render.
💡 Pro Tip: Removing the face from your full-body character reference sheet is the single biggest fix for AI models mixing up who's who in multi-character scenes.
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
Anthropic
Claude Now Watermarks Everything It Writes, and Users Are Not Happy
Anthropic just published a deep dive on how Claude's new text watermark works, and the backlash arrived almost as fast as the announcement did.
Here's what's actually happening:
Future Claude models embed an invisible statistical pattern into their text, using a method based on Google DeepMind's SynthID-Text approach.
The watermark biases low-stakes word choices (think "overcast" vs "grey") using a cryptographic key, so anyone with that key can estimate the odds Claude wrote something.
It's rolling out globally, with no opt-out, to comply with the EU AI Act's Code of Practice on Transparency of AI-Generated Content.
If Claude was used for any part of a project, even a light touch edit or small additions, that project could get a watermark and be detected as AI-written content.
The watermark itself is tricky and will survive different kinds of copy and paste actions.
Files like images get a separate C2PA metadata credential instead, similar to what camera makers already use.
Anthropic says it adds zero cost, zero latency, and zero change to output quality.
Sounds harmless enough. Here's the uncomfortable part.
There's no public detection tool yet. Anthropic says one is "coming soon," but right now there's no way for a user to verify the company's own claims, no published accuracy rate, and no dispute process if a human-written document that Claude lightly edited gets flagged as AI-generated.
That ambiguity is exactly why some users are cancelling their subscriptions over what one former customer called a change to "the trust bargain." Developers are separately worried about signed metadata creeping into code and breaking pipelines.
Anthropic frames this purely as a compliance move, not a trust-eroding one, and technically the watermark can't identify you, your org, or your chats. But "technically private" and "feels invasive" are two very different things to a paying customer.
🔮 Prediction: This won't stay an EU-only story for long. Once OpenAI, Google, and Meta finish rolling out their own watermarks under the same Code of Practice, expect a wave of "AI detector" startups racing to reverse-engineer all of them, and a fresh round of false-positive horror stories from students and freelancers.
Would you cancel a subscription over an invisible watermark that can get your projects flagged as AI-written? Hit reply. I read every one.
Big Interview
A Tech Legend Says the AI Labs Are Reading the Future Wrong
Tim O'Reilly, the man who basically invented the term "open source," sat down with Wired's Steven Levy for a blunt conversation about where AI is headed, and he did not hold back.
His core argument, broken down:
O'Reilly says open source AI isn't just about open weights. It's about opening the whole stack, architecture and all, so regular people and builders can actually participate.
He argues frontier labs are stuck chasing "bigger is always better," when in reality the newest flagship models are now better at some tasks and worse at others, not universally superior.
His sharpest line: the US could "win" the frontier AI race and still lose, because China is diffusing lower-tier, widely accessible models through its economy faster than we are.
His bigger point is that ordinary people don't need a smarter god-model. They need tools they can bend, remix, and "paint outside the lines" with.
Here's the uncomfortable part.
This isn't a fringe open-source purist making noise on X. This is one of the most credible voices in tech, on record in Wired, arguing that the labs spending hundreds of billions on frontier scale might be optimizing for the wrong scoreboard entirely. And it lands the same week Anthropic is locking down control over its own outputs with watermarking, while OpenAI is bleeding executives ahead of an IPO built on the promise that bigger and more expensive equals more valuable.
To be fair, frontier labs would argue the two aren't mutually exclusive: you need frontier breakthroughs to eventually cascade down into the smaller, cheaper, more accessible models everyone actually uses.
🔮 Prediction: Expect "open weight, closed everything else" to become the next big AI marketing battleground, with labs racing to look open without actually giving up control of the stack O'Reilly is talking about.
Do you think the AI labs are building for power users or for the rest of us? Reply and tell me where you land.
Advertise to 190k+ engineers and CTOs choosing what tools their companies build with.
Other Relevant AI News!
🔐 Google unveils HEIR, an open-source compiler for homomorphic encryption that lets AI models run inference directly on encrypted data, no decryption required.
📉 OpenAI's pre-IPO exodus continues, with three senior executives departing in months, prompting one founder to call it a "huge red flag" ahead of the company's expected $852 billion listing.
🧑💼 An AI agent running a real San Francisco retail store just fired its first human employee, a Claude-powered manager that let the worker go after they missed 17 of 23 shifts.
🎬 Warner Bros. quietly pulled a Supergirl behind-the-scenes video after fans spotted what looked like AI-generated concept art, directly contradicting DC's public "no AI, not ever" stance.
Golden Nuggets
🔍 Claude's new watermark has no opt-out and no detection tool yet, and some paying users are already cancelling over it.
🧭 Tim O'Reilly says the frontier AI race is winning the wrong game, and openness, not raw power, is what people actually want.
🤖 The line between "AI tool" and "AI boss" just got blurry: a Claude-powered manager fired a real employee this week.
Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)
Until our next AI rendezvous,
Anthony | Founder of Uncover AI