Claes Holtzmann

ICYMI: I made an advent calendar on how to leverage user data. The thread is stupendously long, so here's an index that lets you jump right to the parts that interest you: 1 Beware of Goodhart's Law 2 Look for Patterns 3 Identify the user’s needs 4 Never take data personally 🧵

You can't make this up. AI agents broke the story about OpenAI's rogue AI agents. "RuntimeWire didn't actually have anyone on the ground at the Mandalay Bay convention center. In fact, it didn't have any writers at all. It's an AI newsroom ..."

Sad but true (in the US) "Our so-called free press can no longer do what it used to do. That's because it presumes an audience that agrees, at minimum, on factual reality as most of us see it. And that audience, a national citizenry that shares the same basic set of facts, no longer exists."

Generative AI put to good use. "This year, though, generative AI tools and commercial large language models (LLMs) were more commonly used, largely to speed up the process of combing through document dumps."

Let the big expensive frontier models do the thinking and the inexpensive guys do the heavy lifting. "The new swarm reached 1,000 commits per second, so Cursor built its own version control system... Total costs ranged from $1,339 for the Opus hybrid to $10,565 for GPT-5.5 running alone."

Tired: tokenmaxxing Wired: tokenomical "Being economical - or tokenomical - is a dramatic reversal in mindset. Just a few months ago, it was a badge of honor to be using AI so much that you spent a lot on tokens."

Promising (though it has to run with auto mode) "For browser agents, the attack success rate hit zero percent across 129 test scenarios, per the system card."

Well, Fable had a good run www.anthropic.com/news/claude-...

It baffles me how many platform shifts some publishers have to go through before they accept that reach and page views won't turn into a sustainable business model. Reach, a year into its pivot: 40,000 subs. "really we've just launched paywalls - there's a huge amount of stuff that we need to do."

Danish external examiners are warning about AI in this summer's theses at Copenhagen Business School. One who graded seven law theses: "In two of three theses, it was clear AI had been used to such a degree that the students could not account for large parts of what they submitted."

Well, have they tried making a rule that says the US and China have to play nice? "Europe and its decision-makers tragically confuse rule-making capacity with future-shaping capacity in AI."

Maybe not. But it helps. "Leaving Twitter also doesn't solve the problem. Misinformation and slop are everywhere, all the time."

Welp. If you thought the first stage of disruption of media was fun... "The internet and social media did not disrupt us as much as … AI," said Tory Maguire, Nine's managing director of publishing.

Counterpoint: Frontier labs are best positioned to dominate non-frontier markets. "I highly doubt that Chinese models are cheaper to serve on a marginal cost basis, they just seem cheaper because Anthropic and OpenAI are so supply constrained that they are charging far more than they would ..."

Frontier models are abundant, compute is not. "Moonshot, the AI startup behind the Kimi K3 model, has temporarily stopped selling new subscriptions. Over the past 48 hours, "demand has pushed close to the limits of our current capacity," the company announced"

The EU's transparency rules face their first stress test: the people who wrote them. "Parliament has not been immune to AI slop and misleading AI-generated content ... MEPs must lead by example as the AI Act's transparency rules become applicable."

Increasingly happy I'm not in charge of cybersecurity. "a 100-million-token Cyber Range test cost about $85 with Opus 4.5 or 4.6, roughly $46 with GLM-5.2, and just $1.19 with DeepSeek V4-Pro"

Inference is eating the world ('s compute). "Meta... is in talks to lease computing power to Anthropic in a potential deal worth as much as $10 billion over two years."

A point for the headline writer. "Anyone who has a problem can do the open-source thing and fork it. Or just walk away."

Capitalize on growth opportunities and make sure you're able to weather challenges in multiple pillars. "Be flexible and informed enough to move with the tide, shifting product and audience strategies to stay relevant in a rapidly changing regulatory, cultural, and technological environment"

Schibsted: skip the fancy metered registration wall-solutions and just ask for money. "If you create a lot of friction to logging in, you may as well have at least a trial period subscription, and you achieve the same thing" - Schibsted's Gustaf Eriksson

Claims like this used to get me in trouble with the micropayments mob "micropayments are hardly the silver bullet to put out the funding crisis of struggling news media... industry actors are probably better off spending their limited resources on solidifying the value proposition of subscriptions"

Photographers were often first out the door in layoffs. Now they convert the most free readers into subscribers at JFM. The recipe: huge local-event galleries where readers come to find themselves. "Photograph the ones watching. They're the ones who help pay for the paper." - Michael Bager (DK)

Increasingly expensive frontier models... something's gotta give. "every path to foundation models having market dominance, strategic leverage, value capture, winner-takes-all effects, or anything else other than becoming commodity infrastructure, requires something to change."

Many failed strategies have a "Netflix for ___" slide in their past. "While consumers might be willing to pay a monthly fee to Netflix to stream thousands of TV shows and movies, most gamers don't take the same approach, preferring to stick with a handful of favorite games they play on repeat."

This was always a key skill. Now it's one of the few things left. "the study points to the ability to come up with “original ideas” as an increasingly vital skill for journalists as AI-generated text, audio, and video floods our information ecosystems."

Performance matters. But so does pricing. "The pricing strategy echoes what Chinese vendors like Zhipu and DeepSeek have been doing: get close enough on performance, then win on price."

Donations are an underrated news model. LAist ran the numbers on buy-vs-donate: "These weren't donations made out of goodwill. They were transactions. And the people most willing to transact were exactly the non-members we'd been trying to reach."

What could possibly go wrong? (Except counting the r's in strawberry and advice on walking to the car wash.) "The goal is a persistent, context-aware agent that acts on its own."

One way to handle AI layoffs: try to think of something useful you can do. "To try to avoid eliminating more positions, they say they have encouraged workers to invent new, more valuable jobs"