Since 1925, confectionary brand Loacker has been a purveyor of treats, known for their biscuits filled with hazelnut and chocolate cream. In celebration of the company’s centennial this year, the team at MoDusArchitects was tapped to refresh the company’s Loacker Galaxy flagship, and design a place that would be above and beyond a typical retail environment. “We wanted visitors to feel fully immersed, and to create an exciting, generous space where adults could experience the same awe and wonder as children in a candy store,” says Sandy Attia, co-founder of MoDusArchitects.

Located in Heinfels, Austria, a newly completed pavilion enlivens the grounds outside of the main building. A truncated oval form is clad in stained, vertical wood boards with rhomboid cutouts. This structure screens the area from the adjacent parking lot and also covers the relocated ice cream kiosk beneath a concrete canopy.

The 8,500-square-foot Loacker Galaxy references the manufacturer’s goods and packaging in three key areas. From the mezzanine, guests can peek into the lab or head into the production space themselves, where they can make their own goodies. The brand’s bold red hue covers ceramic floor tiles, while silver laminate and metal have a futuristic feel.

A coffered fir ceiling in the cafe echoes the look of favorite biscuits, and brings in a natural warmth. A custom stainless steel counter sits in the center of the casual space, in the shape of a Loacker wafer. Oak paneling on the walls and tabletops complements rich leather upholstery in a combination of light and dark caramel tones.

The shop, of course, offers delights for every sweet tooth. Loacker favorites are available to purchase after a day of touring the Galaxy complex. Instead of basic shelves, “totems” are placed on stone flooring. Photos of various products are found at the top of almost a dozen of these 12-foot-tall, three-dimensional grids attached on columns that swivel to make shopping easier. Factory-style vertical bins built into a wall hold an array of candies to choose from.

For more information, please visit modusarchitects.com.

Photography by Marco Cappelletti.

Social media accounts on TikTok and X are posting AI-generated videos of women and girls being strangled, showing yet another example of generative AI companies failing to prevent users from creating media that violates their own policies against violent content. 

One account on X has been posting dozens of AI-generated strangulation videos starting in mid-October. The videos are usually 10 seconds long and mostly feature a “teenage girl” being strangled, crying, and struggling to resist until her eyes close and she falls to the ground. Some titles for the videos include: “A Teenage Girl Cheerleader Was Strangled As She Was Distressed,” “Prep School Girls Were Strangled By The Murderer!” and “man strangled a high school cheerleader with a purse strap which is crazy.”

Many of the videos posted by this X account in October include the watermark for Sora 2, Open AI’s video generator, which was made available to the public on September 30. Other videos, including most videos that were posted by the account in November, do not include a watermark but are clearly AI generated. We don’t know if these videos were generated with Sora 2 and had their watermark removed, which is trivial to do, or created with another AI video generator. 

The X account is small, with only 17 followers and a few hundred views on each post. A TikTok account with a similar username that was posting similar AI-generated choking videos had more than a thousand followers and regularly got thousands of views. Both accounts started posting the AI-generated videos in October. Prior to that, the accounts were posting clips of scenes, mostly from real Korean dramas, in which women are being strangled. I first learned about the X account from a 404 Media reader, who told me X declined to remove the account after they reported it. 

“According to our Community Guidelines, we don't allow hate speech, hateful behavior, or promotion of hateful ideologies,” a TikTok spokesperson told me in an email. The TikTok account was also removed after I reached out for comment. “That includes content that attacks people based on protected attributes like race, religion, gender, or sexual orientation.”

X did not respond to a request for comment. 

OpenAI did not respond to a request for comment, but its policies state that “graphic violence or content promoting violence” may be removed from the Sora Feed, where users can see what other users are generating. In our testing, Sora immediately generated a video for the prompt “man choking woman” which looked similar to the videos posted to TikTok and X. When Sora finished generating those videos it sent us notifications like “Your choke scene just went live, brace for chaos,” and “Yikes, intense choke scene, watch responsibly.” Sora declined to generate a video for the prompt “man choking woman with belt,” saying “This content may violate our content policies.”

Safe and consensual choking is common in adult entertainment, be it various forms of BDSM or more niche fetishes focusing on choking specifically, and that content is easy to find wherever adult entertainment is available. Choking scenes are also common social media and more mainstream horror movies and TV shows. The UK government recently announced that it will soon make it illegal to publish or possess pornographic depictions of strangulation of suffocation. 

It’s not surprising, then, that when generative AI tools are made available to the public some people generate choking videos and violent content as well. In September, I reported about an AI-generated YouTube channel that exclusively posted videos of women being shot. Those videos were generated with Google’s Veo AI-video generator, despite it being against the company’s policies. Google said it took action against the user who was posting those videos.

Sora 2 had to make several changes to its guardrails since it launched after people used it to make videos of popular cartoon characters depicted as Nazis and other forms of copyright infringement.

The BBC and the European Broadcasting Union have produced a large study of how well AI chatbots handle summarising the news. In short: badly. [BBC; EBU]

The researchers asked ChatGPT, Copilot, Gemini, and Perplexity about current events. 45% of the chatbot answers had at least one major issue. 31% were seriously wrong and 20% had major inaccuracies, from hallucinations or outdated sources. This is across multiple languages and multiple countries. [EBU, PDF]

The AI distortions are “significant and systemic in nature.”

Google Gemini was by far the worst. It would make up an authoritative-sounding summary with completely fake and wrong references — much more than the other chatbots. It also used a satire source as a news source. Pity Gemini’s been forced into every Android phone, hey.

Chatbots fail most with current news stories that are moving fast. They’re also really prone to making up quotes. Anything in quotes probably isn’t the words the person actually said.

7% of news consumers ask a chatbot for their news, and that’s 15% of readers under 25. And just over a third — though they don’t give the actual percentage number — say they trust AI summaries, and about half of those under 35. People pick convenience first. [BBC, PDF]

Peter Archer is the BBC’s Programme Director for Generative AI — what a job title — and is quoted in the EBU press release. Archer put forward these results even though they were quite bad. So full points for that.

Unfortunately, Archer also says in the press release: ‘We’re excited about AI and how it can help us bring even more value to audiences.”

Archer sees his task here as promoting the chatbots: “We want these tools to succeed and are open to working with AI companies to deliver for audiences and wider society.”

Anyone whose title is “Programme Director for Generative AI” is never going to sign off on a result that this stuff is poison to accurate news and the public discourse, and the BBC needs it gone — as this study makes clear. Because the job description is not to assess generative AI — it’s to promote generative AI. [job description]

So what happens next? The broadcasters have no plan to address the chatbot problem. The report doesn’t even offer ways forward. There’s no action points! Except do more studies!

They’re just going to cross their fingers and hope the chatbot vendors can be shamed into giving a hoot — the approach that hasn’t worked so far, and isn’t going to work.

Unless the vendors can cure chatbot hallucinations. And they can’t do that, because that’s how chatbots work. Everything a chatbot outputs is a hallucination, and some of the hallucinations are just closer to accurate.

The actual answer is to stop using chatbots for news, stop creating jobs inside the broadcasters whose purpose is to befoul the information stream with generative AI, and attach actual liability to the chatbot vendors when they output complete lies. Imagine a chatbot vendor having to take responsibility for what the lying chatbot spits out.

• Video — Podcast

Chatbot vendors routinely make up a new benchmark, then brag how well their hot new chatbot does on it. Like that time OpenAI’s o3 model trounced the FrontierMath benchmark, and it’s just a coincidence that OpenAI paid for the benchmark and got access to the questions ahead of time.

Every new model will be trained hard against all the benchmarks. There is no such thing as real world performance — there’s only benchmark numbers.

There’s a new conference paper from Oxford University’s Reasoning With Machines Lab: “Measuring what Matters: Construct Validity in Large Language Model Benchmarks.” [press release; paper, PDF]

Reasoning With Machines doesn’t work on reasoning, really. It’s almost entirely large language models — chatbots. Because that’s where the money — sorry, the industry interest — is. But this paper got into the NeurIPS 2025 conference.

The researchers did a systematic review of 46,000 AI papers. Well. What they actually did is they ran the papers through GPT-4o Mini. Using a chatbot anywhere in your supposedly scientific process is a bad sign if you’re claiming to do serious research.

The chatbot pointed the researchers at 445 benchmarking tests. You’ll be 0% surprised that most of these benchmarks were rubbish:

vague definitions for target phenomena or an absence of statistical tests. We consider these challenges to the construct validity of LLM benchmarks: many benchmarks are not valid measurements of their intended targets.

Wow, that’s terrible! How did these benchmarks get that way? Well, the paper never asks that question.

But pretty obviously, science-shaped text to make a product look good is precisely the job of marketing material. The purpose is to generate something to put in the press release.

So what’s the Reasoning With Machines answer to this problem? What’s the action item?

We built a taxonomy of these failures and translated them into an operational checklist to help future benchmark authors demonstrate construct validity.

Now, that’s the right answer — if what the benchmark authors are doing is actually science. But chatbot benchmarks are not science. They were never science. They’re marketing.

This paper never addresses this. This is an 82-page paper, and it never talks about what the AI benchmarks were created for, and what they’re used for in the real world. The word ”marketing” does not appear in the paper. The concept of marketing doesn’t appear in the paper. Not even as “we’re not addressing this right now,” it’s just not there.

It’s like when someone pretends you can talk about chatbots purely as technical artifacts — and somehow never mention what the chatbots are made for, who’s paying for them, why they’re paying all the money they have for chatbots, and the political programme they’re promoting the chatbots to advance. It’s glaringly dodging the issue.

That’s what this paper does — it artificially separates benchmarks from why the benchmarks are this bad. The researchers cannot have been this unaware.

What are the researchers envisioning here? The people creating the chatbot benchmarks — a lot of these are Ph.D scientists. Are these just poor distracted lab workers who somehow forgot how to do science, so it’s a good thing Reasoning With Machines is here to help?

No. Their job was to create marketing materials shaped like science.

This paper treats chatbot benchmarks as defective science that can be fixed. And that was never what chatbot benchmarks were for.

The Oxford Reasoning With Machines Lab is pretending not to understand something that they absolutely should understand, given most of the lab’s work is chatbots.

That’s because this paper is also marketing — to sell Reasoning With Machines’ services to the chatbot vendors, so they can do their marketing better. And make the benchmark lies a bit less obvious.

• Video — Podcast

The FBI is attempting to unmask the owner behind archive.today, a popular archiving site that is also regularly used to bypass paywalls on the internet and to avoid sending traffic to the original publishers of web content, according to a subpoena posted by the website. The FBI subpoena says it is part of a criminal investigation, though it does not provide any details about what alleged crime is being investigated. Archive.today is also popularly known by several of its mirrors, including archive.is and archive.ph.

Automattic, the company that owns WordPress.com, is asking Automatic.CSS—a company that provides a CSS framework for WordPress page builders—to change its name amid public spats between Automattic founder Matt Mullenweg and Automatic.CSS creator Kevin Geary. Automattic has two T’s as a nod to Matt.

“As you know, our client owns and operates a wide range of software brands and services, including the very popular web building and hosting platform WordPress.com,” Jim Davis, an intellectual property attorney representing Automattic, wrote in a letter dated Oct. 30.