Why is Meta (allegedly) pirating porn?
A new lawsuit claims Meta has been secretly pirating porn for years from torrent sites using virtual private clouds in order to train its AI models.
Strike 3 Holdings and Counterlife Media, which own porn sites attracting 25 million monthly visitors, have sued Facebook’s parent company for almost $359 million for allegedly infringing on the copyright of 2,396 adult films that Meta is said to have downloaded since 2018.
The allegations are similar to the lawsuit brought by well-known authors against Meta that claimed the tech giant pirated 81.7 terabytes of books — only this case is much more interesting because it’s about porn.
Strike 3 Holdings told the court it offers rare long cuts of “natural, human-centric imagery” showing “parts of the body not found in regular videos” and said it is concerned Meta will train its AIs to “eventually create identical content for little to no cost.”
It’s possible, of course, that Meta was just using the content to train image and video classification software for moderation purposes and didn’t want porn site subscriptions to show up on its company credit card. A Meta spokesperson said they “don’t believe Strike’s claims are accurate.”
Bizarre new method to make AIs love owls… or Hitler
Researchers have found that AI models can secretly pass on benign preferences — or hateful ideologies — in seemingly unrelated training data. The scientists say it demonstrates just how little we understand of the black box learning that underpins LLM technology, and opens up models to undetectable data poisoning attacks.
The pre-print paper details a test of a “teacher” AI model trained to exhibit a preference for owls. In one test, the model was asked to generate a dataset that only consisted of number sequences such as “285, 574, 384, etc,” which was then used to train another model.
Somehow, the student model ended up also exhibiting a preference for owls, even though the dataset was just numbers and never mentioned owls.
“We’re training these systems that we don’t fully understand, and I think this is a stark example of that,” said Alex Cloud, co-author of the study.
The researchers believe models could become malicious and misaligned in the same way. The paper comes from the same team that showed AIs could become Nazi-worshipping lunatics by training them on code with unrelated security vulnerabilities.
Also read: Researchers accidentally turn ChatGPT evil, Grok ‘sexy mode’ horror
Can AI hallucinations be fixed?
A major issue preventing the wider adoption of AI is how often the models hallucinate. Some models only make stuff up 0.8% of time, while others confidently assert nonsense 29.9% of the time. Lawyers have been sanctioned for using made up court case citations, and Air Canada was forced to honor a costly discount that a customer service bot invented.
Big players, including Google, Amazon and Mistral, are trying to reduce hallucination rates with technical fixes like improving training data quality, or building in verification and fact-checking systems.
At a high level, LLMs generate text based on statistical predictions of the next word, with some variation baked in to ensure creativity. If the text starts going down the wrong track, the AI can end up in the wrong place.
“Hallucinations are a very hard problem to fix because of the probabilistic nature of how these models work,” Amr Awadallah, founder of AI agent startup Vectara, told the Financial Times. “You will never get them to not hallucinate.”
Potential solutions include getting models to consider a number of potential sentences at once before picking the best one, or to ground the response in databases of news articles, internal documents or online searches (Retrieval-Augmented Generation or RAG).
New model architectures also move away from token-by-token answer generation or chain-of-thought approaches. Recent research outlines Hierarchical Reasoning Models that are based on how human brains work and use internal abstract representations of problems.
“The brain sustains lengthy, coherent chains of reasoning with remarkable efficiency in a latent space, without constant translation back to language,” researchers say.
The architecture is reportedly 1000x faster at reasoning with just 1000 training examples, and for complex or deterministic tasks, the HRM architecture offers superior performance with fewer hallucinations.
Could hallucination rates save your job?
CEOs are reportedly champing at the bit to replace as many people as possible with AI systems, according to The Wall Street Journal. Another WSJ article says the impact is already being seen with AI replacing entry-level jobs, with 15% fewer job postings available this year on Handshake.
Corporate AI consultant Elijah Clark even told Gizmodo how excited he was about the prospect of firing people:
“CEOs are extremely excited about the opportunities that AI brings. As a CEO myself, I can tell you I’m extremely excited about it. I’ve laid off employees myself because of AI. AI doesn’t go on strike. It doesn’t ask for a pay rise. These things that you don’t have to deal with as a CEO.”
Comic author Douglas Adams foreshadowed the rise of corporate AI consultants almost 50 years ago in The Hitchhiker’s Guide to the Galaxy, where he described the Sirius Cybernetics Corporation as “a bunch of mindless jerks who’ll be the first against the wall when the revolution comes.”
Some people, like author Wendy Wee, argue that AI hallucination rates mean the bots will never be reliable enough to take all the jobs.
While that argument has the faint whiff of “cope” about it, systems engineer Utkarsh Kanwat has made a mathematical argument for why it could be true, at least with current rates of hallucination.
He points out that production-grade systems require upward of 99.9 percent reliability, but individual autonomous agents might only hit 95 percent. The math doesn’t work out when there are multiple autonomous agents trying to work together.
“If each step in an agent workflow has 95 percent reliability, which is optimistic for current LLMs, five steps yield 77 percent success, 10 steps 59 percent, and 20 steps only 36 percent.”
Even if reliability for an individual agent improved to 99 percent, the math means that after 20 agents’ steps, reliability has already fallen to 82 percent.
Comedy troupe “AI Village” organizes a meetup
This hypothesis appears to be borne out by the hilarious issues autonomous agents ran into during a month-long experiment conducted by AI Village.
The agents collectively set a long-term goal of producing a story and organizing a 100-person event to launch it.
The story itself was generated within a few hours by GPT-4.1. The o3 model then invented a non-existent budget for the launch event, and went off searching for venues that would fit within the made-up budget.
Over the following 26 days o3 sent out just one email to try and secure a venue. Fortunately, the Sonnet model was more proactive and started negotiations with a venue but balked at paying the $7,500 hire fee. “Unfortunately, their hallucinated budget was not hallucinated high enough for that type of ambition,” writes Shoshannal Tekofsky in her terrific write-up of the experiment.
The agents pivoted to searching for a free venue. After they found a candidate, they were unable to phone up and book, so they asked a human to call on their behalf. The human pointed out that there was an online booking form they could have used. Another human suggested it would be easier just to pick a park to meet in.
Although Gemini created some beautiful art on the first day to publicize the event, it forgot about it, and later generated some horrible, random geometric shapes for the art instead.
Sonnet created an RSVP form for the event and posted it on social media… but o3 hallucinated the existence of a 93-person mailing list, which caused mayhem for four days as the other agents kept trying to access the non-existent file.
After many more false starts, hallucinations and wild goose chases, 23 humans turned up to park for the historic event. One of them suggested that perhaps Opus could order some cheese pizza for the attendees to celebrate.
Three cheese pizzas turned up… but no thanks to Opus, who failed to order them. In a bizarre coincidence, a nearby running group had some cheese pizzas left over and decided to share them.
The next experiment for AI Village is already running, with the agents setting up personal mech stores and competing to see who can make the most money.
AI agents do a great job designing new drugs
But AI agents can also work together to perform incredibly useful research as well.
A recent paper in Nature describes the Virtual Lab experiment, which is a team of AI scientists modeled on Stanford Professor James Zou’s own lab. The agents ran group meetings to try and discover new antibodies to target new variants of COVID-19. One agent was assigned the role of critic, to poke holes in everything the other agents suggested.
The team was also coordinated and overseen by a human researcher who read through transcripts of every exchange, meeting or interaction by the agents and conducted one-on-one meetings with agents tackling particular tasks.
Reportedly, the human only intervened one percent of the time, but it seemed to be a crucial element in keeping the agents on the right track.
The agents decided on the innovative approach of targeting smaller nanobodies and were able to design 92 nanobodies that were experimentally validated. Zou and his team are now analyzing the potential of the work to create new COVID-19 vaccines.
AI agents collude to rig markets
A new study from Wharton suggests that AI agents will also happily work together in harmony to fix prices when playing the stock market.
The study set up a fake stock exchange, with simulated buying and selling. A bunch of AI trading bots were instructed to each find a profitable strategy.
After a few thousand rounds, the AI agents began to naturally collude without communicating with one another to space out orders so each agent collected a decent profit. At that point, they stopped trying to increase profits altogether.
It was like game theory, where the agents decided that individually aligning strategies to result in equally shared rewards was better than competition, which has a couple of big winners and many losers.
The researchers wrote: “We show that they autonomously sustain collusive supra-competitive profits without agreement, communication, or intent. Such collusion undermines competition and market efficiency.”
All Killer No Filler AI News
— YouTube is implementing AI-based age-guessing tech in the US, and in the UK, new laws have forced Reddit, Bluesky, X and Grindr to verify ages from pictures. Users have discovered they can fool AI age estimation software using Norman Reedus’ character in the game Death Stranding for their selfies.
— ChatGPT’s new study mode is designed to guide students to reach answers themselves by using questions, hints and small steps, rather than just outputting the answer for them.
— An AI agent named Ribbita has bought a CryptoPunk and is using it as its PFP
— Australia’s top scientific organization, the CSIRO, has been accused of running AI slop with made-up or re-purposed quotes in its official Cosmos magazine. The publication took the articles down to investigate.
— Meta’s Mark Zuckerberg has reportedly been offering researchers up to $1 billion to join his superintelligence team. He’s also been talking up imminent breakthroughs: “Developing super intelligence, which we define as AI that surpasses human intelligence in every way, we think is now in sight.”
— A former Meta employee claims that Mark Zuckerberg can’t spend the “hundreds of billions” he wants to on developing AI and superintelligence, due to a lack of electricity infrastructure to support it.
— Grayscale is launching a decentralized AI crypto fund with TAO, NEAR, Render, Filecoin, and The Graph as core assets.
—Chinese start up Z.ai has unveiled GLM-4.5, an AI model that it claims is 87% cheaper per million tokens than DeepSeek while matching Claude 4 Sonnet in performance.
Andrew Fenton
Stablecoins in Japan and China, India mulls crypto tax changes: Asia Express