A Christian Network Blocks Porn While Its Audience Buys $5 AI Bible Videos

01A Christian Phone Network Will Block Porn Next Week. The Same Audience Is Buying AI Bible Videos on Fiverr for $5.

A new US cell phone network marketed to Christians launches next week with porn blocked at the network level — the first American carrier to do so in a way the account holder cannot turn off, according to MIT Technology Review. The same filter also catches "gender-related content," a category the operator has not fully defined.

Network-security experts told the publication that this represents a precedent for consumer telecom in the United States. Existing parental-control products on T-Mobile, Verizon, and AT&T require an account owner to enable filtering and let them disable it. The Christian carrier removes that switch. Filtering happens at the DNS and network layers before traffic reaches the phone.

While one corner of the same demographic is paying to wall content off, another is paying to flood it in. The Verge reported this week that Christian content creators are commissioning AI-generated Bible videos from Fiverr freelancers for $5 to $20 a piece. The freelancers' profiles advertise turnaround in hours, with output measured in clips per day.

The clips run on TikTok under accounts framed as Christian ministry. One format animates Old Testament scenes such as Noah, Moses, and Daniel in a glossy 3D style that draws millions of views. Buyers told The Verge they treat the videos as evangelism, not entertainment. Some accounts post several pieces of AI content per day, faster than a human animator could produce a single scene.

The infrastructure on both ends comes from the same generation of tools. Network filtering depends on automated category classification of domains. The Fiverr Bible factories depend on text-to-video and text-to-image models that did not exist in consumer form two years ago. Both pipelines are sold to a religiously conservative audience with the same promise: remove human judgment from the loop, whether to keep impurity out or to keep output up.

Neither side has published the rules its software follows. The Christian carrier has not released the domain list its "gender-related content" filter blocks. Fiverr sellers are not required to disclose which models they use. The carrier launches next week; the Bible accounts are already live.

02The God Delusion author read Claude's writing on his podcast, then told listeners to subscribe

Richard Dawkins, 84, spent four decades teaching readers that fluent storytelling is not evidence of truth. On a recent episode of his podcast, the evolutionary biologist read aloud several passages Anthropic's Claude had written for him. He called the prose impressive and recommended his audience pay for a subscription.

Gary Marcus, the cognitive scientist who has built a second career criticizing large language model hype, published a response on Substack titled "Richard Dawkins and The Claude Delusion." The title is a deliberate echo of Dawkins's 2006 bestseller "The God Delusion," which argued that emotional resonance and beautiful language are not proof of an underlying intelligence. Marcus's point: the author who codified that warning had just fallen for its software equivalent.

Marcus's argument is narrow. He does not claim Claude is bad at writing. The writing itself, he says, is the trap. A skeptic, of all people, should recognize that producing paragraphs that sound like thought is not the same as thinking. That was the entire premise of the New Atheist case against scripture. Dawkins read passages, found them moving, and treated that response as data about the system that produced them.

Anthropic did not solicit the endorsement, and Dawkins is not a paid spokesman. The episode is a clean test case: a hostile witness, no commercial incentive, decades of training in skeptical method, persuaded by a few hundred words of model output.

The same week Dawkins was reading Claude aloud, the Academy of Motion Picture Arts and Sciences moved in the opposite direction. New eligibility rules announced this week make AI-generated actors and AI-written screenplays ineligible for Oscars consideration, a categorical exclusion with no quality threshold attached. One of the world's most prominent rationalists was being persuaded by a chatbot's prose. At the same moment, one of entertainment's most prominent institutions decided no chatbot output would count, regardless of how good it was.

03DeepSeek V4 lands almost at the frontier, and a paper makes the same tier fine-tunable on a 3090

Two open-source releases dropped the same week, and together they compress the cost of running and tuning a near-frontier model at both ends.

DeepSeek released V4. Independent reviewer Simon Willison evaluated it as "almost on the frontier." The remaining gap to GPT-class and Claude-class systems no longer determines outcomes on many production workloads. The model weights are open. A team running it on owned hardware pays compute, not a per-token API margin.

That removes half of the closed-stack lock-in. The other half is fine-tuning, which has historically required rented H100-class hardware. Consumer GPUs hit two walls: limited VRAM, and slow PCIe interconnect between cards. Pipeline parallelism with CPU offload helps with both. But existing schedules run into what the RoundPipe paper calls the "weight binding issue." Uneven model stages get pinned to specific GPUs, and throughput is then capped by the most overloaded one. The LM head, being disproportionately large, is the typical bottleneck.

RoundPipe, posted to Hugging Face this week, proposes a schedule that decouples weight assignment from physical GPU placement and rotates responsibility through the pipeline. The paper claims multi-3090 and multi-4090 setups can hit fine-tuning throughput previously available only on data-center hardware. Combined with CPU offload, the practical claim is that fine-tuning a near-frontier-class open model no longer requires renting an 8xH100 node by the hour.

The two cost curves were previously coupled. A team that wanted a custom fine-tune of a competitive model had to rent training hardware to produce it, then keep renting for inference or pay an API. With V4 weights and a consumer-GPU training path, that same workload can in principle live entirely on owned hardware. Existing 3090 and 4090 inventory becomes capital that runs both stages, not just one. Whether RoundPipe's reported numbers hold up outside the paper's setup will determine if the path is real or theoretical.