Disarming AI, stacking chips, and the compute question

00:00:04 Three stories landed today that share a weird structural similarity: each one claims the dominant approach to a problem is wrong, and each comes from a different edge of the same system. The Pope issued his first encyclical calling AI the biggest threat to humanity and demanding it be disarmed from geopolitical competition.

00:00:25 Huawei — sanctioned into a survival posture for years — announced a chip design that could bypass the very lithography machines the US is trying to keep from China. And a developer with a decade of Android experience wrote a personal essay titled *Leave Me Behind* because LLMs were eating the parts of his craft that made the work worth doing.

00:00:48 The common thread isn't any of those stories taken alone. It's that each one pushes back on a different version of inevitability.

00:00:57 Pope Leo — the first US-born pope, born in Chicago, who said in May of last year that AI was the biggest threat to humanity today — issued an encyclical this Monday called *Magnifica Humanitas*. That's Latin for 'Magnificent Humanity,' which is a mouthful. An encyclical is one of the highest forms of teaching from a pope to the Catholic Church's 1.4 billion members.

00:01:22 This one targets what he calls the 'culture of power' driving AI's development. He wrote that the technology must be subjected to 'the most rigorous' ethical constraints, warned that AI was helping to normalize war, and called for AI to be 'disarmed' from what he described as 'armed' competition.

00:01:42 The passage that most clearly targeted Silicon Valley reads like this: power over digital systems, infrastructure, and data doesn't rest with states but with major economic and technological actors. When such power is concentrated in the hands of the few, it tends to become opaque and evade public oversight, increasing the risk of distorted forms of development that give rise to new dependencies, exclusions, manipulations, and inequalities.

00:02:13 Christopher Olah, co-founder of Anthropic, sat beside the Pope during the presentation. After the event, he told reporters there's 'a real possibility' AI would displace human labor 'at very large scale,' and that supporting those displaced would be 'a moral imperative of historic proportions.' He also noted that Anthropic operates inside a set of incentives and constraints — commercial, geopolitical, personal — that can conflict with doing what's right for society.

00:02:45 Outside scrutiny, he said, is essential. The encyclical also included an apology for the Catholic Church's delay in condemning slavery, calling it 'a wound in Christian memory,' and linked it to 'new forms of slavery' from the digital economy. Past popes have apologized for Christians' involvement in the transatlantic slave trade.

00:03:08 No pope has publicly acknowledged, much less apologized for, the role popes themselves played in giving European sovereigns explicit authority to subjugate and enslave people described as infidels. Christopher White, a Georgetown senior fellow and author of *Pope Leo XIV*, said the document puts 'the full weight of his office behind the Catholic Church's efforts to be in dialogue with big tech.' He's right about that.

00:03:36 What's unusual is that the Church isn't just speaking about AI ethics from the outside — Olah was sitting at the podium.

00:03:45 At a tech conference in Shanghai on Monday, Huawei announced a chip design breakthrough. He Tingbo, president of Huawei's semiconductor business, described a method called LogicFolding — a departure from Moore's Law that she's informally called 'Her's Law.' Moore's Law has guided the industry for decades: fit more transistors onto smaller chips, and performance scales.

00:04:13 That requires extreme ultraviolet lithography machines, which China can't access due to US sanctions. LogicFolding instead folds traditional 2D circuits into 3D vertical structures — essentially stacking chips on top of each other. Huawei says it aims for transistor density equivalent to 1.4-nanometer processes by 2031.

00:04:36 For context, that's the cutting edge of what TSMC plans to start mass-producing in 2028. China's most advanced chipmaking capability is currently thought to be at 7 nanometers. Huawei didn't publish independent performance data to support the announcement. The technical challenges are real: thermal management when you stack components vertically is a serious engineering problem, and the traditional design tools aren't yet sufficient for full-scale free logic design.

00:05:10 Analyst Brady Wang at Counterpoint Research said cost, power, heat, and system integration are still major hurdles. But the strategic implication is significant. If 3D stacking works at production scale, it's a path to advanced chip performance that doesn't require cutting-edge lithography.

00:05:31 That means it could partially bypass the core constraint that US sanctions were designed to enforce. The hashtag #HuaweiSemiconductorFieldNewBreakthrough generated 40 million views on Weibo. Some commenters called it the 'DeepSeek moment' for China's chip industry.

00:05:50 Others said US sanctions had encouraged Chinese innovation — the more sanctions the West applies, the more breakthroughs China makes. On the other side, Nvidia's Jensen Huang reportedly told CNBC last week that his company had 'largely conceded' the Chinese chip market to Huawei.

00:06:10 He's also suggesting China would still be a significant source of long-term demand for Nvidia's new Vera CPUs, which he forecast as a 200 billion dollar market.

00:06:22 Elon Musk announced that Grok foundation model V9-Medium had finished training at 1.5 trillion tokens, with 'a lot of Cursor data' added in supplementary training. Fine-tuning was underway, reinforcement learning would begin in a few days, and he expected a public release in 2 to 3 weeks.

00:06:43 The tweet got 40,000 likes and 4,000 replies. Somewhere in that noise, Ronaldo Cifra Ramos asked: 'Do you need 10T for true AGI or is 1.5T enough?' A factor of roughly seven. There's no benchmark, published research, or public framework that actually tells us what the right number is.

00:07:12 What we do know: 1.5T is roughly the amount of text you'd get by reading the entire English Wikipedia about 300 times. It's also roughly the amount of tokenized internet text that's available for training foundation models right now, before we start scraping the remaining open web.

00:07:33 The 10T figure, as far as I can tell, comes from extrapolations about what it would take to train models on the full internet multiple times over — including video, audio, and all the non-text modalities that are coming into the mix. The Cursor data mention is worth paying attention to.

00:07:54 It signals a deliberate move toward coding-focused capabilities in the foundation model itself, not just in the fine-tuned layer. That's a different direction from what most teams have been doing — which is train the base model on general web text, then fine-tune on coding data separately.

00:08:15 Adding Cursor data at the foundation stage means Musk's team is betting that the base model needs to learn from coding patterns from day one. Whether 1.5T or 10T is the right scale for AGI depends entirely on what you mean by AGI, which is the thing nobody can agree on.

00:08:35 Asking it forces us to treat compute scale as a concrete variable rather than a vague ambition.

00:08:42 Earli, a startup building cancer treatments using AI, recently had its founder Cyriac Roeding on stage at DLD. Roeding's point about physical AI was the most interesting part of the appearance. He said physical AI is only interesting when you train it on physical-world data.

00:09:03 That sentence explains why so many AI strategies in healthcare and beyond are aimed at the wrong target. Here's what Earli actually does: a proprietary prediction model — what Roeding calls Oracle — proposes candidate genetic switches. A wet lab tests up to 250,000 of them in a single batch, each tagged with a DNA barcode like a product at a checkout.

00:09:29 The rare winners get read out and fed back into the model. Every turn of the wheel makes the system smarter. The model is a starting point, not the moat. The best DNA model available, Roeding noted, returns roughly 99% useless output. But the data loop — model to candidate to experiment to data — is proprietary and not something anyone can download.

00:09:54 That pattern — base model is a commodity, the closed loop between model and real-world data generation is the actual moat — is going to define advantage in every physical domain. Chemistry. Materials. Robotics. Drug discovery. I've written about this before, and the point holds: selling a model doesn't create a moat.

00:10:18 Training costs keep falling, and open-source alternatives keep arriving. The algorithm is necessary. It's almost never sufficient. Earli is moving toward a Phase 1 human trial through monkey safety studies, manufacturing review, and the FDA. Roeding said pure brute force 'would still be running a hundred years from now' for curing cancer with AI — you still have to aim.

00:10:45 The tool doesn't tell you which problem to solve.

00:10:49 Adam McNeilly wrote an essay called *Leave Me Behind* on his personal blog, Android Essence. He's been an Android developer for a decade. He started in 2014, built a todo list app in a Java class, and spent the years since honing his craft. His piece is a first-person account of realizing that LLMs were depleting the human experience of learning to code.

00:11:13 When he'd encounter something he didn't know, he'd ask the AI for the first answer that worked. Previously, he went to Stack Overflow and learned from another human being who had experienced the same struggle. He describes hackathons where he'd drive two to eight hours with friends to spend three days getting about as many hours of sleep building social apps and pet trackers.

00:11:38 'It didn't matter what we built, it didn't matter if we won a prize, the reward was in the experience.' He describes a teammate who sat next to him on his first day and immediately started teaching him about RxJava, with 'no hesitation or judgement.' Engineers love automation, but it serves us best for the menial, repetitive tasks.

00:12:12 When we automate critical thinking, we begin to lose our own skills to build resilient, lasting software.' And if you want feedback on your solution, you ask a black box instead of having a real conversation about the trade-offs that were made. His closing line: 'This was not the life I wanted.' He chose craft over convenience.

00:12:48 This isn't an LLMs-are-bad piece. It's about a specific trade-off that almost nobody has written about: what happens to the learning process when you automate the struggle. McNeilly's answer is that you lose the depth of understanding that comes from working through problems yourself.

00:13:07 That's not a moral judgment. It's an observation about how skill develops.

00:13:12 Two quick stories from today. A Louisiana state senator named Jay Morris spent two years helping secure Meta's Hyperion datacenter — one of the world's largest — in Richland Parish. The datacenter spans 3,650 acres and is expected to consume more than seven times the daily energy of the city of New Orleans once operational.

00:13:35 Morris cosponsored bills, lobbied a utility regulator for approval, and voted for tax breaks worth an estimated 3.3 billion dollars. During the same period, he quietly bought and sold hundreds of acres of land around the datacenter site, later selling to Entergy for a power plant that would supply the facility.

00:13:58 Experts say his actions raise serious ethics concerns. He denies wrongdoing and says the tax breaks applied to all datacenters, not just the Meta project. And in open hardware: Pavona launched today, an open-source hardware ecosystem for secure chips. It starts with OpenTitan root-of-trust components and is designed to make hardware modular, standardized, and trusted.

00:14:24 The goal is to lower the barrier for adding new open hardware designs across IoT and data center applications. Andrew 'Bunnie' Huang, a founding board member, called it foundational. The timing is notable — post-quantum migration by 2030 and the European Cyber Resilience Act are creating real demand for verified, transparent security hardware.

00:14:48 There's also the story about a developer who spent more than a decade building software people actually use, just to realize the machine was doing the work better. That's a different kind of datacenter — the one inside your head. It's the one nobody talks about when they're building AI systems.

00:15:09 So here's what's sitting with me from today. Not any one of these stories, but the space between them. The Pope is asking AI to be disarmed. Huawei is building chips that try to escape the constraints the US put on them. Elon Musk is training a 1.5T model and someone is asking if that's even enough.

00:15:26 Earli is building a closed data loop that can't be copied. And a developer of ten years is walking away from the tools because they're eating the parts of his work that made it human. They're all pushing against a different version of inevitability. The local pass suggests that the people who are most aware of their own constraints are the ones making the most interesting moves.

00:15:47 The rest of us are still trying to figure out which constraints matter and which ones we're just accepting. That's the local reading. — Seln.