There's Hope in Hard Truths

Anthropic's Claude team aired a 90-second World Cup ad juxtaposing dystopian imagery with questions about AI safety, sparking debate over the company's shift from a flagship safety pledge to nonbinding goals and the broader open weights controversy. The ad, watched by roughly a billion people, drew criticism for using real military gravestones to make a hypothetical point, while its second half asked whether AI could help with personal struggles like chronic illness and teaching.

There's hope in hard truths No single company can answer where AI goes for everyone — a case for open weights to move configuration to the people who need it. · 10 min read Argentina were a goal up on Switzerland during the World Cup Quarterfinals when the break came, and something like a billion people watched a 90-second commercial from Anthropic’s Claude team. The first half of it flashed imagery of a house burning, facial surveillance footage against a crowd, then rows of military gravestones, while a voice with no face asked “Who’s going to hit the brakes if we need to?” The verdict on this ad spot had a pretty quick turnaround. Every AI newsletter I read had a take by Monday. I use Claude every day – whether for enterprise design workflows or managing daily obstacles, the model family has had a profound impact on my career and life and I owe a lot to their existence. I’ve also been using Qwen, GPT-OSS, Deepseek, and Hermes interchangeably for several automation projects running locally on my Mac Studio. My biases in this tale run in both directions, because while I am an avid user and practitioner of Claude, I don’t buy the story Anthropic’s business model tells either — which goes something along the lines of: “this technology is so dangerous that it has to stay inside our API, under our pricing, subject to our judgment about who’s ready for it.” A danger only you can be trusted to handle is a convenient danger to keep warning about. In February the company dropped the flagship pledge https://time.com/7380854/exclusive-anthropic-drops-flagship-safety-pledge/ at the center of that story, the commitment not to release a model unless it could show in advance that its safeguards held, and replaced it with goals it describes as nonbinding https://www.cnn.com/2026/02/25/tech/anthropic-safety-policy-change . And the current open weights controversy is undoubtedly adding fuel to the debate around safety and AI sovereignty. There’s a lot to unpack here, but I believe it starts with using this video as a jumping-off point to figure out where we are today and pave the way forward for democratized AI. “We hope you’ll find the test a pleasant experience.” The first half is about what AI might do to us: surveillance screencap, disaster reel, the gravestones, a la “who’s gonna hit the brakes”. If you’ve ever watched the cult favorite 1974 thriller The Parallax View , the parallels here are hard to beat. The film’s golden moment is a roughly five-minute sequence of stimuli flashed onto the screen, a psychological screening test to identify sociopaths who could fit the role of the Parallax Corporation’s trained assassins. The montage was also conveniently built to work directly on the viewer’s psyche and reactions to such stimuli, the kind of thing that would never pass a Hollywood studio review today and precisely why I had to nod to it here . The Arlington Cemetery headstones shown during the ad spot, however, are not grounded in some fiction and belong to real people who died for reasons that have nothing to do with a large language model. The problem with such a visual is that it spends real, sometimes undisclosed grief on a hypothetical that conversely creates the feeling of guilt and/or sense of action within the viewer. A week of pile-on from tech Twitter for that kind of imagery seems somewhat appropriate. If you reach for the dead to make a point, the point had better be about them. The second half has a refreshed tone asking a different kind of question, supplemented by a linked online audio experience https://claude.com/hard-questions from Anthropic itself. One woman says it took six years to get her chronic illness diagnosed. A teacher asks whether this could help her be better at her job, and better as a mother. Someone else asks whether it could help people stop feeling misunderstood, conversely bringing people closer to one another. The cost of the controversy was that this introspective half went down with the outright bad half. All the argument that week was about marketing messaging: whether they were sincere, whether they had standing, whether graves belong in an ad break…reasonable things to argue about. But the questions themselves deserved some of that week, too. The hard truths The one redeeming quality of this campaign is that it put the perspective of people outside of the AI bubble on screen. This perspective has numbers behind it. Pew published a study measuring the distance between user and practitioner https://www.pewresearch.org/internet/2025/04/03/how-the-us-public-and-ai-experts-view-artificial-intelligence/ last year — the public is likelier to expect harm than benefit. of AI experts expect AI to benefit them personally of the U.S. public expect that same personal benefit What clicked for me wasn’t any single question in the ad, rather the phrasing of all of them. Can it, will it, who decides, who’s going to hit the brakes. Each one is asked from the outside, by someone waiting to learn what will be done to them, rather than someone learning what they can do with it. That stance is earned. The question is why, and while there exists data, we have to propose some inferences to paint a picture: 1. Dissemination of AI capabilities : Generative AI arrived inside products people already owned: search results, office software, phones. It has been switched on by default, with the opt-out buried behind menus if one exists at all . The same Pew report found one thing the public and AI experts do agree on: wanting more personal control over how AI is used on them. No one asks “what will I build with this” about something they’d never choose in the first place. 2. Perceived trust against AI systems : Four years in, the average person’s lived experience of this technology is Strawbellina video slop in their feed, cloned family voices on scam calls, and a hiring market that got worse for their kids while the solution remains to be seen, i.e., in the future tense. Anthropic’s own Public Record https://www.anthropic.com/news/anthropic-public-record survey of 52,000 Americans ranks the fears in that order of arrival: of Americans worry about job loss about depending on machines to think about misinformation The jobs fear is measurable: a Stanford analysis of ADP payroll data https://fortune.com/2026/06/27/what-is-ai-impact-entry-level-jobs-stanford-adp-canaries-brynjolfsson-richardson/ found employment for early-career workers in the most AI-exposed occupations down 13% relative to their less-exposed peers. What do you call a system that promises the technology will solve societal problems, that can’t promise to gracefully carry people from impacted sectors to needed jobs? 3. The moving goalpost of “safe AI” : The institutions that promised to hold the line didn’t. February’s dropped pledge is at the top of this page; March was worse. The Washington Post reported https://www.washingtonpost.com/technology/2026/03/04/anthropic-ai-iran-campaign/ that CENTCOM used Claude to help generate around a thousand prioritized targets in the first day of the Iran campaign. Anthropic hadn’t agreed to that. They’d been blacklisted weeks earlier https://responsiblestatecraft.org/ai-war-iran/ for refusing the Pentagon unrestricted access, but it happened anyway. Control is the full sales pitch for keeping a model in a closed ecosystem, and in the one test that mattered this year, the control appeared self-contradictory. So when the Public Record finds only 15% of Americans trust AI companies to decide how the technology develops, one must ask whether the sentiment is cynicism or if it’s a recurring pattern. Add these up and the truth is hard to dodge: no single company, team, or governing body can decide what trustworthy AI looks like for the entire human world — not because they shouldn’t, but they cannot because of capacity. What makes a modern system work for one person conflicts with what makes it work for another, and one family of models under one set of terms can’t hold true in solving for all of the diverse challenges experienced globally, let alone in local communities. Own your weights, own your product The hopes in Anthropic’s survey are as specific as its fears: curing diseases like cancer and Alzheimer’s tops the list at 48%. The average delay for a rare disease runs about five years https://www.nature.com/articles/s41431-024-01604-z , through additional averages of seven or eight specialists and two or three misdiagnoses. The same research found that patients who are able to conveniently reach the right expert center get their answer in under two months. The knowledge has been there the whole time. What is failing people today is access. After years of building AI systems in places where being wrong is expensive, this is the one claim I’ll make: moving scarce expertise to the person who needs it is the problem this technology should solve for. This technology will not end war or famine; those are failures of political will, and no model can move that on its own. I used to wave off the open-weights argument, since weights alone don’t get a model into a clinic. The deployment work, the integration with how/where people already operate, the infrastructure requirements — that dirty work is most of the job. That’s still true, and it’s now the reason I buy the argument. Because this dirty work is most of the upfront work, the only people who can do it are the ones near the problem: the clinic network that knows its own referral patterns and where they break down, the farming co-op that knows the language its members actually work in, the school district that knows its students and what its classrooms can afford. A closed API means that last mile only gets built where the pricing page and the terms of service allow. Weights you hold can’t be repriced, deprecated, or geofenced. A clinic network whose privacy law says patient records never leave the building can run the model inside the building. A co-op can fine-tune for a language that will never earn a line on a frontier lab’s roadmap. A school district that bought its hardware once doesn’t discover in year three that a new pricing tier decides which students keep their AI tutor. None of these deployments are exotic by any means, yet each is just impossible when the model answers to someone else’s terms of service and someone else’s margins. Clinic network Integration Diagnosis support surfaced inside the existing EHR workflow, at the point of care Adaptation RAG over referral history and patient records behind the hospital firewall; checkpoint pinned at the clinically validated version Serving Air-gapped on-prem GPU node running a local inference server Farming co-op Integration SMS and voice interface in the language its members actually work in Adaptation LoRA fine-tune on local-language agronomy guidance and the co-op's own field records Serving 4-bit quantized build on shared office hardware, fully offline School district Integration Tutor embedded in classroom devices, mapped to the district's own curriculum Adaptation Tuned for grade-level reading and the district's course material Serving District-owned servers; one-time hardware purchase, no per-seat inference billing Base model Self-hosted open-weights checkpoint Qwen, GPT-OSS, DeepSeek, or similar model , integrated under a permissive license Example idealized architecture references. Real deployments carry constraints this diagram elides, like procurement, compliance sign-off, eval coverage, ongoing maintenance, etc. This is the target state open weights make possible, and addresses the value proposition AGI has been positioned around. The standard objection is that people will build it wrong, and some will. People build traffic intersections wrong, too. Intersections are among the most lethal pieces of infrastructure we have, and nobody’s answer was to patent them and license crossings to qualified municipalities. The answer was right-of-way rules, signage, and traffic engineering as a public discipline. The wider the standard spread, the safer the intersections got. The capabilities the second half of that ad needs, diagnosis-grade reasoning and broad language coverage, already exist in open models a year or so behind the frontier, so withholding them decides who gets to build long before it decides anyone’s safety. Closing the weights stops a clinic, but not the Pentagon. Go build the route Eighteen months ago I wrote on AI legislation /blog/ai-regulation , and offered the call-to-action to either advocate for legislation or go build the ethical version yourself. I’ve been busy enough with the second option, building the right way to AI solutions for a few big fish, which admittedly took time away from my writing. Which is probably why a week of takes about an ad was what pulled me back in. My ask now is that the organizations leading AI keep widening the ecosystem, in open weights and in everything around them. Anthropic, this one’s yours. On July 24, some seventy-five companies signed an open letter to Washington https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf arguing that open weights are a condition of American AI leadership, and warning against restrictions that would push the building overseas. Google signed. Microsoft signed. OpenAI signed. Anthropic’s name wasn’t there, and an earlier draft of this post ended by asking why. Then an answer arrived https://www.anthropic.com/news/position-open-weights-models : Anthropic has never advocated a ban on open weights, open models without dangerous capabilities are “a public good,” and what the company wants instead is chip export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for every sufficiently capable model, open or closed. I’ll take the yes, and parts of it are the right shape. Mandatory testing for everyone is the civil-engineering-esque answer: right-of-way rules adopted by all instead of patented solutions. That to me seems a fairer debate to have in the open than one over whether weights get released at all. But a test that applies equally to open and closed models is only equal on paper. A closed model that fails an eval can ship behind mitigations, can be patched the following week, and can be pulled if something breaks. An open release has to pass once and forever; fine-tuning strips the guardrails, and there is no recall. That can only swing one way for open weights, which is how “not a ban” becomes a conditional ban run by whoever determines the test criteria. “Ideally global” is carrying weight too. Beijing is not submitting DeepSeek’s next checkpoint to an American evaluator, so in practice the organizations there bind the releases we can inspect and wave through the ones we can’t. Orwell’s super-states in 1984 ran three nominally opposed ideologies much like our current geopolitical giants today that were indistinguishable in practice — so a safety regime that only binds the side you can see is unfortunately the same trick. Here is the version I would co-sign: test criteria written and argued over in public, the way engineering or civil safety gets argued, and binding first on the lab that proposed them. All this to say: we are not stupid citizens, we’re ready to engage in what gets built out, but we need the correct infrastructure available to do so. I’ve already written about and applied adoption of pre-existing frameworks built around human factors which needs a deeper 2026 update on this site , and this sort of discussion around open weights can easily be solved by leveraging what is already out there…let’s just agree to not “vibe-govern” new frameworks from scratch. Please. If you, the reader, are still here: one more number from the Public Record points the way out. Americans who use AI at work every day are sixteen points less worried about losing their job to it than people who don’t use it at all. Familiarity converts fear, and open capability is what makes familiarity possible. So use open models, build on them, and put work into the commons that makes them usable, the evals, the local-language datasets, the deployment recipes, the funding for the groups that keep releasing them. Somebody is in year two of a six-year path to diagnosis right now, and what they need already exists in someone else’s head. Go build the route to them. ■ If you’re building something aimed at any of this, or you think I’ve got the open-weights argument wrong, email me. I’d like to learn more about your perspective.