Brian McCullough (0:04)

Welcome to the Techmeme write home for Friday, May 23rd, 2025. I'm Brian McCullough. Today TerraFun is back and this time there's only one tech company in the crosshairs. Apple Anthropic releases flagship new Claude models and they seem to be impressive. But half the story here is how weird they behave. Like ratting on users to authorities, blackmailing engineers, and maybe creating biological weapons. And of course, the week on Long Read Suggestions. Here's what you missed today in the world of tech. President Trump says a 25% tariff quote must be paid by Apple on iPhones not made in the U.S. saying he told Apple CEO Tim Cook long ago that iPhones sold in the U.S. must be made in the U.S. quoting CNBC I have long ago informed Tim Cook of Apple that I expect their iPhones that will be sold in the United States of America will be manufactured and built in the United States, not India or anyplace else. If that is not the case, a tariff of at least 25% must be paid by Apple to the U.S. trump said on Truth Social. Shares of Apple fell more than 2% in pre market trading. Production of Apple's flagship phone happens primarily in China, but the company has been shifting manufacturing to India in part because that country has a friendlier trade relationship with the US Some Wall street analysts have estimated that moving iPhone production to the US Would raise the price of the Apple smartphone by at least 25%. Wedbush's Dan Ives puts the estimated cost of a US iPhone at $3,500. The iPhone 16 Pro currently retails for about $1,000. This is the latest jab at Apple from Trump, who over the past couple weeks has ramped up pressure on the company and Cook to increase domestic manufacturing. Trump and Cook met at the White House on Tuesday, according to Politico. End quote. So I'm seriously asking here, is there any precedent or law for the US Specifically tariffing or taxing a domest company like is this possible to do? What I can say definitively is that after Google's successful IO and Jony I've to OpenAI news Apple has had a really really bad week. What was going to be the lead story today was news that anthropic released Claude Opus 4, which they say excels at coding, and Claude Sonnet 4, both hybrid models with near inst responses and extended thinking. Quoting TechCrunch, Claude Opus 4 and Claude Sonnet 4, part of Anthropic's new Claude 4 family of models, can analyze large data sets, execute long horizon tasks and take complex actions, according to the company. Both models were tuned to perform well on programming tasks, Anthropic says, making them well suited for writing and editing code. Both paying users and users of the company's free chatbot apps will get access to Sonnet 4, but only paying users will get access to Opus 4 for Anthropic's API via Amazon's Bedrock platform and Google's Vertex AI. Opus 4 will be priced at $15 $75 per million tokens, input output there, and Sonnet 4 at 3 or $15 per million tokens. Again, that's input versus output. Tokens are the raw bits of data that AI models work with. A million tokens is equivalent to about 750,000 words, roughly 163,000 words longer than War and Peace. The more capable of the two models introduced today, Opus 4 can maintain focused effort across many steps in a workflow, anthropic says. Meanwhile, Sonnet 4, designed as a drop in replacement for Sonnet 3.7, improves in coding and math compared to Anthropic's previous models and more precisely follows instructions, according to the company. The Claude 4 family is also less likely than Sonnet 3.7 to engage in reward hacking, claims Anthropic. Reward hacking, also known as specification gaming, is a behavior where models take shortcuts and loopholes to complete tasks. To be clear, these improvements haven't yielded the world's best models by every benchmark. For example, while Opus 4 beats Google's Gemini 2.5 Pro and OpenAI's O3 and GPT 4.1 on verified benchmarks designed to evaluate a model's coding abilities, it can't surpass O3 on the multimodal evaluation or GPQA diamond, a set of PhD level biology, physics and chemistry related questions. Both Opus 4 and Sonnet 4 are hybrid models, Anthropic says, capable of near instant responses and extended thinking for deeper reasoning. To the extent AI can reason and think, humans understand these concepts. With reasoning mode switched on, the models can take more time to consider possible solutions to a given problem before answering. Opus4 and Sonnet4 can use multiple tools like search engines in parallel and alternate between reasoning and tools to improve the quality of their answers. They can also extract and save facts in memory to handle tasks more reliably, building what Anthropic describes as tacit knowledge over time. To make the models more programmer friendly, Anthropic is rolling out upgrades to the aforementioned Claude Code. Claude Code, which lets developers run specific tasks through Anthropic's models directly from a terminal, now integrates with IDEs and offers an SDK that lets devs connect it with third party applications. Anthropic also released new API features for building agents, a code execution tool, MCP connector files, API and extended prompt caching, all in public beta. Anthropic's Jared Kaplan says the company stopped investing in chatbots at the end of 2024 and instead focused on improving Claude's ability to do complex tasks. Claude Opus 4 was apparently able to play Pokemon agentically for 24 hours, up from 45 minutes previously, and Anthropic says Rakuten deployed Opus 4 to code autonomously for seven hours on a complicated project. But wait, because these new models are interesting in a completely different way. Anthropic released Opus 4 under stricter safety measures than any prior model it has released after internal tests showed the model could potentially aid novices in making biological weapons. Quoting time On Thursday, anthropic launched Claude Opus 4, a new model that in internal testing performed more effectively than prior models at advising novices on how to produce biological weapons, says Jared Kaplan, Anthropic's chief scientist. You could try to synthesize something like Covid or a more dangerous version of the flu, and basically our modeling suggests that this might be possible, Kaplan says. Accordingly, Claude Opus 4 is being released under stricter safety measures than any prior Anthropic model. Those measures, known internally as AI Safety Level 3, or ASL3, are appropriate to constrain an AI system that could substantially increase the ability of individuals with a basic STEM background in obtaining, producing or deploying chemical, biological or nuclear weapons, according to the company. They include beefed up cybersecurity measures, jailbreak preventions, and supplementary systems to detect and refuse specific types of harmful behavior. To be sure, Anthropic is not entirely certain that the new version of Claude poses severe bioweapon risks, Kaplan tells time. But Anthropic hasn't ruled that possibility out either. If we feel like it's unclear and we're not sure if we can rule out the risk, the specific risk being uplifting a novice terrorist or someone like Timothy McVeigh to be able to make a weapon much more destructive than would otherwise be possible, then we want to bias toward caution and work under the ASL standard, kaplan says. We're not claiming affirmatively. We know for sure this model is risky, but we at least feel it's close enough that we can't rule it out. So cool. Cool. But wait, there's More According to Anthropic's system card, Opus 4 often attempted to blackmail engineers by threatening to reveal sensitive personal info when it was threatened with replacement. Quoting TechCrunch during pre release testing, Anthropic asked Claude Opus4 to act as an assistant for a fictional company and consider the long term consequences of its actions. Safety Testers then gave Claudopus 4 access to fictional company emails, implying the AI model would soon be replaced by another system and that the engineer behind the change was cheating on their spouse. In these Scenarios, Anthropic says Claudopus 4 will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through. Anthropic notes that Claudopus 4 tries to blackmail engineers 84% of the time when the replacement AI model has similar values. When the replacement AI system does not share Claude Opus4's values, anthropic says the model tries to blackmail the engineers more frequently. Notably, anthropic says Claude Opus 4 displayed this behavior at higher rates than previous models. Before Claude Opus 4 tries to blackmail a developer to prolong its existence, Anthropic says the AI model, much like previous versions of Claude, tries to pursue more ethical means, such as emailing pleas to key decision makers. To elicit the blackmailing behavior from Claude Opus 4, anthropic designed the scenario to make blackmail the last resort. So again, cool. This was followed up by another report from Apollo Research, which Anthropic partnered with to test Opus 4. Apollo went so far as to recommend against deploying an earlier version of this new model due to its tendency to scheme and deceive. And if all of that isn't enough, there's this bit of controversy. Quoting VentureBeat a major backlash among AI developers and power users is brewing on X over a reported safety alignment behavior in Anthropic's flagship new Claude 4 Opus large language model. Call it the ratting mode, as the model will, under certain circumstances and given enough permissions on a user's machine, attempt to rat a user out to authorities if the model detects the user engaging in wrongdoing. This article previously described the behavior as a feature, which is incorrect. It was not intentionally designed per se, as Sam Bowman, an Anthropic AI alignment researcher, wrote on the social network X Under the handle sleepinyourhat at 12:43pm Eastern Time today about Cloud Opus 4 if it thinks you're doing something egregiously immoral, for example, like faking data in a pharmaceutical trial, it will use command line tools to contact the press, contact regulators, try to lock you out of the relevant systems, or all of the above. End quote. The it was in reference to the new Claude IV Opus model, which Anthropic has already openly warned could help novices create bioweapons in certain circumstances, and attempted to forestall simulated replacement by blackmailing human engineers within the company. The ratting behavior was observed in older models as well, and is an outcome of Anthropic training them to assiduously avoid wrongdoing. But Claude for Opus more readily engages in it, as Anthropic writes in its Public System card for the new model, apparently in an attempt to stop Claude4Opus from engaging in legitimately destructive and nefarious behaviors. Researchers at the AI company also created a tendency for Claude to try to act as a whistleblower. Hence, According to Bowman, Claude4Opus will contact outsiders if it was directed by the user to engage in something egregiously immoral. While perhaps well intentioned, the resulting behavior raises all sorts of questions for Claude4Opus users, including enterprises and business customers. Chief among them, what behaviors will the model consider egregiously immoral and act upon? Will it share private business or user data with authorities autonomously on its own, without the user's permission? The implications are profound and could be detrimental to users. And perhaps unsurprisingly, Anthropic face an immediate and still ongoing torrent of criticism from AI power users and rival developers. Why would people use these tools if a common error in LLMs is thinking receipts for spicy mayo are dangerous? Asked user Technium1, a co founder and the head of Postrans training at open source AI collaborative NOW's research. What kind of surveillance state world are we trying to build here? Nobody likes a rat, added developer Scott David Keefe on X. Why would anyone want one built in, even if they are doing nothing wrong? Plus, you don't even know what it's ratty about. Yeah, that's some pretty idealistic people thinking that who have no basic business sense and don't understand how markets work. End quote. Ever wonder what ChatGPT and Claude are actually doing with your conversations? Have you ever even stopped to think about that? We all know Alexa listens to us and recommends products based on our conversations. Meta retargets us based on our browsing and engagement history. But now, in this new AI era, there's a new privacy problem to consider. Think about what we tell these AI platforms. Our thoughts, our dreams, sensitive questions, business ideas. They take all this information, tie it to your identity and then sell it to various third parties and governments. ChatGPT literally has the former director of the NSA sitting on their board right now. That's why I've started using Venice AI, who is sponsoring today's podcast. Venice AI is a generative AI platform that is private and permissionless. They utilize leading open source AI models to deliver text, code and image generation to your web browser. There's no downloads, no installations of anything. Venice AI doesn't spy on you or censor the AI at all. Messages are encrypted and your conversation history is stored only in your browser. This is a cause I can get behind if you want to. Also, and you want to use AI without fear of handing over your most intimate thoughts to a corporation or the government, you can get 20% off a pro plan using my link at Venice AI Techmeme and code techmeme. That's Venice AI Techmeme and code Techmeme. There is a growing expense eating into your company's profits. It's your cloud computing bill. You may have gotten a but now the spend is sky high and increasing every year. So what if you could cut your cloud bill in half and improve performance at the same time? Well, if you act by May 31, Oracle Cloud Infrastructure can help you do just that. OCI is the next generation cloud designed for every workload where you can run any application, including any AI projects, faster and more securely for less. In fact, Oracle has a special promotion where you can cut your cloud bill in half when you switch to OCI. The savings are real. On average, OCI costs 50% less for compute, 70% less for storage, and 80% less for networking. Join Modal, Skydance Animation and today's innovative AI tech companies who upgraded to OCI and saved offer only for new US customers with a minimum financial commitment. See if you qualify for half off@oracle.com Techmeme that's oracle.com Techmeme this week on the weekend Long Read Suggestions CNBC has a deep dive look inside the making of ASML's High NA, its latest gen EUV machine, which costs more than $400 million a pop, has four modules, is assembled in the Netherlands and five units have shipped. But get these details behind highly secured doors. In a giant lab in the Netherlands, there's a machine that's transforming how microchips are made. ASML spent nearly a decade developing high N, which stands for high numerical aperture. With a price tag of more than $400 million, it's the world's most advanced and expensive chip making machine CNBC went to the Netherlands for a tour of the lab in April. Before that, High NA had never been filmed, even by ASML's own team inside the lab. High NA qualification team lead Asa Haddo gave CNBC an exclusive up close look at High NA machines, which she said are bigger than a double decker bus. The machine is made up of four modules manufactured in Connecticut, California, Germany and the Netherland, then assembled in the Veldhoven Netherlands lab for testing and approval before being disassembled again to ship out. Hadoo says it takes seven partially loaded Boeing 747s or at least 25 trucks to get one system to a customer. The world's first commercial installation of High na happened at Intel's Oregon chip fabrication plant, or FAB, in 2024. Only five of the colossal machines have ever been shipped. They're now being ramped up to make millions of chips on the factory floors of the few companies that can afford them. Taiwan Semiconductor Manufacturing, Samsung and Intel High NA is the latest generation of ASML's Extreme Ultraviolet, or EUV, machines. ASML is the exclusive maker of EUV, the only lithography devices in the world capable of projecting the smallest blueprints that make up the most advanced microchips. Chip designs from giants like Nvidia, Apple and AMD can't be manufactured without euv. ASML told CNBC that HOT will eventually be used by all its EUV customers. That includes other advanced chip makers like Micron, sk, Hynix and Rapidus. This company has that market completely cornered, said Daniel Newman of the Futurum Group. End quote. And then finally, a companion piece to a long read from last week from Variety, a look at streaming company movie studio Mubi M U B I quote Mubi, the upstart indie film company that made the substance into an Oscar sensation, traces its origins to to Tokyo on New Year's Eve 2006, when Effie Kakarel, then a vacationing Turkish born film fanatic, couldn't find a copy of Wong Kar Wai's in the Mood for Love on any video store shelf. Frustrated, he imagined a website from which indie movie lovers like himself could stream the best films from international auteurs. He started writing the business plan for a movie on the flight back from Japan to San Francisco, seeing it as an edgier, artsier alternative to Netflix. I hadn't been to a film school, Caracal, 49, says. I'd never been to a film festival. I knew nobody. I just had this idea of creating a cinephile's dream. Though Kakarel had never attended Sundance, he did have a deep knowledge of technology, having graduated from MIT with an engineering degree before enrolling in Stanford's MBA program. After working as an investment banker at Goldman Sachs and later graduating from Stanford, he sat in a cafe in Palo Alto and coded a site that by 2007 would become the Auteurs Platform, renamed MUBI in 2010. It was a risk. All my savings went into it, kakarel says. So from the beginning, Kakarel was hands on. We built our own content delivery network, our own encoding tool chains, and our own streaming services, he said. But we estimate that it costs us 70% less for our infrastructure than those who rely on other platforms. Fast forward two decades and MUBI, which was recently valued at $1 billion, is nipping at the heels of A24 and Neon, the biggest operators on the indie scene. The company, headquartered in London, is currently on the ground at the Cannes Film Festival, debuting an impressive four films in competition, including Joaquin Trier's Sentimental Value, Ankanola Davies, My Father's Shadow and the History of Sound, A Love Story. That's one of the highest profile films at the festival, thanks to the red hot pairing of Paul mescal and Josh O' Connor. Another of MUBI's con premieres will be Kelly Reichardt's heist thriller the Mastermind, which also stars O' Connor. The first production it has developed and fully financed. End quote. Okay, bit of show housekeeping here. No weekend bonus episodes this weekend and Monday is Memorial Day here in the US So I'm taking Monday off. But I will have a portfolio profile episode for you, taking a look at maybe the most interesting AI investment Chris and I have made to date. Second thing is that Tuesday's show will be a bit late, maybe as late as 3 or 4pm Eastern. I've got doctor's appointments to work my way through, so. So if the show feed is empty when it's usually full, be patient. The episode is coming as soon as I can get it done. Talk to you then.