TECH INTELLIGENCE
← Back to all articles
AI & LLM review

Anthropic’s Model Shutdown, Midjourney’s Medical Scanner, and the Rise of Open-Source AI

The Regulatory Cliff: Anthropic’s Forced Shutdown

The most significant event in the artificial intelligence sector recently was not a breakthrough in capability, but an abrupt regulatory intervention. The US government effectively forced Anthropic to suspend access to its advanced models, Fable 5 and Mythos 5, for all users globally. This decision came just days after these models were released, ending their brief window of availability. The catalyst for this shutdown was a complex interplay between export controls and safety concerns. The US government mandated that Anthropic suspend access to any foreign national, regardless of location. Because verifying citizenship in real-time is technically nearly impossible for an online service, the only compliant path forward was to disable the models entirely for everyone. Anthropic initially cited a minor vulnerability as the reason for the shutdown. However, reports suggest that cybersecurity defenders and trusted partners within the industry raised alarms about potential jailbreaks of the model’s guardrails. One notable figure in this chain of events was Amazon CEO Andy Jassy, who reportedly alerted senior administration officials to security risks associated with Anthropic's most advanced models. This is particularly striking given Amazon’s status as one of Anthropic’s largest investors and vendors. The situation highlights a growing tension between AI developers and government regulators. Dario Amodei, Anthropic’s CEO, had previously argued in an essay that frontier AI models should be subject to strict regulation similar to aviation safety standards. He suggested the government should have the power to block or reverse model releases if they pose public safety threats. Ironically, when the government exercised this exact power due to a compliance issue, it created significant frustration within the company. The broader implication of this event is substantial for the AI industry. With Anthropic valued at nearly one trillion dollars, the precedent that the US government can abruptly shut down a commercial product overnight serves as a stark warning for other foundation labs. It underscores the fragility of operating in an unregulated frontier where safety protocols and export laws are still being defined in real-time.

Midjourney Enters Healthcare with Low-Cost Imaging

While Anthropic dealt with regulatory headwinds, Midjourney took a surprising pivot into physical health technology. The company announced the launch of "Midjourney Medical," a new branch dedicated to using advanced imaging technology for healthcare diagnostics. The centerpiece of this initiative is a device designed to replace traditional MRI machines. Instead of expensive magnetic resonance equipment, Midjourney has developed a system that uses ultrasound waves within a specialized tank or ring structure containing thousands of transducers. These devices bounce sound waves through water and around the body to create detailed images of internal structures, such as the spinal cord and abdominal muscles. Midjourney claims this technology can detect biological issues at one-sixtieth of the speed and significantly lower cost compared to conventional MRIs. The company’s long-term vision involves creating "Midjourney Spas," starting with a location in San Francisco scheduled for 2027. These facilities would offer body scanners alongside traditional spa amenities, aiming to democratize health monitoring by encouraging frequent, low-cost screenings rather than reactive medical visits. The move is not without controversy. Critics, including tech commentator Hank Green, have pointed out that ultrasound technology cannot replicate all functions of an MRI or CT scan. Ultrasound struggles with imaging air-filled organs like lungs and bones, making it unsuitable for certain types of cancer screening. However, the potential for affordable, rapid soft-tissue imaging remains a compelling innovation. Midjourney’s entry into this space is driven by its founder David Holz’s background in sensor technology and motion tracking. By combining his expertise in sensors with Midjourney’s image generation capabilities, the company aims to collect vast amounts of biometric data that could eventually train more accurate predictive health models. Notably, Midjourney remains bootstrapped, funding this ambitious healthcare venture through its own profits rather than external investors.

Testing ZAI’s GLM 5.2: The Open-Source Challenger

In the realm of large language models, a new contender has emerged from ZAI with the release of GLM 5.2. This open-weight model is designed specifically for long-horizon coding and agentic tasks, boasting a massive one-million-token context window that rivals frontier proprietary models. GLM 5.2 features 753 billion parameters and operates under an MIT license, allowing developers to download, fine-tune, and deploy the model freely. While running such a large model locally is impractical for most users due to hardware requirements, its availability on cloud platforms makes it accessible for enterprise use. Benchmark results place GLM 5.2 in direct competition with top-tier models like Claude Opus 4.8 and GPT 5.5. On the SweBench Pro benchmark, which tests coding proficiency, GLM 5.2 outperforms GPT 5.5 but trails slightly behind Claude Opus 4.8. In blind user testing via Code Arena, it secured second place overall, beating both GPT 5.5 and Claude Opus in specific web development tasks. The most compelling aspect of GLM 5.2 is its pricing structure. With input costs at $1.40 per million tokens and output costs at $4.40, it offers performance close to Claude Opus at a fraction of the price. This cost efficiency makes it an attractive option for developers building large-scale applications where token usage can quickly become expensive. In practical testing, GLM 5.2 demonstrated strong capabilities in generating code and creating structured content like slide presentations. While its initial attempts at game development showed some jankiness requiring iterative prompting, the model’s ability to understand complex instructions and generate clean, functional output confirms its status as a state-of-the-art open-weight option.

Industry Updates: OpenAI, Meta, and Perplexity

Several other major players have rolled out significant updates this week that are reshaping how users interact with AI tools. OpenAI has introduced "Record and Replay" within its Codex platform. This feature allows the system to watch a user perform a repetitive task on their computer—such as uploading files or formatting data—and then learn to replicate those steps automatically in the future. It essentially creates custom automation skills based on video observation of human behavior, streamlining workflows for routine tasks. Meta has integrated new AI features into Facebook, including an "AI Mode" that provides answers grounded in public posts and discussions across its platform. This mirrors functionality seen in other social media-driven search tools but focuses heavily on community sentiment and real-time information. Additionally, Meta introduced AI-powered photo editing capabilities, allowing users to swap clothing or hair styles within images using generative templates. Perplexity Computer has launched a "self-improving memory" system for its agents. This feature builds a context graph of the work performed by the agent over time, reviewing it at set intervals to teach itself how to perform tasks more efficiently. By learning from mistakes and user corrections, the agent becomes faster and more accurate with continued use, reducing the need for repeated model calls. Adobe continues to expand its AI assistant across Premiere Pro, Illustrator, and InDesign, allowing users to generate sequences or images via text prompts rather than manual editing. Meanwhile, Google has introduced "Ask Ad Manager," a chatbot interface that helps advertisers optimize their campaigns by analyzing performance data and suggesting improvements, similar to tools already available in YouTube Studio.

Public Sentiment on AI

Recent surveys from Pew Research highlight a growing paradox in public opinion regarding artificial intelligence. Approximately half of US adults now report using AI chatbots regularly, with ChatGPT remaining the dominant platform among general users. However, skepticism about the technology’s long-term impact is also rising. Many Americans predict that AI will have a negative effect on society and their personal lives, even as they continue to use it for productivity and convenience. This duality suggests that while people appreciate specific applications of AI—such as coding assistance or information retrieval—they remain wary of broader implications like job displacement, misinformation, and the saturation of media with synthetic content.
RELATED

More in this category

AI & LLM

ChatGPT Work Review: Building Live Websites Without Coding

Explore ChatGPT’s new "Work" mode, which shifts from conversational assistance to autonomous task execution. This review examines how users can build and host complete websites using natural language prompts, analyzing the setup process, prompt engineering techniques, and the practical limitations of this agentic AI tool for non-technical creators.

AI & LLM

OpenAI Overhauls ChatGPT Memory and Merges Codex: What Changes for Users?

OpenAI has significantly upgraded ChatGPT’s memory system to rival competitors like Claude, offering a more structured context management experience. Simultaneously, the company announced plans to merge its powerful Codex agent capabilities directly into the main app, signaling a major shift toward actionable AI workflows for everyday users and professionals alike.