AI Daily Briefing — 2026-06-22

The White House last Friday ordered Anthropic to halt exports of its advanced AI models, Fable and Mythos, citing unspecified national security risks. The company complied within hours, cutting off access for foreign nationals and overseas users. This marks the first major test of whether the U.S. can effectively restrict frontier AI through export controls—a strategy that has historically struggled to contain encryption and spyware. The outcome could set a precedent for how other AI labs navigate global markets.

Anthropic had marketed Mythos as a powerful cyber defense tool, limiting access to roughly 150 vetted organizations. The ban was reportedly triggered by two events: Anthropic granted access to a South Korean telecom (widely reported as SK Telecom, which denies ties to China), and Amazon CEO Andy Jassy alerted officials after researchers found a way around Fable 5’s safeguards. Anthropic disputes the severity of the breach, calling it a narrow, patched issue, but the Commerce Department issued an export directive anyway.

This isn't the first time the U.S. has tried such controls. In the 1990s, the government investigated PGP creator Phil Zimmermann for violating arms export laws, fearing encryption would block intelligence surveillance. Zimmermann fought back by publishing PGP's source code as a printed book, sparking the “Crypto Wars.” The investigation was later dropped, paving the way for end-to-end encryption used by billions today.

Similarly, in the early 2010s, Western-made spyware was found targeting dissidents abroad, leading to expanded Wassenaar Arrangement controls. Yet those measures often failed to stop proliferation. As Anthropic navigates this standoff, history suggests export controls alone may not contain the spread of powerful AI—but the stakes are higher than ever.

A startup named Subquadratic has emerged from stealth mode, asserting it has solved a key mathematical bottleneck that has limited large language models for nearly a decade. The company claims its approach drastically reduces the number of computations required by transformers, resulting in faster, cheaper, and more energy-efficient AI. While experts remain skeptical, Subquadratic has begun sharing evidence that suggests its method may be worth serious attention, as reported by MIT Technology Review AI.

In other tech news, brain-computer interface (BCI) trials are rapidly expanding. Casey Harrell, a man with ALS, has become a pioneering user of a brain implant that has allowed him to work, reconnect with loved ones, and read to his daughter. He describes the device as "nothing short of revolutionary." The number of trial volunteers has soared recently, with China becoming the first nation to approve a BCI for medical use, signaling a shift from lab research toward commercial applications.

Meanwhile, Amazon engineers who advocated for limits on data center expansion now face potential termination after company investigations. In science, a new fossil discovery challenges 150 years of evolutionary theory, suggesting early land vertebrates skipped the tadpole stage. Senator Bernie Sanders has proposed legislation to give Americans direct ownership in AI firms through a sovereign wealth fund, funded by a one-time tax on AI company stock. These developments highlight the accelerating pace of change in technology and science.

A Miami-based AI startup called Subquadratic has emerged from stealth with a bold assertion: it has solved a major mathematical hurdle that has limited large language models for nearly a decade. The company unveiled SubQ, a new type of LLM it claims is faster, cheaper, and far more energy-efficient than existing models. Early skepticism was high due to a lack of evidence, but the startup has now released independent test results that lend credibility to its claims.

According to Subquadratic, SubQ can process up to 12 times more text at once than most models, enabling tasks like analyzing entire code bases or hundreds of documents. The company says it matches the performance of top models from Google DeepMind, OpenAI, and Anthropic on key benchmarks like coding. Third-party evaluator Appen ran tests that largely validated these results, with Appen’s Jeanine Sinanan-Singh calling the architecture a potential “game changer” for speed and efficiency.

Despite the promising data, Subquadratic has yet to make SubQ widely available, fueling comparisons to Theranos. Co-founder Alex Whedon acknowledged the skepticism, noting that releasing third-party benchmarks earlier would have helped. CEO Justin Dangel believes the breakthrough could shift the industry away from transformer-based models, stating, “We hope we’re kicking off a new age of efficiency.”

Source: MIT Technology Review AI

The Trump administration has ordered Anthropic to take its two newest AI models, Fable 5 and Mythos 5, offline, citing undisclosed national security concerns. The move, which came via an export control order, forced the company to remove the models entirely to avoid violating restrictions on foreign national access. Reports indicate the White House acted after Amazon CEO Andy Jassy raised alarms about researchers allegedly bypassing Fable 5’s guardrails. The lack of transparency around the specific risks has drawn sharp criticism from cybersecurity experts, who signed an open letter urging the administration to revoke the order, warning it could harm U.S. network defenders.

On the latest episode of TechCrunch’s Equity podcast, analysts examined the motivations behind the crackdown. Sean O’Kane noted that Anthropic’s strained relationship with the Trump administration sets it apart from other leading AI labs, suggesting rivals may face less scrutiny. However, Rebecca Bellan highlighted that the security risks tied to Anthropic’s models are not unique, raising questions about whether the administration is using policy to target a specific company. The episode also explored how the controversy could inadvertently boost Anthropic’s profile, with Bellan remarking that “everybody loves a bad boy.”

The situation underscores a volatile regulatory landscape for AI firms. While competitors might benefit from Anthropic’s troubles, the lack of clear guidelines leaves the entire industry vulnerable to sudden, opaque enforcement actions. As the administration juggles other priorities, including negotiations over a war in Iran, the Anthropic case serves as a cautionary tale about the intersection of politics, national security, and emerging technology.

A new website called In the Weights is offering a fresh twist on the classic vanity search, promising to reveal how well artificial intelligence models remember you. Created by former OpenAI designers Thomas Dimson and Joey Flynn, the platform queries multiple large language models—including Grok, Gemini, and several versions of GPT—asking them to identify a given name. It then clusters the responses and assigns a strength score, essentially measuring how deeply a person is encoded in an AI's training data. The founders argue that traditional web searches no longer capture a person's digital footprint, as chatbots increasingly become the primary source of information about individuals.

The site's name refers to the numerical parameters that shape an AI model's behavior. As the creators explain, being "in the weights" means your existence was deemed important enough to be remembered during the process of building superhuman artificial intelligence. Early results show a competitive leaderboard, with celebrities like Macaulay Culkin and Luciano Pavarotti topping the charts. Even a humble tech blogger received a score of 641, placing them in the top 6% of names, though colleagues quickly outperformed them.

Dimson told TechCrunch AI that the project was born from a desire to "get the creative juices flowing again" after leaving OpenAI. He noted that Google vanity searches have become obsolete as more traffic shifts to large language models. The site also highlights potential hallucinations, such as one model claiming a common name refers to multiple people. While some critics dismiss the tool as simply asking multiple chatbots for self-descriptions, its retro Nintendo-inspired design and addictive comparison feature have struck a nerve with users eager to see if they live forever in the superintelligence.