Two Developments Reshaping AI’s Near-Term Trajectory
A startup that emerged from stealth last month says it has solved a core mathematical problem limiting large language models, while brain-computer interface research is reaching patients outside lab settings at a pace researchers hadn’t anticipated.
Subquadratic’s Claim and the Skepticism It Has Not Fully Silenced
Subquadratic, an AI startup, surfaced publicly in May with a striking assertion: it had resolved a mathematical bottleneck that has constrained large language models for nearly a decade. The company’s approach targets the transformer architecture itself – specifically the number of computations a transformer must execute to produce an output. By cutting that number down sharply, Subquadratic says its models run faster, cost less to operate, and consume significantly less energy than competing systems currently on the market.
The efficiency problem in LLMs is not abstract. Transformers scale their computational demands quadratically with context length – meaning that processing longer inputs requires exponentially more compute. That relationship has pushed energy costs and hardware requirements upward as models have grown. If Subquadratic has genuinely altered that mathematical relationship, the downstream effects on infrastructure costs and energy consumption would be substantial.
Many researchers responded to the initial announcement with skepticism, which is a reasonable default when a startup claims to have fixed something the field has struggled with for years. What has changed since the stealth exit is that Subquadratic has begun sharing supporting data. Researchers who have reviewed that data suggest the approach merits serious attention, though significant doubts remain about whether the results will hold across real-world deployment conditions.
The tension here is familiar in AI research: a promising result in controlled settings does not automatically transfer to production environments with unpredictable inputs, diverse use cases, and the full weight of enterprise-scale demand. Subquadratic has not yet fully answered those questions, and some researchers remain unconvinced even after seeing the preliminary evidence.
Brain Implants Move from Experiments to Everyday Life
The trajectory of brain-computer interface research has shifted from controlled academic trials to something closer to functional independence for patients. Casey Harrell, a man living with ALS, has become what researchers are calling “the first power user” of a brain implant. Through the device, Harrell has maintained an income, rebuilt connections with friends and family, and read to his daughter – activities that ALS would otherwise have made impossible. He described the technology as “nothing short of revolutionary.”
Harrell’s case is one point in a broader pattern. Over the past two years, the number of volunteers enrolling in BCI clinical trials has increased sharply. The scale of participation has grown fast enough that researchers are now dealing with questions about long-term device performance, user training, and feature expansion that were previously theoretical concerns.
In a development with significant regulatory implications, China became the first country to approve a BCI device for medical use in 2025. That approval marks a transition point – BCIs shifting from experimental procedures requiring research exemptions to sanctioned medical treatments with defined clinical pathways. The regulatory recognition changes how manufacturers, hospitals, and insurers approach the technology.
Engineering advances are also allowing BCI developers to pack more features into implanted systems. Increased electrode density, improved signal processing, and better software interfaces have expanded what patients can control and communicate through their devices. The gap between what early BCI prototypes could do and what current systems offer is wide enough that comparisons between generations are becoming difficult.
What remains unresolved is the long-term durability question. Brain implants operate inside living tissue, which responds to foreign objects over time. Scar tissue formation, signal degradation, and device longevity are challenges that clinical trials are only beginning to generate meaningful data on. As the number of trial participants grows and implants remain in place for longer periods, the field will have more answers – but also more complications to navigate publicly.
Other Stories Worth Tracking
Amazon engineers who testified at meetings advocating for limits on data center expansion are now reportedly under investigation by the company and may face termination. The workers filed a joint complaint with Seattle’s Office for Civil Rights. Separately, new research published in Nature indicates that AI may already be weakening the skills of doctors and engineers who rely on it heavily – a finding that adds pressure to ongoing debates about how professionals should integrate AI tools without ceding core judgment to them. Research on attention and AI use has flagged related concerns about dependency and cognitive offloading.
On the policy side, Senator Bernie Sanders has introduced legislation to create an AI sovereign wealth fund, financed through a one-time tax on AI companies’ stock, that would distribute annual payments directly to Americans. Meanwhile, a Nature-adjacent finding from Quanta suggests the structural properties of the human genome may interfere with AI models designed to analyze biology and disease – which would limit the usefulness of an entire category of AI health applications before those applications reach clinical scale. Tech workers who previously pushed their AI usage to the maximum are now pulling back, as spiraling token costs have made unrestricted use financially unsustainable for individuals and teams.
