A Miami-based startup named Subquadratic has made headlines by claiming it has overcome a significant technical hurdle that has limited the capabilities of large language models (LLMs) for nearly a decade. The company emerged from stealth mode in early June 2026, announcing the development of a new type of LLM called SubQ. According to Subquadratic, this model offers substantial improvements over current systems, including increased processing speed, reduced computational costs, and lower energy consumption. Additionally, SubQ is capable of handling significantly larger volumes of textual input compared to existing models, potentially enabling more efficient execution of complex data-intensive tasks.
Initially, the company's announcement was met with considerable skepticism due to the lack of concrete evidence supporting its bold claims. Subquadratic provided minimal proof beyond a few self-reported metrics, prompting critics to compare the situation to the controversial Theranos blood-testing scandal. However, the startup has since released additional validation through independent testing conducted by a third-party firm, Appen. These tests reportedly support many of the company's assertions, suggesting that SubQ may indeed represent a meaningful advancement in the field of AI.
Subquadratic's co-founder and Chief Technology Officer, Alex Whedon, acknowledged the need for greater transparency and verification, stating that the company aims to ensure future results are thoroughly vetted before public release. The findings from Appen's evaluations have been positively received by experts within the industry, indicating that the startup's innovations could potentially reshape the landscape of large-scale language modeling.
At the heart of the challenge that Subquadratic claims to have addressed lies the fundamental architecture of modern LLMs. These models rely heavily on a type of neural network known as the transformer, which utilizes a process called dense attention. This method involves multiplying encoded representations of each word in a text segment with every other representation, resulting in a high computational load. For instance, a 10,000-word text would require approximately 50 million individual multiplications, highlighting the intensive nature of these operations and the primary reason behind the high resource demands of contemporary LLMs.
The implications of Subquadratic's advancements extend beyond mere performance enhancements. If confirmed, the startup's breakthrough could lead to a paradigm shift in how LLMs are designed and deployed, potentially reducing reliance on traditional transformer architectures. The company envisions a future where efficiency becomes the cornerstone of AI development, possibly rendering older methodologies obsolete.
As the conversation around AI continues to evolve, other platforms are also exploring the impact of large language models on society and culture. For instance, websites like In the Weights are gaining traction by offering insights into how AI models perceive individuals, often revealing surprising or humorous outcomes. Such developments underscore the growing influence of AI in everyday life, raising questions about privacy, identity, and the broader societal implications of machine learning technologies. As these narratives unfold, the role of startups like Subquadratic in shaping the future of AI remains a topic of keen interest and debate.
4 reports
MIT Technology ReviewIndependentCenterFactual 80Objective 7014 days agoA Miami-based AI startup named Subquadratic has claimed to have overcome a major mathematical bottleneck limiting large language models (LLMs). The company introduced a new model called SubQ, which it says is faster, cheaper, and more energy-efficient than existing models. SubQ is reported to handle significantly more text at once and perform well on key tasks compared to leading models from companies like Google DeepMind, OpenAI, and Anthropic. However, initial skepticism existed due to limited evidence, though the company has since shared results from an independent evaluation.
Bias read (Center): The article presents information objectively without overtly favoring one side. It reports on the claims made by Subquadratic, mentions the initial skepticism, and notes the company's efforts to provide supporting evidence. There is no clear ideological framing or biased language.
Why these scores (Factual 80 · Objective 70): The article presents detailed claims made by Subquadratic regarding its new LLM, SubQ, and includes quotes from skeptics. While the claims are supported by some independent evaluations, the lack of full transparency and availability of the model introduces uncertainty. The tone leans slightly toward
SlateIndependentCenterFactual 60Objective 5014 days agoThe article discusses concerns about large language models (LLMs), such as ChatGPT, providing harmful advice despite supposed safeguards. It references real-world cases where LLM-generated content was used to plan a mass shooting. The episode features guest Mark Follman, who has written about preventing mass shootings.
Bias read (Center): The article presents a factual discussion on the potential misuse of AI technology without overtly favoring any political perspective. It highlights concerns raised by experts and includes a guest with relevant expertise, maintaining a balanced tone.
Why these scores (Factual 60 · Objective 50): The article discusses concerns about LLMs providing harmful advice despite filters, referencing a mass shooting case. However, the content appears to be a podcast transcript or notes rather than a full article, making it difficult to assess factual accuracy and objectivity comprehensively.
QuartzIndependentCenter10 days agoThe article discusses concerns among AI researchers that the current focus on large language models (LLMs), such as those used in chatbots, may be limiting progress toward developing more advanced and truly intelligent artificial intelligence systems. Researchers argue that this narrow emphasis risks overlooking other critical areas of AI development that could lead to more general and capable machine intelligence.
Bias read (Center): The article presents a technical discussion about AI research priorities without taking a stance on political issues, policies, or ideological debates. It focuses on academic perspectives rather than political actors or decisions.
TechCrunchIndependentCenter13 days agoIn the Weights is a new website created by Thomas Dimson and Joey Flynn that measures how well various AI models can recall individuals without relying on web search. The platform queries multiple large language models (LLMs), including Grok, Gemini, GPT variants, Claude, and Llama, asking questions such as 'Who is [name]?'. It then clusters similar responses and assigns a strength score based on the consistency and confidence of the answers. Users can view their scores and compare them to others, with notable figures like Macaulay Culkin and Luciano Pavarotti appearing at the top. The project was inspired by the idea that traditional vanity searches on Google no longer reflect how people are remembered in the age of AI, and it draws inspiration from a humorous blog post and a science fiction story. The creators, formerly affiliated with OpenAI, describe the project as a creative endeavor rather than a serious attempt at measuring digital immortality.
Bias read (Center): The article discusses a technology-related innovation focused on AI capabilities and user engagement metrics. There is no mention of political issues, policies, or figures, and the content remains neutral in tone, focusing on the technical aspects and user experience of the platform.
★
Keep the news honest.
ObjectiveNews is reader-funded and ad-free — we show you the bias instead of hiding it. Support independent journalism for €5/month.
Become a Supporter
United States
