Relevance: supporting · Type: background
Confidence100%
Subquadratic is an artificial intelligence startup based in Miami.
Relevance: primary · Type: event
Confidence100%
Subquadratic announced its exit from stealth mode last month.
Relevance: primary · Type: action
Confidence100%
Subquadratic developed a large language model called SubQ.
Relevance: primary · Type: action
Confidence100%
Subquadratic claims SubQ can process up to 12 times as much text at once as most other models.
Relevance: primary · Type: action
Confidence100%
Subquadratic claims SubQ matches the performance of models from Google DeepMind, OpenAI, and Anthropic on key coding tasks.
Relevance: primary · Type: action
Confidence100%
Subquadratic shared the results of an independent evaluation of SubQ conducted by third-party firm Appen.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"We expected healthy skepticism," says Alex Whedon.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism, which is why we’re taking the time to make sure any future results are fully verified before putting them out," says Alex Whedon.
Relevance: primary · Type: action
Confidence100%
Appen evaluated SubQ on standard tests.
Relevance: primary · Type: event
Confidence100%
Appen found that SubQ was 56 times faster than models using FlashAttention in a baseline speed test.
Relevance: primary · Type: event
Confidence100%
SubQ scored 89.7% on LiveCodeBench, a test assessing performance on competitive coding problems.
Jeanine Sinanan-Singh, director of generative AI research
Relevance: supporting · Type: quote
Confidence100%
"This model continues to provide frontier-level performance in coding," says Jeanine Sinanan-Singh.
Relevance: primary · Type: background
Confidence100%
SubQ has a context window of up to 12 million tokens.
Relevance: supporting · Type: background
Confidence100%
Most top large language models today have context windows of one million tokens.
Relevance: supporting · Type: background
Confidence100%
Justin Dangel is the cofounder and CEO of Subquadratic.
Justin Dangel, CEO
Relevance: supporting · Type: quote
Confidence100%
"We hope we’re kicking off a new age of efficiency," says Justin Dangel.
Justin Dangel, CEO
Relevance: supporting · Type: quote
Confidence100%
"We don’t think anybody will be building on transformers in a few years," says Justin Dangel.
Relevance: primary · Type: background
Confidence100%
Subquadratic uses sparse attention instead of dense attention in its model architecture.
Relevance: primary · Type: background
Confidence100%
Subquadratic dynamically selects which words to focus on during processing.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"Sparse attention says not all of those relationships are important, because they’re not," says Alex Whedon.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"Historically, most mechanisms have used fixed patterns, like always comparing the first word to the fifth," says Alex Whedon.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"That’s pretty limiting," says Alex Whedon.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"Language is too sophisticated for that," says Alex Whedon.
Alex Whedon, chief technology officer
Relevance: supporting · Type: quote
Confidence100%
"And so, one of the things that makes our mechanism unique is that we dynamically select which ones are important," says Alex Whedon.
Relevance: supporting · Type: background
Confidence100%
Dan McAteer is an artificial intelligence engineer.
Dan McAteer, artificial intelligence engineer
Relevance: supporting · Type: quote
Confidence100%
"SubQ is either the biggest breakthrough since the Transformer ... or it’s AI Theranos," says Dan McAteer.
Relevance: supporting · Type: background
Confidence100%
Will Depue is an independent AI researcher who previously worked at OpenAI.
Will Depue, independent AI researcher
Relevance: supporting · Type: quote
Confidence100%
"Pretty much everything under the sun has been attempted," says Will Depue.
Will Depue, independent AI researcher
Relevance: supporting · Type: quote
Confidence100%
"It’s not impossible, but it’s akin to running a four-minute mile," says Will Depue.
Relevance: primary · Type: action
Confidence100%
Subquadratic claims it costs eight dollars to run SubQ through the RULER 128 test.
Relevance: supporting · Type: action
Confidence100%
Justin Dangel stated that it costs $2600 to run Anthropic's LLM Opus 4.6 through the RULER 128 test.
Relevance: supporting · Type: background
Confidence100%
RULER 128 is a test developed by Nvidia to assess a model's ability to retrieve information from large data sets.
Relevance: supporting · Type: background
Confidence100%
SubQ is not yet widely available for public use.
Relevance: supporting · Type: background
Confidence100%
Jeanine Sinanan-Singh is Appen’s director of generative AI research.
Jeanine Sinanan-Singh, director of generative AI research
Relevance: supporting · Type: quote
Confidence100%
"That was really exciting to me, it validated their architecture," says Jeanine Sinanan-Singh.
Jeanine Sinanan-Singh, director of generative AI research
Relevance: supporting · Type: quote
Confidence100%
"I was like, ‘Wow, this could be a game changer,’ because models struggle with speed and inefficiency," says Jeanine Sinanan-Singh.
Jeanine Sinanan-Singh, director of generative AI research
Relevance: supporting · Type: quote
Confidence100%
"But when you have kind of shocking results, it’s really not as credible when you say it yourself," says Jeanine Sinanan-Singh.
forum Comments (0)
No comments yet. Be the first to comment.