Author: GAEA Talks - Bewertung: 20x - Views:1526
Forrester technology and innovation forums (Austin, London, New York). Promo code "GAEATECH" for 10% off. Visit https://forrester.com/events/#tech to book your place.
This week on GAEA Talks, Graeme Scott sits down with Alex Whedon - Co-Founder and CTO of Subquadratic, former Head of Generative AI at Tribe AI, former engineer at Meta, Instagram and Stitch Fix, and the architect behind SubQ, the first LLM to break the transformer's quadratic scaling barrier.
Alex has spent ten years working with language models - since before transformers, back when LSTMs were state of the art. His path has been unconventional by design: he dropped out of his bachelor's degree, yet went on to publish at conferences like AAAI, build AI products serving 200 million users at Academia.edu, create the first text-based recommendation system at Stitch Fix, and work on creative monetisation at Instagram. Most recently, as Head of Generative AI at Tribe AI, he led dozens of enterprise implementations for companies including Anthropic, New Relic and Mars, seeing first-hand where AI breaks down inside real businesses. That view convinced him the hardest problems in enterprise AI could only be solved by rethinking the model architecture itself - so he co-founded Subquadratic to do exactly that.
In this episode, Alex explains why the entire AI industry is downstream from one algorithm - the transformer - and why its two fundamental flaws are quietly capping what AI can do. He breaks down quadratic compute scaling in plain terms (10x the input, 100x the compute), the memory wall where context can cost more than the model itself, and how Subquadratic's linear-scaling architecture claims to cut compute by up to 1,000x at extreme context lengths without sacrificing quality. Along the way he makes a bracing macroeconomic argument: that electricity, water, minerals and capital are the real limits on AI's growth, that the industry is being far too stingy with tokens, and that efficiency is not a nice-to-have but an inevitability. This is one of the most genuinely first-principles conversations GAEA Talks has ever recorded.
• Why the whole AI space is downstream from a single algorithm - and what happens when you change it
• Quadratic compute scaling explained simply - why 10x the input means 100x the compute
• The memory wall - how, at a few million tokens, the context can require more memory than the entire model
• What "subquadratic" actually means, and why linear scaling (10x input, 10x compute) is the unlock
• How Subquadratic claims a ~1,000x compute reduction at 12 million tokens without quality trade-offs
• Why most enterprise applications are data-heavy - and why transformers are the worst possible fit for them
• The four axes of better context: larger, higher-intelligence, cheaper and faster - and why "context rot" matters
• Why the first movers built for consumers, not enterprises - and how "no rules" became a business model
• The real constraints on AI's future: electricity, water, precious minerals and capital
• Why we are being "too stingy with the tokens" - and why per-token cost and wattage cannot keep rising
• The DeepSeek lesson - how a lack of GPUs forced a different kind of thinking, and why incumbents didn't learn from it
• Why "first is never usually best" - the hidden burden of educating the market and raising too much capital
• Alex's core philosophy: solve the first-principles problem and you can make every downstream problem irrelevant
About Alex Whedon: Alex Whedon is Co-Founder and CTO of Subquadratic, a frontier AI research and infrastructure company building a new class of large language models with subquadratic scaling laws. In May 2026 the company emerged from stealth with $29M in seed funding and announced SubQ, which it describes as the first frontier model built on a fully subquadratic Sparse Attention (SSA) architecture, with a research context window of up to 12 million tokens. Before Subquadratic, Alex was Head of Generative AI at Tribe AI, where he led more than 40 enterprise AI implementations for companies including Anthropic, New Relic and Mars. Earlier he was a software engineer at Meta and Instagram, built the first text-based recommendation system at Stitch Fix, developed AI products serving 200 million users at Academia.edu, and researched with the World Bank and Blue Cross Blue Shield. He has published at leading AI conferences including AAAI, despite dropping out of his undergraduate degree.
LinkedIn: https://www.linkedin.com/in/alexander-whedon/
Subquadratic: https://subq.ai
GAEA Talks: https://gaeatalks.ai
#AI #ArtificialIntelligence #GAEATalks #EnterpriseAI #Subquadratic #SubQ #LLM #Transformers #SparseAttention #ContextWindow #AIefficiency #AIscaling #AIinfrastructure #MachineLearning #AIcompute #FrontierAI #AIarchitecture #FirstPrinciples #DeepSeek #AIeconomics #FutureOfAI #AIresearch
SOCIAL SHARE CARD GENERATOR